WS2015/2016, 09.12.2015

Intelligent Systems

Introduction to Machine Learning

Carsten Rother, Dmitrij Schlesinger

09.12.2015Intelligent Systems: Introduction 2

Machine Learning everywhere

Recommender Systems

Slide credits: Bernt Schiele, Stefan Roth, Peter Gehler …

09.12.2015Intelligent Systems: Introduction 3

Machine Learning everywhere

AdWords predictions

09.12.2015Intelligent Systems: Introduction 4

Machine Learning everywhere

Ranking web queries

09.12.2015Intelligent Systems: Introduction 5

Machine Learning everywhere

Image segmentation

09.12.2015Intelligent Systems: Introduction 6

Machine Learning everywhere

Body part detection on Kinect

09.12.2015Intelligent Systems: Introduction 7

What is Machine Learning

Semantics and Structure

Database of Text

Images and Videos

Symbolic Data and Measurements

2D, 3D microscopy

09.12.2015Intelligent Systems: Introduction 8

What is Machine Learning

Goal of machine learning:

Machines that learn to perform a task from experience

We can formalize this as:

is called output variable,

the input variable and

the model parameters

learn... adjust the parameters

… to perform …

... a task ... the function

... from experience using a training dataset , where of either

or

09.12.2015Intelligent Systems: Introduction

Performance: “99% correct classification”

- of what ?

- i.e. on speech recognition (correct words, characters, speakers

identification etc.)

- over which dataset ?

- remember on “precision/recall” from the previous lectures

9

What does it mean “to perform” ?

“The car drives without human intervention 99% of

the time on country roads”

Is 99% good enough ?

1% false alarm for 300.000 passengers at airport Frankfurt ?

09.12.2015Intelligent Systems: Introduction

• Supervised learning: there is a completely labelled (annotated) dataset

• Unsupervised learning: no annotations at all

• Reinforcement learning: there is a reward for correct action/recognition

Some variants:

• Semi-Supervised learning: dataset is partially labelled

• Transductive learning: no model, just find the “true” answer

• On-line learning: the training samples are not available at a time

• Large-scale learning: lots of data

• Active learning: how to ask a “teacher” for an annotation ?

10

Different learning scenarios

09.12.2015Intelligent Systems: Introduction

Training:

training data

Testing:

Testing data (different from training)

11

General Paradigm (supervised learning)

Unknown parameter

Learn

Predict0,1,2 …

prediction

09.12.2015Intelligent Systems: Introduction

Assume, it is necessary to place a (predefined) number of radio

stations in order to supply the signal of the best quality.

We should:

• Assign each house to a station

• find the best station positions (in order to minimize e.g. the total

distance)

If the assignments are known, it would be a supervised learning. Now

however, we have to find both classification (house assignments) and

unknown parameters (station positions) simultaneously.

12

Unsupervised learning

09.12.2015Intelligent Systems: Introduction

Given a situation find an action to maximize a reward function

No optimal output action is presented

Each action yields a reward (which may have to be approximated)

Exploration: try out new actions

Exploitation: use known actions that yield high rewards

Find a good trade-off between exploration and exploitation

Examples: learning robot motor primitives: ball in a cup, elevator

scheduling, Playing Backgammon/Go …

13

Reinforcement Learning

09.12.2015Intelligent Systems: Introduction 14

Going back to 1973• Sir James Lighthill report to the British Parliament

Full report on Youtube: http://www.youtube.com/watch?v=FLnqHzpLPws&list=PL27303EC6EC90FD5A

09.12.2015Intelligent Systems: Introduction 15

Going back to 1973• Sir James Lighthill report to the British Parliament

He specifically mentioned the problem of "combinatorial explosion" or "intractability",

which implied that many of AI's most successful algorithms would grind to a halt on

real world problems and were only suitable for solving "toy" versions.

09.12.2015Intelligent Systems: Introduction 16

Questions in Machine Learning

• What is the modelling language:

undirected / directed Graphical models; unstructured models, …

• How does the model look like:

What is the structure?

How do the functions look like?

• Can we learn the Model from Data:

Learn structure

Learn potential functions

• Probabilistic Learning / Discrimantive Learning

• How do we optimize the model (perform inference):

fast, approximate

• Exactly solvable?

09.12.2015Intelligent Systems: Introduction 17

What is Machine Learning

Goal of machine learning:

Machines that learn to perform a task from experience

We can formalize this as:

is called output variable,

the input variable and

the model parameters

learn... adjust the parameters

… to perform …

... a task ... the function consider „probabilistic models“

... from experience using a training dataset , where of either

or

09.12.2015Intelligent Systems: Introduction

09.12: Introduction to Machine Learning, Probability Theory

16.12: Decision Making, Statistical Learning (Schlesinger)

06.01: Directed Graphical Models (Rother)

13.01: Undirected Graphical Models (Rother)

20.01: Neural Networks (Schlesinger)

27.01: Reinforcement Learning (Rother)

02.02: Wrap-up (Rother)

First exercise: 14.12. about Probability Theory (Heidrich)

Course homepage: http://cvlab-dresden.de/courses/intelligente-systeme/

18

Lecture overview