introduction to artificial intelligencesite.iugaza.edu.ps/mtastal/files/ch0-ai-dl-ml.pdfintroduction...

56
Chapter 1: Introduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial Intelligence Winter 2019

Upload: others

Post on 01-Jan-2021

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Chapter 1

Introduction to Artificial Intelligence

Prepared by

Dr Mohammed Taha El Astal

TMUL 3310

Artificial Intelligence

Winter 2019

Content

Syllabus and course objective

Motivation toward AI

AI definition

AI evolution

AI Why now

AI ML and DL

Supervised and Unsupervised Learning

ML flow

Facebook advancements in AI

Concerns about AI

Course outlines

Valentino Zocca et all Denis Python Deep Learning Packt

Publishing 2nd edition January 2019

Course Outlines

Introduction to Neural Networks

Introduction to PyTorch

Convolutional Neural Networks

Style Transfer

Recurrent Neural Networks

Sentiment Prediction with an RNN

Deploying PyTorch Models

Artificial Intelligence Definition

bull John McCarthy widely recognized as one of the godfathers of AI defined it as ldquothe science and engineering of making intelligent machinesrdquo

A branch of computer sciencedealing with the simulation ofintelligent behavior in computers

The capability of a machine toimitate intelligent human behavior

A computer system able to perform tasks that normally require humanintelligence such as visual perception speech recognition decision-making and translation between languages

FAIR

bull Discussion what fields

you realize it use AI

bull FAIR Facebook Artificial

Intelligence Research

bull More than 460 solid

publication in less than 8

years

bull PhD and internship

programs

The Rise of Artificial Intelligence | Documentary HDhttpswwwyoutubecomwatchv=2EWDDUGNhKM

Motivation toward AI

Evolution of AI

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 2: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Content

Syllabus and course objective

Motivation toward AI

AI definition

AI evolution

AI Why now

AI ML and DL

Supervised and Unsupervised Learning

ML flow

Facebook advancements in AI

Concerns about AI

Course outlines

Valentino Zocca et all Denis Python Deep Learning Packt

Publishing 2nd edition January 2019

Course Outlines

Introduction to Neural Networks

Introduction to PyTorch

Convolutional Neural Networks

Style Transfer

Recurrent Neural Networks

Sentiment Prediction with an RNN

Deploying PyTorch Models

Artificial Intelligence Definition

bull John McCarthy widely recognized as one of the godfathers of AI defined it as ldquothe science and engineering of making intelligent machinesrdquo

A branch of computer sciencedealing with the simulation ofintelligent behavior in computers

The capability of a machine toimitate intelligent human behavior

A computer system able to perform tasks that normally require humanintelligence such as visual perception speech recognition decision-making and translation between languages

FAIR

bull Discussion what fields

you realize it use AI

bull FAIR Facebook Artificial

Intelligence Research

bull More than 460 solid

publication in less than 8

years

bull PhD and internship

programs

The Rise of Artificial Intelligence | Documentary HDhttpswwwyoutubecomwatchv=2EWDDUGNhKM

Motivation toward AI

Evolution of AI

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 3: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Course outlines

Valentino Zocca et all Denis Python Deep Learning Packt

Publishing 2nd edition January 2019

Course Outlines

Introduction to Neural Networks

Introduction to PyTorch

Convolutional Neural Networks

Style Transfer

Recurrent Neural Networks

Sentiment Prediction with an RNN

Deploying PyTorch Models

Artificial Intelligence Definition

bull John McCarthy widely recognized as one of the godfathers of AI defined it as ldquothe science and engineering of making intelligent machinesrdquo

A branch of computer sciencedealing with the simulation ofintelligent behavior in computers

The capability of a machine toimitate intelligent human behavior

A computer system able to perform tasks that normally require humanintelligence such as visual perception speech recognition decision-making and translation between languages

FAIR

bull Discussion what fields

you realize it use AI

bull FAIR Facebook Artificial

Intelligence Research

bull More than 460 solid

publication in less than 8

years

bull PhD and internship

programs

The Rise of Artificial Intelligence | Documentary HDhttpswwwyoutubecomwatchv=2EWDDUGNhKM

