Introduction to Deep Learning Models

Tanujit Chakraborty
@ Sorbonne
Webpage: https://www.ctanujit.org
Lecture Outline
• Machine learning basics
 Supervised and unsupervised learning
 Linear and non-linear classification methods
• Introduction to deep learning
• Elements of neural networks (NNs)
 Activation functions
• Training NNs
 Gradient descent
 Regularization methods
• NN architectures
 Convolutional NNs
 Recurrent NNs
 LSTM
 Transformers
 GAN

* Most of the slides are adapted from DL and D2L Book and several online resources and slides that are acknowledged at the end of the presentation. 2
Machine Learning Basics

• Artificial Intelligence is a scientific field concerned with the development of algorithms that allow computers to
learn without being explicitly programmed
• Machine Learning is a branch of Artificial Intelligence, which focuses on methods that learn from data and make
predictions on unseen data

Labeled Data Machine Learning algorithm

Training
Prediction

Labeled Data Learned model Prediction

3
Machine Learning Types

• Supervised: learning with labeled data

 Example: email classification, image classification
 Example: regression for predicting real-valued outputs

• Unsupervised: discover patterns in unlabeled data

 Example: cluster similar data points

• Reinforcement learning: learn to act based on feedback/reward

 Example: learn to play Go

class A

class B

Regression Clustering
Classification

4
Supervised and Unsupervised Learning

• Supervised learning categories and techniques • Unsupervised learning categories and techniques
 Numerical classifier functions  Clustering
o Linear classifier, perceptron, logistic regression, support o k-means clustering
vector machines (SVM), neural networks o Mean-shift clustering
 Parametric (probabilistic) functions o Spectral clustering
o Naïve Bayes, Gaussian discriminant analysis (GDA),  Density estimation
hidden Markov models (HMM), probabilistic graphical
models o Gaussian mixture model (GMM)
 Non-parametric (instance-based) functions o Graphical models
o k-nearest neighbors, kernel regression, kernel density  Dimensionality reduction
estimation, local regression o Principal component analysis (PCA)
 Symbolic functions o Factor analysis
o Decision trees, classification and regression trees (CART)
 Aggregation (ensemble) learning
o Bagging, boosting (Adaboost), random forest

5
Machine Learning Timeline
Nearest Neighbor Classifier

• Nearest Neighbor – for each test data point, assign the class label of the nearest training data point
 Adopt a distance function to find the nearest neighbor
o Calculate the distance to each data point in the training set, and assign the class of the nearest data point (minimum distance)
 It does not require learning a set of weights

Test Training
Training example examples
examples from class
from class 2
1

7
Nearest Neighbor Classifier

• For image classification, the distance between all pixels is calculated (e.g., using ℓ1 norm, or ℓ2 norm)
 Accuracy on CIFAR-10: 38.6%

• Disadvantages:
 The classifier must remember all training data and store it for future comparisons with the test data
 Classifying a test image is expensive since it requires a comparison to all training images

ℓ1 norm
(Manhattan distance)

Picture from: https://cs231n.github.io/classification/ 8

k-Nearest Neighbors Classifier

• k-Nearest Neighbors approach considers multiple neighboring data points to classify a test data point
 E.g., 3-nearest neighbors
o The test example in the figure is the + mark
o The class of the test example is obtained by voting (based on the distance to the 3 closest points)

x2
x
x
x o
x x
x x
+ o
o x
o + x
o
o o
o
o

x1

9
Linear Classifier

• Linear classifier
 Find a linear function f of the inputs xi that separates the classes
𝑓 𝑥𝑖 , 𝑊, 𝑏 = 𝑊𝑥𝑖 + 𝑏

 Use pairs of inputs and labels to find the weights matrix W and the bias vector b
o The weights and biases are the parameters of the function f
 Several methods have been used to find the optimal set of parameters of a linear classifier
o A common method of choice is the Perceptron algorithm, where the parameters are updated until a minimal error is reached (single
layer, does not use backpropagation)
 Linear classifier is a simple approach, but it is a building block of advanced classification algorithms, such as SVM and
neural networks
o Earlier multi-layer neural networks were referred to as multi-layer perceptrons (MLPs)

10
Linear Classifier
• The decision boundary is linear
 A straight line in 2D, a flat plane in 3D, a hyperplane in 3D and
higher dimensional space
• Example: classify an input image
 The selected parameters in this example are not good, because the
predicted cat score is low

Picture from: https://cs231n.github.io/classification/ 11

Support Vector Machines
• Support vector machines (SVM)
 How to find the best decision boundary?
o All lines in the figure correctly separate the 2 classes
o The line that is farthest from all training examples
will have better generalization capabilities
 SVM solves an optimization problem:
o First, identify a decision boundary that correctly
classifies the examples

o Next, increase the geometric margin between the

boundary and all examples
 The data points that define the maximum
margin width are called support vectors
 Find W and b by solving:

