angular and deep learning
TRANSCRIPT
Deep Learning and Angular
Angular Meetup (06/14/2017)
Google (Mountain View)
Oswald Campesato
The Data/AI Landscape
Gartner Hype Curve: Where is Deep Learning?
The Impact of AI
“Robot trucks will kill far fewer people (if any).
Machines don’t get distracted or look at phones
instead of the road.
Machines don’t drink alcohol, do drugs, or things that
contribute to accidents.”
Robot trucks don’t need salaries, vacations, health
insurance, rest periods, or sick time.
The only costs will be upkeep of the machinery.
AI/ML/DL: How They Differ
Traditional AI (20th century):
based on collections of rules
Led to expert systems in the 1980s
The era of LISP and Prolog
AI/ML/DL: How They Differ
Machine Learning:
Started in the 1950s (approximate)
Alan Turing and “learning machines”
Data-driven (not rule-based)
Many types of algorithms
Involves optimization
AI/ML/DL: How They Differ
Deep Learning:
Started in the 1950s (approximate)
The “perceptron” (basis of NNs)
Data-driven (not rule-based)
large (even massive) data sets
Involves neural networks (CNNs: ~1970s)
Lots of heuristics
Heavily based on empirical results
The Rise of Deep Learning
Massive and inexpensive computing power
Huge volumes of data/Powerful algorithms
The “big bang” in 2009:
”deep-learning neural networks and NVidia GPUs"
Google Brain used NVidia GPUs (2009)
AI/ML/DL: Commonality
All of them involve a model
A model represents a system
Goal: a good predictive model
The model is based on:
Many rules (for AI)
data and algorithms (for ML)
large sets of data (for DL)
A Basic Model in Machine Learning
Let’s perform the following steps:
1) Start with a simple model (2 variables)
2) Generalize that model (n variables)
3) See how it might apply to a NN
Linear Regression
One of the simplest models in ML
Fits a line (y = m*x + b) to data in 2D
Finds best line by minimizing MSE:
m = average of x values (“mean”)
b also has a closed form solution
Linear Regression in 2D: example
Linear Regression: alternatives
Fitting a polynomial (degree 2, 3, …)
Can lead to overfitting
Polynomials diverge faster than lines
Can reduce predictive accuracy
NB: Linear Regression != Curve Fitting
Linear Regression: example #1
One feature (independent variable):
X = number of square feet
Predicted value (dependent variable):
Y = cost of a house
A very “coarse grained” model
We can devise a much better model
Linear Regression: example #2
Multiple features:
X1 = # of square feet
X2 = # of bedrooms
X3 = # of bathrooms (dependency?)
X4 = age of house
X5 = cost of nearby houses
X6 = corner lot (or not): Boolean
a much better model (6 features)
Linear Multivariate Analysis
General form of multivariate equation:
Y = w1*x1 + w2*x2 + . . . + wn*xn + b
w1, w2, . . . , wn are numeric values
x1, x2, . . . , xn are variables (features)
Properties of variables:
Can be independent (Naïve Bayes)
weak/strong dependencies can exist
Neural Network with 3 Hidden Layers
Neural Networks: equations
Node “values” in first hidden layer:
N1 = w11*x1+w21*x2+…+wn1*xn
N2 = w12*x1+w22*x2+…+wn2*xn
N3 = w13*x1+w23*x2+…+wn3*xn
. . .
Nn = w1n*x1+w2n*x2+…+wnn*xn
Similar equations for other pairs of layers
Neural Networks: Matrices
From inputs to first hidden layer:
Y1 = W1*X + B1 (X/Y1/B1: vectors; W1: matrix)
From first to second hidden layers:
Y2 = W2*X + B2 (X/Y2/B2: vectors; W2: matrix)
From second to third hidden layers:
Y3 = W3*X + B3 (X/Y3/B3: vectors; W3: matrix)
Apply an “activation function” to y values
Neural Networks (general)
Multiple hidden layers:
Layer composition is your decision
Activation functions: sigmoid, tanh, RELU
https://en.wikipedia.org/wiki/Activation_function
Back propagation (1980s)
https://en.wikipedia.org/wiki/Backpropagation
=> Initial weights: small random numbers
Activation Functions (Examples)
import numpy as np
...
# Python sigmoid example:
z = 1/(1 + np.exp(-np.dot(W, x)))
...# Python tanh example:
z = np.tanh(np.dot(W,x));
# Python ReLU example:
z = np.maximum(0, np.dot(W, x))
What’s the “Best” Activation Function?
Initially sigmoid was popular
then tanh became popular
Now RELU is preferred (better results)
NB: sigmoid + tanh are used in LSTMs
Sample Cost Function #1
Sample Cost Function #2
How to Select a Cost Function
1) Depends on the learning type:
=> supervised/unsupervised/RL
2) Depends on the activation function
3) Other factors
Example:
cross-entropy cost function for supervised
learning on multiclass classification
GD versus SGD
SGD (Stochastic Gradient Descent):
+ involves a SUBSET of the dataset
+ aka Minibatch Stochastic Gradient Descent
GD (Gradient Descent):
+ involves the ENTIRE dataset
More details:
http://cs229.stanford.edu/notes/cs229-notes1.pdf
What are Hyper Parameters?
higher level concepts about the model such as
complexity, or capacity to learn
Cannot be learned directly from the data in the
standard model training process
must be predefined
Hyper Parameters (examples)
# of hidden layers in a neural network
the learning rate (in many models)
# of leaves or depth of a tree
# of latent factors in a matrix factorization
# of clusters in a k-means clustering
How Many Layers in a DNN?