Motivation toward AI

Evolution of AI

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 4: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Artificial Intelligence Definition

bull John McCarthy widely recognized as one of the godfathers of AI defined it as ldquothe science and engineering of making intelligent machinesrdquo

A branch of computer sciencedealing with the simulation ofintelligent behavior in computers

The capability of a machine toimitate intelligent human behavior

A computer system able to perform tasks that normally require humanintelligence such as visual perception speech recognition decision-making and translation between languages

FAIR

bull Discussion what fields

you realize it use AI

bull FAIR Facebook Artificial

Intelligence Research

bull More than 460 solid

publication in less than 8

years

bull PhD and internship

programs

The Rise of Artificial Intelligence | Documentary HDhttpswwwyoutubecomwatchv=2EWDDUGNhKM

Motivation toward AI

Evolution of AI

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 5: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

FAIR

bull Discussion what fields

you realize it use AI

bull FAIR Facebook Artificial

Intelligence Research

bull More than 460 solid

publication in less than 8

years

bull PhD and internship

programs

The Rise of Artificial Intelligence | Documentary HDhttpswwwyoutubecomwatchv=2EWDDUGNhKM

Motivation toward AI

Evolution of AI

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 6: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Evolution of AI

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 7: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

AI Why now

Huge computational and storage capabilitiesavailable currently

Its all about Big Data was available2020 44 ZB (Zeta=)

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 8: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

AI Why now (Cont)

GPU Graphical Processing Unit Specific purpose circuits

designed to perform specifictasks

like video rendering gaming Thousands of relatively

simple cores amp parallelprocessing

CPU Central Processing Unit It is designed to be general

purpose IC in order to performsa verity of operations

like computation watchingmovies internet surfingetc

Few cores amp Sequential serialprocess

TPU Tensor Processing Unit fine-tuned PU for deep

learning Refer to self-reading

httpscloudgooglecomblogproductsai-machine-learningwhat-makes-tpus-fine-tuned-for-deep-learning

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 9: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

AI Why now (Cont)

bull Facebook Infrastructure case (32018)

bull Open Sourced Hardware

Big Sur Big Basin

Big Basin Volta (32018)

7 Tflops 106 Tflops and

even more in Volta version

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 10: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

bull Google cloud TPU case (32018)

AI Why now (Cont)

Build and calculate your modelrsquos Pricehttpscloudgooglecomproductscalculator httpswwwyoutubecomwatchv=78P0pBj-i4c

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 11: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

AI Why now (Cont)

Data Source Examples

bull Crowd sourcing platform

bull Captioning in video correction

bull Hashtag comments postweakly supervised

bull Not only bigger amount ofimages but also large number oflabels and categories because ofhashtag

This allow AI to become real

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 12: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

The Difference between AI Machine and Deep Learning

bull Although the three terminologies are usually used interchangeably theydo not quite refer to the same things

bull AI can be loosely interpreted tomean incorporating humanintelligence to machines

bull AI is the broader concept thatconsists of everything from GoodOld-Fashioned AI all the way tofuturistic technologies such as deeplearning

bull AI-powered machines are usuallyclassified into two groups mdash generaland narrow

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 13: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

The Difference between AI Deep and Machine Learning (Cont)

bull ML can be loosely interpreted to mean empowering computer with theability to ldquolearnrdquo

bull It is to enable machines to learn bythemselves using the provided dataand make accurate predictions

bull ML is a subset of AI in fact itrsquos simplya technique for realizing AI

bull Training in machine learning entailsgiving a lot of data to the algorithmand allowing it to learn more aboutthe processed information

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 14: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

The Difference between AI Deep and Machine Learning (Cont)

bull DL is a subset of ML itrsquos simply a technique for realizing ML DL is the nextevolution of machine learning

bull DL algorithms are roughly inspired by theinformation processing patterns found inthe human brain

bull while DL can automatically discover thefeatures to be used for classification MLrequires these features to be providedmanually

bull In contrast to ML DL needs high-endmachines and considerably big amountsof training data to deliver accurateresults