12
Linear vs Non-linear Techniques

• Linear classification techniques • For some tasks, input data

 Linear classifier can be linearly separable,
 Perceptron and linear classifiers can be
suitably applied
 Logistic regression
 Linear SVM
 Naïve Bayes

• Non-linear classification techniques • For other tasks, linear

 k-nearest neighbors classifiers may have
 Non-linear SVM difficulties to produce
 Neural networks adequate decision
 Decision trees
boundaries
 Random forest

13
Non-linear Techniques

• Non-linear classification
 Features 𝑧𝑖 are obtained as non-linear functions of the inputs 𝑥𝑖
 It results in non-linear decision boundaries
 Can deal with non-linearly separable data

Inputs: 𝑥𝑖 = 𝑥𝑛1 𝑥𝑛2

2 2
Features: 𝑧𝑖 = 𝑥𝑛1 𝑥𝑛2 𝑥𝑛1 ∙ 𝑥𝑛2 𝑥𝑛1 𝑥𝑛2

Outputs: 𝑓 𝑥𝑖 , 𝑊, 𝑏 = 𝑊𝑧𝑖 + 𝑏
14
Non-linear Support Vector Machines

• Non-linear SVM
 The original input space is mapped to a higher-dimensional feature space where the training set is linearly separable
 Define a non-linear kernel function to calculate a non-linear decision boundary in the original feature space

Φ: 𝑥 ↦ 𝜙 𝑥

15
Binary vs Multi-class Classification
• A classification problem with only 2 classes is
referred to as binary classification
 The output labels are 0 or 1
 E.g., benign or malignant tumor, spam or no-spam
email
• A problem with 3 or more classes is referred to as
multi-class classification

• Both the binary and multi-class classification

problems can be linearly or non-linearly
separated

 Figure: linearly and non-linearly separated data

for binary classification problem

16
Computer Vision Tasks

• Computer vision has been the primary area of interest for ML

• The tasks include: classification, localization, object detection, instance segmentation

17
No-Free-Lunch Theorem
• Wolpert (2002) - The Supervised Learning No-Free-Lunch Theorems
• The derived classification models for supervised learning are simplifications of the reality
 The simplifications are based on certain assumptions
 The assumptions fail in some situations
o E.g., due to inability to perfectly estimate ML model parameters from limited data

• In summary, No-Free-Lunch Theorem states:

 No single classifier works the best for all possible problems
 Since we need to make assumptions to generalize

Why is DL Useful?
• DL provides a flexible, learnable framework for representing visual, text, linguistic information
 Can learn in supervised and unsupervised manner

• DL represents an effective end-to-end learning system

• Requires large amounts of training data
• Since about 2010, DL has outperformed other ML techniques
 First in vision and speech, then NLP, and other applications
18
ML vs. Deep Learning
• Deep learning (DL) is a machine learning subfield that uses multiple layers for learning data representations
 DL is exceptionally effective at learning patterns

• Conventional machine learning methods rely on human-designed feature representations

 ML becomes just optimizing weights to best make a final prediction

Picture from: https://www.xenonstack.com/blog/static/public/uploads/media/machine-learning-vs-deep-learning.png 19

ML vs. Deep Learning

• DL applies a multi-layer process for learning rich hierarchical features (i.e., data representations)
 Input image pixels → Edges → Textures → Parts → Objects

Low- Mid- High- Trainabl

Output
Level Level Level e
Features Features Features Classifie
r

20
Deep Learning Timeline
Representational Power
• NNs with at least one hidden layer are universal approximators
 Given any continuous function h(x) and some 𝜖 > 0, there exists a NN with one
hidden layer (and with a reasonable choice of non-linearity) described with the
function f(x), such that ∀𝑥, ℎ 𝑥 − 𝑓(𝑥) < 𝜖
 I.e., NN can approximate any arbitrary complex continuous function

• NNs use nonlinear mapping of the inputs x to the

outputs f(x) to compute complex decision boundaries
• But then, why use deeper NNs?
 The fact that deep NNs work better is an empirical
observation
 Mathematically, deep NNs have the same
representational power as a one-layer NN

22
Introduction to Neural Networks

• Handwritten digit recognition (MNIST dataset)

 The intensity of each pixel is considered an input element
 Output is the class of the digit
Input Output

x1 y1
0.1 is 1
x2
y2
0.7 is 2
The image is “2”
…
…

…
…
…
…
x256 y1
0.2 is 0
16 x 16 = 256
0
Ink → 1 Each dimension represents
the confidence of a digit
No ink →0
23
Introduction to Neural Networks

