Module 1

Learning Objectives
You will be able to:

▪ Understand differences between Machine Learning and Deep Learning

▪ Understand Deep Learning concepts

2
AI / Machine Learning / Deep Learning

Artificial Intelligence (AI):

Broader concept of machines Artificial Intelligence
being able to carry out 'smart'
tasks
Machine Learning: Machine Learning
Current application of AI that
machines learn from data using
mathematical, statistical models Deep
Learning
Deep Learning: (Hot!)
Using Neural Networks to solve
some hard problems
Advances in AI

Many of the latest advances in AI are due to Deep Learning

 Self-driving cars
 Image recognition
 Language translation
About Deep Learning

Deep Learning uses Neural Networking techniques

Neural Networks fell out of favor in the ‘90s as statistics-based methods yielded
better results
Now making a comeback due to:
 Big Data: we have so much data to train our algorithms
 Big Data ecosystems: Hadoop*, Spark*, Cloud
 Big Compute: cloud computing
 Advances in hardware: CPU, GPU and TPU
ML vs. DL
Machine Learning Deep Learning

Data size Performs reasonably well on small / medium data Need large amount of data for reasonable
(see next slide for performance
graph)
Scaling Doesn't scale with large amount of data Scales well with large amount of data

Compute power Doesn't need a lot of compute (works well on single Needs a lot of compute power (usually runs on
machines) clusters)
CPU/GPU Mostly CPU bound Can utilize GPU for certain computes (massive
matrix operations)
Feature engineering Features need to be specified manually by experts DL can learn high-level features from data
automatically
Execution time Training usually takes seconds, minutes, hours Training takes lot longer (days)

Interpretability Easy to interpret Hard to understand the final result

Scaling With Data

ML algorithms tend to not

scale well for massive
amount of datasets

See next slides for Jeff

Dean's (Google*) take on
this
Scaling With Data

Now that we have massive amount of data, DL algorithms are outperforming ML

algorithms

Now
A Deep Learning Example : Image Recognition

Here each layer is doing some

computation and sending the
results to the next layer
Subsequent layers can recognize
complex shapes like eyes, noses,
etc.
A Brief History: Modeling After Neurons in Brain

The human brain is a bunch of interconnected neurons

The neuron is like a “gate” – produces an output

12
History, 1943: McCulloch Pitts Neural Model

McCulloch and Pitts defined a simple

model of a neuron
It consisted of N inputs In and N
Weights
Go to a transfer (sum) function, apply a
threshold to an output
Limitations:
 Binary output
 No way to automatically train
weights WN
13
History, 1957: Frank Rosenblatt

Frank Rosenblatt invented the

Perceptron:
- Simplest type of feedforward
neural network

14
Perceptron (Single Layer Perceptron)

The perceptron is a linear model used for binary

classification with a simple input–output
relationship

Mark 1 Perceptron
A Very Simple Perceptron

Here this perceptron

is deciding if I should
go to a concert.
It considers various
inputs (weather, if a
friend will join, etc.)
And different weights
If the final score is >
50, then the answer is
YES
A Very Simple Perceptron

What is the outcome if

 Weather is good
 And a friend can
join?

What is the outcome if

 You can go by
public transit
 And a friend can
join
Perceptron (Generalized)
Perceptron will have multiple
inputs and an output
Terminolog Explanation
y
Xi Input

Wi Weight for the input

x.w

n Number of inputs

b Bias term (does not depend

on input values, shifts
decision boundary from
origin)
Constructing Neural Networks

We add multiple layers

Each layer can have many neurons
'Deep' Neural Networks

If the network has more than

ONE hidden layer, it is called
DEEP network (or more than 3
layers total including input and
output)
More layers gives the network
the ability to adapt to complex
data
However, more layers take
more time to train
Examples of 'Deep' Neural Networks

Image Recognition
 ResNet (from 2015) with 152 layers
A CNN (Convolutional Neural Network) from Microsoft that won ILSRVC 2015
competition
 AlexNet (8 layers)
Won ImageNet Challenge (millions of images in 20,000 categories) in 2012

Machine Translation
 Google Translate* uses Deep Neural Networks (DNN) to translate between
languages (English to French, etc.) with very high accuracy
Examples of Deep Neural Networks

Reinforcement Learning
Bots learning to play games automatically (at superhuman levels!)

Demo1 : Deep Mind playing Atari* Breakout

Demo2 : Deep Mind AlphaGo*
Demo3 : Open AI bots playing Dota 2* and beating pro human teams!
Sizing Layers

Each Layer can have many neurons

How do we decide the number of neurons per layer?
Let's start with an example:
 Assume our input has 3 dimensions / Features (A,B,C)
 And we are classifying them into 2 classes X & Y
Sizing Layers

Input and Output are straight

forward
Input layer is sized to match
number of features
 in this case 3, for inputs A,B,C
Output layer is sized to match
number of output classes
 2 for outputs X,Y
Sizing Hidden Layers

How about Hidden Layer sizing? How

many neurons in Hidden Layer?
This is a 'hyper parameter' for the
network
We can try a few choices and evaluate
the results
Note: Large number of neurons can
'overfit' the data
Activation Function

 In this example, the

activation function is a
pretty simple 'comparison
function’
- If the (dot product) sum
of weight and input < 50 ,
output is NO
- If the sum is >= 50,
output is YES
- So we are converting a
simple number input
(score) into a binary
YES/NO answer
Activation Function

 Activation functions
propagate the output
of one layer’s nodes
forward to the next
layer
 Allows us to create
different output values
 Allows us to model
non-linear functions
Activation Function

Input sum = Σ (Wi . Xi) + bias term

Output = activation_function (input sum)