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 15: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

The Difference between AI Deep and Machine Learning (Cont)

httpswwwyoutubecomwatchv=q7bKMHdxtPU

bull This video illustrates simply the boarders between AI ML and DL

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 16: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

ML Execution Flow

Data Features Training Evaluation Inference

OnlineOffline

High Storage Capabilities High Computing Capabilities

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 17: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Artificial Neural Networks (ANN)

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 18: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Simple NN demonstration

bull httpswwwyoutubecomwatchv=3JQ3hYko51Yampt=34s

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 19: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Deep Learning

Deep Learning

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 20: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Supervised Learning

Expert realndashstate agent

Junior agents

Just by looking

Can estimate price

By his experience or previous data can obtain some training labeled data

area room location Price

1 400 3 rural 220K

2 240 2 Near city 340k

3 110 2 downtown 510K

This data can teach juniors his experience

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 21: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Supervised Learning (Cont)

Can you deduce what type of operations were in quiz

Human know + -

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 22: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Even if it is not complete but you caneven deduce some useful information

Unsupervised Learning

area room location Price

1 400 3 rural

2 240 2 Near city

3 110 2 downtown

Large house means out of cityInside city means small number of rooms

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 23: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Unsupervised Learning

Can you deduce what these number means

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 24: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Supervised amp Unsupervised Learning

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 25: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Semi-Supervised Learning

bull Problems where you data (X) and only some of the data is labeled (Y) arecalled semi-supervised learning problems

bull A good example is a photo archive whereonly some of the images are labeled (egdog cat person) and the majority areunlabeled

bull Many real world machine learningproblems fall into this area This isbecause it can be expensive or time-consuming to label data as it may requireaccess to domain experts Whereasunlabeled data is cheap and easy tocollect and store

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 26: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Semi-Supervised Learning (Cont)

You can use unsupervised learning techniques to discover and learn the

structure in the input variables

You can also use supervised learning techniques to make best guess

predictions for the unlabeled data feed that data back into the supervised

learning algorithm as training data and use the model to make predictions

on new unseen data

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 27: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Part II Example of inventions amp Apps use AI of

Google Facebook Tesla NVIDIA

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 28: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Vision Detectron

bull Detectron State-of-the-art platform for

object detection research

bull Started in July 2016 and released via

Github in Jan 22 2018

bull Uses Mask R-CNN Faster R-CNN

RPN Fast R-CNNetc

bull Dataset Common Objects in Context

(COCO) 330K15M object instances httpswwwyoutubecomwatchv=_zZe27JYi8YFor illustration purpose only it uses other product

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 29: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

CONTNew of 2018

bull Improved in term of

bull accuracy

bull speed

bull and smooth with mobiles

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 30: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Using Detectron

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 31: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Vision DensePose

bull Needed to understand the 3d of video

bull DensePose is designed to map all human

pixels of an RGB image to a 3D surface-

based representation of the human body

bull 50K humans collecting more than 5M

bull Announced 2018

bull Coming soon via Github

httpswwwyoutubecomwatchv=_zZe27JYi8Y

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 32: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Speechwav2letter

bull Wav2letter is an end-to-end system to transcribe speech

bull Published in literature 2016 amp 2017

bull Available at Github

bull Robustness of language variance

bull DataSet Openslrorg (1000 hours) corpus of read English speech

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 33: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Languagemanys

bull GitHub translate

bull 6000 languages and its variant

bull GitHub muse Early work differentiate betweenlanguages

bull GitHubfasttext is a for text representation andclassification

bull GitHubParlAI is conversational models acrossmultiple tasks

bull GitHubFairseq translation summarization andother text generation tasks

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 34: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

M Messenger

bull M messenger enhanced with Fb

AI chatbot to assist users with

comfortable experience while

communication

bull New Recommendation real-

time translation (ie Buyer amp

seller) 559 Billions day

social recommendation

Plan recommendationhellipetc

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 35: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