• Handwritten digit recognition

x1 y1
x2
y2
… Machine “2”
…

…
…
x256 𝑓: 𝑅256 → 𝑅10 y1
0

The function 𝒇 is represented by a neural network

24
Elements of Neural Networks

• NNs consist of hidden layers with neurons (i.e., computational units)

• A single neuron maps a set of inputs into an output number, or 𝑓: 𝑅 𝐾 → 𝑅

z  a1w1  a2 w2    aK wK  b
a1 w1
𝑎=𝜎 𝑧
a2 w2
z  z 

…

a
wK
…

output
aK Activation
weights function
b
input
bias

25
Elements of Neural Networks

• A NN with one hidden layer and one output layer

Weights Biases

𝒉𝒊𝒅𝒅𝒆𝒏 𝒍𝒂𝒚𝒆𝒓 𝒉 = 𝝈(𝐖𝟏 𝒙 + 𝒃𝟏 )

𝒐𝒖𝒕𝒑𝒖𝒕 𝒍𝒂𝒚𝒆𝒓 𝒚 = 𝝈(𝑾𝟐 𝒉 + 𝒃𝟐 )

Activation functions

4 + 2 = 6 neurons (not counting inputs)

[3 × 4] + [4 × 2] = 20 weights
𝒚 4 + 2 = 6 biases
26 learnable parameters

𝒉
26
Elements of Neural Networks
• A neural network playground link

27
Elements of Neural Networks
• Deep NNs have many hidden layers
 Fully-connected (dense) layers (a.k.a. Multi-Layer Perceptron or MLP)
 Each neuron is connected to all neurons in the succeeding layer

Input Layer Layer Layer Output

x1 1 2 … L y1
…
x2 … y2
…

…
…
…
…

…
…

…
…

…
…
xN … yM
…
Input Output
Layer Hidden Layer
Layers 28
Elements of Neural Networks

• A simple network, toy example

1 ∙ 1 + −1 ∙ −2 + 1 = 4

0.98 Sigmoid Function

1 4
1
-2
 z  
1
1 1  ez
 z 
-1 -2 0.12
-1
1 z
0

1 ∙ −1 + −1 ∙ 1 + 0 =-2
29
Elements of Neural Networks

• A simple network, toy example (cont’d)

 For an input vector [1 −1]𝑇 , the output is [0.62 0.83]𝑇

1 4 0.98 2 0.86 3 0.62

1
-2 -1 -1
1 0 -2
-1 -2 0.12 -2 0.11 -1 0.83
-1
1 -1 4
0 0 2

1 0.62
𝑓: 𝑅2 → 𝑅2 𝑓 =
−1 0.83
30
Matrix Operation

• Matrix operations are helpful when working with multidimensional inputs and outputs

1 4 0.98
1 𝜎 W x + b = a
-2
1
1 −2 1 1 0.98
-1 -2 0.12 𝜎 + =
−1 1 −1 0 0.12
-1
1
4
0
−2

31
Matrix Operation

• Multilayer NN, matrix calculations for the first layer

 Input vector x, weights matrix W1, bias vector b1, output vector a1

x1 … y1
…
x2 W1 … y2
b1 …

…
…
…
…

…
…

…
…

…
…
xN x a1 … yM
…

a1 = 𝜎 W1 x + b1

32
Matrix Operation
• Multilayer NN, matrix calculations for all layers

x1 … y1
…
x2 W1 W2 …
WL y2
b L
b1 b2 …

…
…
…
…

…
…

…
…

…
…
a2… yM
xN x a1 y
…

𝜎 W1 x + b1
𝜎 W2 a1 + b2
𝜎 WL aL-1 + bL
33
Matrix Operation
• Multilayer NN, function f maps inputs x to outputs y, i.e., 𝑦 = 𝑓(𝑥)

x1 … y1
…
x2 W1 W2 …
WL y2
b L
b1 b2 …

…
…
…
…

…
…

…
…

…
…
xN x
a1 a2
… y yM
…

y =𝑓 x =𝜎 WL …𝜎 W2 𝜎 W1 x + b1 + b2 … + bL

34
Softmax Layer
• In multi-class classification tasks, the output layer is typically a softmax layer
 I.e., it employs a softmax activation function
 If a layer with a sigmoid activation function is used as the output layer instead, the predictions by the NN may not be
easy to interpret
o Note that an output layer with sigmoid activations can still be used for binary classification