Algorithm #1 (from Geoffrey Hinton):
1) add layers until you start overfitting your
training set
2) now add dropout or some another
regularization method
Algorithm #2 (Yoshua Bengio):
"Add layers until the test error does not improve
anymore.”
How Many Hidden Nodes in a DNN?
Based on a relationship between:
# of input and # of output nodes
Amount of training data available
Complexity of the cost function
The training algorithm
Use Cases for Neural Networks
CNNs (Convolutional NNs):
Good for image processing
2000: CNNs processed 10-20% of all checks
=> Approximately 60% of all NNs
RNNs (Recurrent NNs):
Good for NLP and audio
CNN: Sample Filters
CNN Filters (examples)
Types of RNNs
LSTMs (Long Short Term Memory)
GRUs
ResNets (Residual NNs)
Features of LSTMs
Used in Google speech recognition + Alpha Go
input/output/forget gates
they avoid the vanishing gradient problem
Can track 1000s of discrete time steps
Used by international competition winners
Often combined with CTC
Inside an LSTM
Inside an LSTM
Inside an LSTM
Keras/LSTM Code Snippet
import numpy
from keras.datasets import imdb
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers.embeddings import Embedding
from keras.preprocessing import sequence
...
GANs: Generative Adversarial Networks
GANs: Generative Adversarial Networks
Make imperceptible changes to images
Can consistently defeat all NNs
Can have extremely high error rate
Some images create optical illusions
https://www.quora.com/What-are-the-pros-and-cons-of-using-generative-adversarial-networks-a-type-of-neural-network
ML/DL Frameworks
Caffe (templates instead of code)
Theano (influenced TensorFlow)
Tensorflow
TensorFlow Lite (release date?)
Keras (“layer” over Theano+TF)
Tefla (mini framework over TF)
Torch (Lua) + PyTorch (Facebook)
MxNET (Amazon)
CNTK (Microsoft)
Languages for ML/DL
Popular languages for ML:
R (popular among statisticians)
Python (sklearn/pandas/etc)
Popular languages for DL:
Python (Keras/Theano/TF modules)
some Java/C++/Go
“Challenges” in Deep Learning
overfitting/underfitting of a model
vanishing/exploding gradient
learning rate (too high or too low)
Debugging NNs (good luck)
Miscellaneous Topics
* Data versus algorithms:
Option A: good data + average algorithm
Option B: average data + good algorithm
=> Option A is preferred over Option B
• “Cleaning” a dataset:
De-duplicate and fix invalid/missing data (how?)
* Dimensionality reduction:
eliminate “unimportant” features (columns)
Miscellaneous Topics
* XOR requires two hidden layers to solve (why?)
• A dataset whose columns are interchangeable cannot be
solved with a CNN (why?)
• Second generation TPUs
• TensorFlow Lite (open source later in 2017)
www.tensorflow.org/tutorials
D3 Fun Samples
D3 Animation effects:
MouseMoveFadeAnim1Back1.html
SVG tiger:
svg-tiger-d3.svg
D3 and SVG tiger:
svg-tiger-d3.html
Deep Learning Playground
TF playground home page:
http://playground.tensorflow.org
Demo #1:
https://github.com/tadashi-aikawa/typescript-
playground
Converts playground to TypeScript
D3/TypeScript/Deep Learning
Download playground_master.zip
npm install
npm start
Demo converts playground to TypeScript
D3/TypeScript/Deep Learning
TypeScript files in ‘src’ directory:
state.ts
seedrandom.d.ts
playground.ts
linechart.ts
heatmap.ts
dataset.ts
nn.ts (<= activations/nodes in a neural net)
Activations in TypeScript (nn.ts)
export class Activations {
public static TANH: ActivationFunction = {
output: x => (Math as any).tanh(x),
der: x => {
let output = Activations.TANH.output(x);
return 1 - output * output;
} }; public static RELU: ActivationFunction = {
output: x => Math.max(0, x), der: x => x <= 0 ? 0 : 1
};
Activations in TypeScript (nn.ts)
public static SIGMOID: ActivationFunction = {
output: x => 1 / (1 + Math.exp(-x)), der: x => {
let output = Activations.SIGMOID.output(x);
return output * (1 - output);
} }; public static LINEAR: ActivationFunction = {
output: x => x, der: x => 1
}; }
Angular/Deep Learning App (Demo #2)
Create NGDeepLearning via ‘ng’
Copy ./src/*ts files from playground_master into NGDeepLearning/src subdirectory
Merge the two package.json files
Merge the two index.html files
install d3: npm install d3 --save
Angular/Deep Learning
Add import * as d3 from 'd3’; to the files:
dataset.ts
heatmap.ts
linechart.ts
playground.ts
Launch the app: ng serve
Deep Learning and Art/”Stuff”
“Convolutional Blending” images:
=> 19-layer Convolutional Neural Network
www.deepart.io
Bots created their own language:
https://www.recode.net/2017/3/23/14962182/ai-learning-language-open-ai-research
https://www.fastcodesign.com/90124942/this-google-engineer-taught-an-algorithm-to-make-train-footage-and-its-hypnotic
About Me
I provide training for the following:
=> Deep Learning/TensorFlow/Keras
=> Android
=> Angular 4
Recent/Upcoming Books
1) HTML5 Canvas and CSS3 Graphics (2013)
2) jQuery, CSS3, and HTML5 for Mobile (2013)
3) HTML5 Pocket Primer (2013)
4) jQuery Pocket Primer (2013)
5) HTML5 Mobile Pocket Primer (2014)
6) D3 Pocket Primer (2015)
7) Python Pocket Primer (2015)
8) SVG Pocket Primer (2016)
9) CSS3 Pocket Primer (2016)
10) Android Pocket Primer (2017)
11) Angular Pocket Primer (2017)