ContChabot

bull Based on NLP and Witai

bull 300000 bots developed for last two years

bull Businesses and customers now exchange 8 billion

messages a day on Messenger

bull Easy to deploy not more 30 minutes for

experienced developers

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 36: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

bull Three main blocks of this

1 object detection amp recognition

2 precise location determination

3 3d effects inclusion

ContAR Camera effects

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 37: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

ContAR Camera Experience with bots

bull Last year at F8-2017 one of Facebookrsquos mission is to

make the camera the first augmented reality

platform

bull Progress toward that goal continued today with an

introduction of AR for Instagram and Facebook

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 38: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

3D memories constructing

bull Using vision AI (Image amp

Video) then VR

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 39: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Blood Donation

bull For Social Impact

bull NLP

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 40: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Suicide Prevention

bull For Social Impact

bull NLP and Caffee2

bull Score threshold amp

weighting for priority

bull 2X faster 1K+ higher

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 41: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Protecting from Bad contents

bull Uses AI to identify objectionable

content or fake accounts

bull Dataset 10s of millions reportsweek

bull This takes gt05M scam accounts

down last month

bull Fairness Flow that can measure for

potential biases for or against

particular groups of people (gender

equity of jobs recommendations)

bull Election bull fake account behavior (large activity activity

associated with non-origin of the account )andor

its starting

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 42: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Protecting You

bull Your profile image used

will notify you

bull Some country with no

Image Profile protect you

too

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 43: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

DeSIGN

bull Teaching machine creativity is

not easy task

bull lsquoDeSignrsquo system employs GANs

bull Published 42018

bull GAN unsupervised machine

learning implemented by

2 neural networks contesting with

each other in a zero-sum

game framework

bull DataSet 20K Images with clothes

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 44: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Visual Dialog

bull Visual Dialog (of A) uses

1 Detectron

2 Dialog ParlAI hellipetc (GitHub)

3 Dataset VisDial 120K Image

from COCO 1 dialog with 10

questions each 12M QampA

Available in visualdialogorgdata

4 Applications Aid visually

Impaired people and A project

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 45: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

House3D

bull Some scenarios useful for

reasoning is house3D

bull Released in F8-2018 via GitHub

bull httpsgithubcomfacebookrese

archHouse3D

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 46: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

A-STAR

bull Project ASTAR (A) uses AI to

build agent can see talk act and

reason

bull Dataset

VisDial + 767 homes times

12rooms times 50 objects times 5281

questions

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 47: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

TorchCraft

bull Strategic Games are the

best place for reasoning

training

bull Reasoning is something

we do unconscious

bull Ready now for

researcher amp developers

httpsgithubcomTorch

CraftTorchCraft

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 48: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

ELF

bull FAIR open sourced ELF OpenGo

whose defeated world

champion

bull Trained model and the code

used to create it are available

now via

httpsgithubpytorchelf

bull Source Opened on 252018 to

help in mission of developing

self-learning one

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 49: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

bull One of most interestingapplication of AI in real life

bull Many companies work onthis

Tesla

Google Waymo 360

Nvidia

httpswwwyoutubecomwatchv=vmgDAa1Fy7E

Self-Driving Cartesla

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 50: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Google Waymo 360

httpswwwyoutubecomwatchv=B8R148hFxPwampt=25s

Self-Driving Cargoogle

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 51: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Human Concerns about AI

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 52: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

18 Million jobs

ge500K jobs

23 Million jobs

AI impact by 2020

Elimination Creation

bull Source Gartner research firm

Based on Forecasting

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 53: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Based on History

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 54: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Singularity

bull The Singularity is the

hypothetical future creation of

super-intelligent machines

bull Superintelligence is defined as a

technologically-created cognitive

capacity far beyond that possible

for humans

bull Should the Singularity occurhttpswwwyoutubecomwatchv=Gm9d-VOv7Mo

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019

Page 55: Introduction to Artificial Intelligencesite.iugaza.edu.ps/mtastal/files/ch0-AI-DL-ML.pdfIntroduction to Artificial Intelligence Prepared by Dr. Mohammed Taha El Astal TMUL 3310 Artificial

Dr Mohammed Taha El AstalDrmastalgmailcom

4022019