A Layer with Sigmoid Activations

3
z1
0.95
 y1   z1 
z2
1
 0.73
 
y2   z 2

z3
-
3

0.05
 
y3   z 3

35
Softmax Layer

• The softmax layer applies softmax activations to output a probability Probability:

value in the range [0, 1]
 The values z inputted to the softmax layer are referred to as logits
 0 < 𝑦𝑖 < 1
 𝑖 𝑦𝑖 = 1
A Softmax Layer
0.88 3
 e
3 20
y1  e
z1 z1 zj
z1 e e
j 1
0.12 3
z2 1
e e z2 2.7  y2  e z2
e
zj

j 1
0.05 ≈0

3
z3 -3
e
z3
e y3  e z3 zj
e
3 j 1

 e
zj

j 1 20.75
36
Activation Functions

• Non-linear activations are needed to learn complex (non-linear) data representations

 Otherwise, NNs would be just a linear function (such as W1 W2 𝑥 = 𝑊𝑥)
 NNs with large number of layers (and neurons) can approximate more complex functions
o Figure: more neurons improve representation (but, may overfit)

37
Activation: Sigmoid

• Sigmoid function σ: takes a real-valued number and “squashes” it into the range between 0 and 1
 The output can be interpreted as the firing rate of a biological neuron
o Not firing = 0; Fully firing = 1
 When the neuron’s activation are 0 or 1, sigmoid neurons saturate
o Gradients at these regions are almost zero (almost no signal will flow)
 Sigmoid activations are less common in modern NNs

𝑓 𝑥 ℝ𝑛 → 0,1

𝑥
38
Activation: Tanh

• Tanh function: takes a real-valued number and “squashes” it into range between -1 and 1
 Like sigmoid, tanh neurons saturate
 Unlike sigmoid, the output is zero-centered
o It is therefore preferred than sigmoid
 Tanh is a scaled sigmoid: tanh(𝑥) = 2 ∙ 𝜎(2𝑥) − 1

𝑓 𝑥 ℝ𝑛 → −1,1

39
Activation: ReLU

• ReLU (Rectified Linear Unit): takes a real-valued number and thresholds it at zero 𝑓 𝑥 = max(0, 𝑥)

ℝ𝑛 → ℝ𝑛+
 Most modern deep NNs use ReLU activations
 ReLU is fast to compute
o Compared to sigmoid, tanh 𝑓 𝑥
o Simply threshold a matrix at zero
 Accelerates the convergence of gradient descent
o Due to linear, non-saturating form
 Prevents the gradient vanishing problem

40
Activation: Leaky ReLU

• The problem of ReLU activations: they can “die”

 ReLU could cause weights to update in a way that the gradients can become zero and the neuron will not activate again
on any data
 E.g., when a large learning rate is used

• Leaky ReLU activation function is a variant of ReLU

 Instead of the function being 0 when 𝑥 < 0, a leaky ReLU has a small negative slope (e.g., α = 0.01, or similar)

 This resolves the dying ReLU problem 𝛼𝑥 for 𝑥 < 0

𝑓 𝑥 =
 Most current works still use ReLU 𝑥 for 𝑥 ≫ 0
o With a proper setting of the learning
rate, the problem of dying ReLU can
be avoided

41
Activation: Linear Function
• Linear function means that the output signal is proportional to the input signal to the neuron

ℝ𝑛 → ℝ𝑛
 If the value of the constant c is 1, it is also called identity
activation function 𝑓 𝑥 = 𝑐𝑥
 This activation type is used in regression problems
o E.g., the last layer can have linear activation function, in
order to output a real number (and not a class
membership)

42
Training NNs

• The network parameters 𝜃 include the weight matrices and bias vectors from all layers

𝜃 = 𝑊 1 , 𝑏1 , 𝑊 2 , 𝑏 2 , ⋯ 𝑊 𝐿 , 𝑏 𝐿
 Often, the model parameters 𝜃 are referred to as weights

• Training a model to learn a set of parameters 𝜃 that are optimal (according to a criterion) is one of the greatest
challenges in ML

x1 … y1
0.1 is 1
…

Softmax
x2 … y2
0.7 is 2
…
…
…

…
…

…
…
x256 … y1
0.2 is 0
16 x 16 = 256
… 0
Slide credit: Hung-yi Lee – Deep Learning Tutorial 43
Training NNs
• Data preprocessing – helps convergence during training
 Mean subtraction, to obtain zero-centered data
o Subtract the mean for each individual data dimension (feature)
 Standardization
o In zero-centered data, divide each feature by its standard deviation
• To obtain standard deviation of 1 and mean of 0 for each data dimension (feature)
 Normalization
o Scale the data within the range [0,1] or [-1, 1]
• E.g., image pixel intensities are divided by 255 to be scaled in the [0,1] range

Picture from: https://cs231n.github.io/neural-networks-2/ 44

Training NNs
• To train a NN, set the parameters 𝜃 such that for a training subset of images, the corresponding elements in the
predicted output have maximum values

Input: y1 has the maximum value

Input: y2 has the maximum value

.
.
.