= g (input_sum)
One liner

Output = g (Σ (Wi . Xi) + b)

Types of Activation Functions

The following Activation Functions are in use:

 Linear
 SIGMOID*
 Tanh
 Rectified Linear Unit (ReLU) (and its variants)
Activation Function: Linear

Very simple
Used for Linear Regression
Output = weight* Input +
Intercept
Y = aX + b
Activation Function : SIGMOID*

Comes from Logistic Regression

It reduces the input values to an
output of 0 to 1 (probability)
SIGMOID* was the oldest / first
Activation Function used
However, now it has been eclipsed by
other Activation Functions that
produce better results
Activation Function : Tanh

Tanh function ranges from -1 to +1

(SIGMOID* function is between 0 and
+1)
So Tanh function can better deal with
negative numbers compared to
Sigmoid
Activation Function : Rectified Linear

ReLU is a very simple and effective

Activation Function
If input is above zero, it passes it on
If input goes below zero, it is clipped
at zero
ReLU is the current state of the art, as
it has been proven to work in many
different datasets
ReLu is also very fast to compute!
SIGMOID* / Tanh

SIGMOID* and Tanh both suffer from the Vanishing Gradient Problem:
 The derivative of a Sigmoid is less than .25
 As we propagate that through many layers, that gradient becomes smaller
and smaller and eventually 'vanishes' (too small)
ReLU to Rescue!

ReLU doesn't suffer from

 Vanishing Gradient Problem
 Exploding Gradient Problem
Here, simpler is actually better
Activation Function : Leaky ReLU

 Leaky ReLU fixes an inherent

problem with ReLU
- When input dips below zero, ReLU
produces ZERO
- This is 'dying ReLU' problem
 When input is below zero, Leaky
ReLU will produce a 'small signal'
Activation Function: SoftPlus

SoftPlus is 'smooth version of

ReLU'
Activation Function: SoftMax*

 SoftMax* is a generalized Logistic Regression

- Logistic Regression can do binary outcome (0 / 1)
- SoftMax* gives out probabilities for multiple classes

digit 0 1 2 3 4 5 6 7 8 9
(outcome)
Probability 0.9 0 0 0 0 0 0 0 0 0.1
Activation Functions - Comparison
Loss Functions

Loss functions quantify how close a

given Neural Network is to the truth
How much the predictions deviate from
expected outputs
 Convert this difference from Vectors
to a 'single number'
Goal is to minimize Loss Function
 Zero would mean the model is
'perfect'
 It is hard to achieve in practice, but
we can get pretty close to zero
42
Loss Functions For Regression: Mean Squared Error
(MSE) Loss Function
Very much like 'Ordinary Least Squares' in Linear Regression
Pros
 Very simple to understand
 Fast to compute
Cons
 Can be prone to outliers

43
Loss Functions For Regression: Mean Absolute Error Loss
(MAE)
Averages 'absolute error' across all data points

44
Loss Functions for Regressions Takeaway

Both 'Mean Squared Error' (MSE) and 'Mean Absolute Error' (MAE) are used
widely
And they perform pretty well on most scenarios
If the inputs span a large range (X1 = 1 to 10, X2 = 1000 to 1000000), then
consider normalizing them first
Other options to consider:
 Mean squared log error loss (MSLE)
 Mean absolute percentage error loss (MAPE)

45
Loss Function for Classification

 Classifiers segment data into 'buckets' (fraud / not-fraud, spam/not-spam)

 Most of the time, the classifier gives a probability (80% spam, 20% not spam)
 Loss Functions:
- Hinge Loss
- Logistics Loss

46
Loss Functions for Classification: Hinge Loss

 Hinge Loss is used heavily when the network does hard binary classification
(0 or 1)
 Can be extended for doing 'multiclass classification’

47
Loss Function for Classification : Logistic Loss

 Logistics Loss is preferred when we get probabilities instead of hard

classifications
 For example, model says 70% spam, 30% not spam, instead of giving a 0 or
1
 Last Layer uses 'SoftMax*' activation function

digit 0 1 2 3 4 5 6 7 8 9
(outcome)
Probability 0.75 0 0 0 0 0 0 0 0.15 0.10

48
Hyperparameters

• Models and networks have parameters that we adjust during optimization

- Goal is to minimize error and maximize performance
• Hyperparameters
- Learning rate
- Regularization
Hyperparameter : Learning Rate (alpha, 𝛂 )

 Learning rate controls the amount by which we

adjust parameters during optimization
 It is a coefficient that scales the size of steps
 During back propagation we multiply error
gradient by learning rate
Hyperparameter : Learning Rate (alpha, 𝛂 )

 Larger learning rate (close to 1.0) means:

- Bigger steps and less number of iterations
- But can overshoot and miss the minimum
 Smaller learning rate (0.0001) means:
- Smaller leaps, and lots of iterations
- Can find the minimum, but can take a long time
 What is the 'perfect’ learning rate?
- Experiment (it varies for each dataset and
computation)
Hyperparameters : Regularization (lambda, ƛ)

 Regularization helps avoid 'over fitting'

 It does this by controlling coefficients so they don't get too large
 L2 (Ridge) Regularization
- Shrinks coefficients to avoid over fitting
 L1 (Lasso) Regularization
- Shrinks coefficients like L2, but can also make them zero
- Effectively removing the variable from consideration (Variable Selection)
Summary

We learned about:
 Deep Learning vs. Machine Learning
 How Neural Nets have evolved
 Activation functions
 Loss functions
 Hyperparameter tuning
Recommended Resources

 Intel® AI Academy : https://software.intel.com/ai-academy

 https://software.intel.com/en-us/ai-academy/basics

 "Getting Started with Deep Learning"

by Josh Patterson, Adam Gibson
Link1

56