Input: y9 has the maximum value

Input: y10 has the maximum value

45
Training NNs
• Define a loss function/objective function/cost function ℒ 𝜃 that calculates the difference (error) between the
model prediction and the true label
 E.g., ℒ 𝜃 can be mean-squared error, cross-entropy, etc.

x1 … y1 0.2 1
…
x2 … y2 0.3 0
… Cost
…
…

…
…
…
…

…
…
…
…
…
…
x256 … y10 0.5 ℒ(𝜃) 0
…
True label “1”

46
Training NNs

𝑁
• For a training set of 𝑁 images, calculate the total loss overall all images: ℒ 𝜃 = 𝑛=1 ℒ 𝑛 𝜃
• Find the optimal parameters 𝜃 ∗ that minimize the total loss ℒ 𝜃

ℒ1 𝜃
x1 NN 𝑦1 y1

ℒ2 𝜃
x2 NN 𝑦2 y
2

ℒ3 𝜃
x3 3
NN 𝑦 y3
…
…
…
…

…
…
…
…
ℒ𝑛 𝜃
xN NN 𝑦𝑁 yN
47
Loss Functions

• Classification tasks

Training Pairs of 𝑁 inputs 𝑥𝑖 and ground-truth class labels 𝑦𝑖

examples

Output Softmax Activations

Layer [maps to a probability distribution]

𝑁 𝐾
1 (𝑖) (𝑖) (𝑖)
Loss function Cross-entropy ℒ 𝜃 =− 𝑦𝑘 log 𝑦𝑘 + 1 − 𝑦𝑘 log 1 − 𝑦𝑘𝑖
𝑁
𝑖=1 𝑘=1

Ground-truth class labels 𝑦𝑖 and model predicted class labels 𝑦𝑖

48
Loss Functions

• Regression tasks

Training Pairs of 𝑁 inputs 𝑥𝑖 and ground-truth output values 𝑦𝑖

examples

Output
Layer Linear (Identity) or Sigmoid Activation

𝑛
1 2
Mean Squared Error ℒ 𝜃 = 𝑦 (𝑖) − 𝑦 (𝑖)
𝑛
𝑖=1
Loss function
𝑛
1
Mean Absolute Error ℒ 𝜃 = 𝑦 (𝑖) − 𝑦 (𝑖)
𝑛
𝑖=1

49
Training NNs
• Optimizing the loss function ℒ 𝜃
 Almost all DL models these days are trained with a variant of the gradient descent (GD) algorithm
 GD applies iterative refinement of the network parameters 𝜃
 GD uses the opposite direction of the gradient of the loss with respect to the NN parameters (i.e.,𝛻ℒ 𝜃 = 𝜕ℒ 𝜕𝜃𝑖 )
for updating 𝜃
o The gradient of the loss function 𝛻ℒ 𝜃 gives the direction of fastest increase of the loss function ℒ 𝜃 when the parameters 𝜃
are changed

ℒ 𝜃
𝜕ℒ
𝜕𝜃𝑖

𝜃𝑖

50
Training NNs

• The loss functions for most DL tasks are defined over very high-dimensional spaces
 E.g., ResNet50 NN has about 23 million parameters
 This makes the loss function impossible to visualize

• We can still gain intuitions by studying 1-dimensional and 2-dimensional examples of loss functions

1D loss (the minimum point is obvious) 2D loss (blue = low loss, red = high
loss)
Picture from: https://cs231n.github.io/optimization-1/ 51
Gradient Descent Algorithm
• Steps in the gradient descent algorithm:
1. Randomly initialize the model parameters, 𝜃 0
2. Compute the gradient of the loss function at the initial parameters 𝜃 0 : 𝛻ℒ 𝜃 0
3. Update the parameters as: 𝜃 𝑛𝑒𝑤 = 𝜃 0 − 𝛼𝛻ℒ 𝜃 0
o Where α is the learning rate
4. Go to step 2 and repeat (until a terminating criterion is reached)

𝜕ℒ
Loss ℒ Initial Gradient 𝛻ℒ =
parameters 𝜃
𝜕

𝜃0

Parameter update: 𝜃 𝑛𝑒𝑤 = 𝜃 − 𝛼𝛻ℒ 𝜃 0

Global loss minimum ℒ𝑚𝑖𝑛

Parameters 𝜃
52
Gradient Descent Algorithm
• Example: a NN with only 2 parameters 𝑤1 and 𝑤2 , i.e., 𝜃 = 𝑤1 , 𝑤2
 The different colors represent the values of the loss (minimum loss 𝜃 ∗ is ≈ 1.3)

1. Randomly pick a
starting point 𝜃 0
2. Compute the gradient
at 𝜃 0 , 𝛻ℒ 𝜃 0
𝜃∗
𝑤2 3. Times the learning
𝜃1 rate 𝜂, and update 𝜃,
𝜃 𝑛𝑒𝑤 = 𝜃 0 − 𝛼𝛻ℒ 𝜃 0
𝜃1 = 𝜃 0 − 𝛼𝛻ℒ 𝜃 0
−𝛻ℒ 𝜃 0 4. Go to step 2, repeat
𝜃0
0 𝜕ℒ 𝜃 0 /𝜕𝑤1
𝛻ℒ 𝜃 =
𝑤1 𝜕ℒ 𝜃 0 /𝜕𝑤2
53
Gradient Descent Algorithm
• Example (contd.)

Eventually, we would reach a minimum …..

2. Compute the gradient at

𝜃 𝑜𝑙𝑑 , 𝛻ℒ 𝜃 𝑜𝑙𝑑

𝜃2
3. Times the learning rate 𝜂,
1 1
𝜃 − 𝛼𝛻ℒ 𝜃 and update 𝜃,
𝑤2 𝜃 2 − 𝛼𝛻ℒ 𝜃 2
𝜃 𝑛𝑒𝑤 = 𝜃 𝑜𝑙𝑑 − 𝛼𝛻ℒ 𝜃 𝑜𝑙𝑑
𝜃1
4. Go to step 2, repeat

𝜃0

𝑤1 54
Gradient Descent Algorithm

• Gradient descent algorithm stops when a local minimum of the loss surface is
reached
 GD does not guarantee reaching a global minimum
 However, empirical evidence suggests that GD works well for NNs

ℒ 𝜃

Picture from: https://blog.paperspace.com/intro-to-optimization-in-deep-learning-gradient-descent/ 55

Gradient Descent Algorithm

• For most tasks, the loss surface ℒ 𝜃 is highly complex (and non-convex)

• Random initialization in NNs results in different

0
ℒ
initial parameters 𝜃 every time the NN is trained
 Gradient descent may reach different minima at
every run
 Therefore, NN will produce different predicted
outputs
• In addition, currently we don’t have algorithms
that guarantee reaching a global minimum for an
arbitrary loss function
𝑤1 𝑤2

56
Backpropagation
• Modern NNs employ the backpropagation method for calculating the gradients of the loss function 𝛻ℒ 𝜃 =
𝜕ℒ 𝜕𝜃𝑖
 Backpropagation is short for “backward propagation”

• For training NNs, forward propagation (forward pass) refers to passing the inputs 𝑥 through the hidden layers
to obtain the model outputs (predictions) 𝑦
 The loss ℒ 𝑦, 𝑦 function is then calculated
 Backpropagation traverses the network in reverse order, from the outputs 𝑦 backward toward the inputs 𝑥 to calculate
the gradients of the loss 𝛻ℒ 𝜃
 The chain rule is used for calculating the partial derivatives of the loss function with respect to the parameters 𝜃 in the
different layers in the network
• Each update of the model parameters 𝜃 during training takes one forward and one backward pass (e.g., of a
batch of inputs)
• Automatic calculation of the gradients (automatic differentiation) is available in all current deep learning
libraries
 It significantly simplifies the implementation of deep learning algorithms, since it obviates deriving the partial
derivatives of the loss function by hand

57
Mini-batch Gradient Descent

• It is wasteful to compute the loss over the entire training dataset to perform a single parameter update for large
datasets
 E.g., ImageNet has 14M images
 Therefore, GD (a.k.a. vanilla GD) is almost always replaced with mini-batch GD

• Mini-batch gradient descent

 Approach:
o Compute the loss ℒ 𝜃 on a mini-batch of images, update the parameters 𝜃, and repeat until all images are used
o At the next epoch, shuffle the training data, and repeat the above process
 Mini-batch GD results in much faster training
 Typical mini-batch size: 32 to 256 images
 It works because the gradient from a mini-batch is a good approximation of the gradient from the entire training set

58
Stochastic Gradient Descent

• Stochastic gradient descent

 SGD uses mini-batches that consist of a single input example
o E.g., one image mini-batch
 Although this method is very fast, it may cause significant fluctuations in the loss function
o Therefore, it is less commonly used, and mini-batch GD is preferred
 In most DL libraries, SGD typically means a mini-batch GD (with an option to add momentum)

59
Problems with Gradient Descent

• Besides the local minima problem, the GD algorithm can be very slow at plateaus, and it can get stuck at
saddle points
cost ℒ 𝜃

Very slow at the plateau

Stuck at a saddle point

Stuck at a local minimum

𝛻ℒ 𝜃 ≈ 0 𝛻ℒ 𝜃
=0 𝛻ℒ 𝜃 = 0
𝜃
60
Gradient Descent with Momentum
• Gradient descent with momentum uses the momentum of the gradient for parameter optimization

cost ℒ 𝜃
Movement = Negative of Gradient + Momentum

Negative of Gradient
Momentum
Real Movement

𝜃
Gradient = 0
61
Gradient Descent with Momentum
• Parameters update in GD with momentum at iteration 𝑡: 𝜃 𝑡 = 𝜃 𝑡−1 − 𝑉 𝑡
o Where: 𝑉 𝑡 = 𝛽𝑉 𝑡−1 + 𝛼𝛻ℒ 𝜃𝑡−1
o I.e., 𝜃𝑡 = 𝜃𝑡−1 − 𝛼𝛻ℒ 𝜃 𝑡−1 − 𝛽𝑉 𝑡−1
• Compare to vanilla GD: 𝜃 𝑡 = 𝜃 𝑡−1 − 𝛼𝛻ℒ 𝜃 𝑡−1
 Where 𝜃 𝑡−1 are the parameters from the previous iteration 𝑡 − 1
• The term 𝑉 𝑡 is called momentum
 This term accumulates the gradients from the past several steps, i.e.,
𝑉 𝑡 = 𝛽𝑉 𝑡−1 + 𝛼𝛻ℒ 𝜃 𝑡−1
= 𝛽 𝛽𝑉 𝑡−2 + 𝛼𝛻ℒ 𝜃 𝑡−2 + 𝛼𝛻ℒ 𝜃 𝑡−1
= 𝛽2 𝑉 𝑡−2 + 𝛽𝛼𝛻ℒ 𝜃 𝑡−2 + 𝛼𝛻ℒ 𝜃 𝑡−1
3 𝑡−3
=𝛽 𝑉 + 𝛽 2 𝛼𝛻ℒ 𝜃 𝑡−3 + 𝛽𝛼𝛻ℒ 𝜃 𝑡−2 + 𝛼𝛻ℒ 𝜃 𝑡−1
 This term is analogous to a momentum of a heavy ball rolling down the hill

• The parameter 𝛽 is referred to as a coefficient of momentum

 A typical value of the parameter 𝛽 is 0.9

• This method updates the parameters 𝜃 in the direction of the weighted average of the past gradients

62
Nesterov Accelerated Momentum

• Gradient descent with Nesterov accelerated momentum

 Parameter update: 𝜃 𝑡 = 𝜃 𝑡−1 − 𝑉 𝑡
o Where: 𝑉 𝑡 = 𝛽𝑉 𝑡−1 + 𝛼𝛻ℒ 𝜃 𝑡−1 + 𝛽𝑉 𝑡−1
 The term 𝜃 𝑡−1 + 𝛽𝑉 𝑡−1 allows to predict the position of the parameters in the next step (i.e., 𝜃 𝑡 ≈ 𝜃 𝑡−1 + 𝛽𝑉 𝑡−1 )
 The gradient is calculated with respect to the approximate future position of the parameters in the next iteration, 𝜃 𝑡 ,
calculated at iteration 𝑡 − 1

GD with Nesterov
GD with momentum
momentum

63
Adam

• Adaptive Moment Estimation (Adam)

 Adam combines insights from the momentum optimizers that accumulate the values of past gradients, and it also
introduces new terms based on the second moment of the gradient
o Similar to GD with momentum, Adam computes a weighted average of past gradients (first moment of the gradient),
i.e., 𝑉 𝑡 = 𝛽1 𝑉 𝑡−1 + 1 − 𝛽1 𝛻ℒ 𝜃 𝑡−1
o Adam also computes a weighted average of past squared gradients (second moment of the gradient),
2
i.e., 𝑈 𝑡 = 𝛽2 𝑈 𝑡−1 + 1 − 𝛽2 𝛻ℒ 𝜃 𝑡−1
𝑡 𝑡−1 𝑉𝑡
 The parameter update is:𝜃 = 𝜃 −𝛼
𝑈 𝑡 +𝜖
𝑉𝑡 𝑈𝑡
o Where: 𝑉 𝑡 = 1−𝛽1 and 𝑈𝑡 = 1−𝛽2
o The proposed default values are 𝛽1 = 0.9, 𝛽2 = 0.999, and 𝜖 = 10−8
• Other commonly used optimization methods include:
 Adagrad, Adadelta, RMSprop, Nadam, etc.
 Most commonly used optimizers nowadays are Adam and SGD with momentum

64
Learning Rate

• Learning rate
 The gradient tells us the direction in which the loss has the steepest rate of increase, but it does not tell us how far along
the opposite direction we should step
 Choosing the learning rate (also called the step size) is one of the most important hyper-parameter settings for NN
training

LR too small
LR too large

65
Learning Rate
• Training loss for different learning rates
 High learning rate: the loss increases or plateaus too quickly
 Low learning rate: the loss decreases too slowly (takes many epochs to reach a solution)

66
Learning Rate Scheduling

• Learning rate scheduling is applied to change the values of the learning rate during the training
 Annealing is reducing the learning rate over time (a.k.a. learning rate decay)
o Approach 1: reduce the learning rate by some factor every few epochs
• Typical values: reduce the learning rate by a half every 5 epochs, or divide by 10 every 20 epochs
o Approach 2: exponential or cosine decay gradually reduce the learning rate over time
o Approach 3: reduce the learning rate by a constant (e.g., by half) whenever the validation loss stops improving
• In TensorFlow: tf.keras.callbacks.ReduceLROnPleateau()
• Monitor: validation loss, factor: 0.1 (i.e., divide by 10), patience: 10 (how many epochs to wait before applying it), Minimum learning rate: 1e-6 (when to stop)

 Warmup is gradually increasing the learning rate initially, and afterward let it cool down until the end of the training

Exponential decay Cosine decay Warmup

67
Vanishing Gradient Problem

• In some cases, during training, the gradients can become either very small (vanishing gradients) of very large
(exploding gradients)
 They result in very small or very large update of the parameters
 Solutions: change learning rate, ReLU activations, regularization, LSTM units in RNNs

x1 … y1
…
x2 … y2
…
…
…

…
…
…
…

…
…

…
…
xN … yM
…
Small gradients, learns very slow

68
Generalization

• Underfitting
 The model is too “simple” to represent all the relevant
class characteristics
 E.g., model with too few parameters
 Produces high error on the training set and high error
on the validation set

• Overfitting
 The model is too “complex” and fits irrelevant
characteristics (noise) in the data
 E.g., model with too many parameters
 Produces low error on the training error and high error
on the validation set

69
Overfitting

• Overfitting – a model with high capacity fits the noise in the data instead of the
underlying relationship

• The model may fit the training data very well, but
fails to generalize to new examples (test or
validation data)

70
Regularization: Weight Decay

• ℓ𝟐 weight decay
 A regularization term that penalizes large weights is added to the loss function

Data loss Regularization loss

ℒ𝑟𝑒𝑔 𝜃 = ℒ 𝜃 + 𝜆 𝜃𝑘2
𝑘
 For every weight in the network, we add the regularization term to the loss value
o During gradient descent parameter update, every weight is decayed linearly toward zero
 The weight decay coefficient 𝜆 determines how dominant the regularization is during the gradient computation

71
Regularization: Weight Decay

• Effect of the decay coefficient 𝜆

 Large weight decay coefficient → penalty for weights with large values

72
Regularization: Weight Decay

• ℓ𝟏 weight decay
 The regularization term is based on the ℓ1 norm of the weights

ℒ𝑟𝑒𝑔 𝜃 = ℒ 𝜃 + 𝜆 𝑘 𝜃𝑘
 ℓ1 weight decay is less common with NN
o Often performs worse than ℓ2 weight decay
 It is also possible to combine ℓ1 and ℓ2 regularization
o Called elastic net regularization
2
ℒ𝑟𝑒𝑔 𝜃 = ℒ 𝜃 + 𝜆1 𝑘 𝜃𝑘 + 𝜆2 𝑘 𝜃𝑘

73
Regularization: Dropout
• Dropout
 Randomly drop units (along with their connections) during training
 Each unit is retained with a fixed dropout rate p, independent of other units
 The hyper-parameter p needs to be chosen (tuned)
o Often, between 20% and 50% of the units are dropped

74
Regularization: Dropout

• Dropout is a kind of ensemble learning

 Using one mini-batch to train one network with a slightly different architecture

minibatch minibatch minibatch minibatch

1 2 3 n

……
75
Regularization: Early Stopping
• Early-stopping
 During model training, use a validation set along with a training data
o E.g., validation/train ratio of about 25% to 75% (often)
 Stop when the validation accuracy (or loss) has not improved after n subsequent epochs
o The parameter n is called patience

Stop training

validation

76
Batch Normalization

• Batch normalization layers act similar to the data preprocessing steps mentioned earlier
 They calculate the mean μ and variance σ of a batch of input data, and normalize the data x to a zero mean and unit
variance
𝑥−𝜇
 i.e., 𝑥 =
𝜎

• BatchNorm layers alleviate the problems of proper initialization of the parameters and hyper-parameters
 Result in faster convergence training, allow larger learning rates
 Reduce the internal covariate shift

• BatchNorm layers are inserted immediately after convolutional layers or fully-connected layers, and before
activation layers
 They are very common with convolutional NNs

77
Hyper-parameter Tuning

• Training NNs can involve setting many hyper-parameters

• The most common hyper-parameters include:
 Number of layers, and number of neurons per layer
 Initial learning rate
 Learning rate decay schedule (e.g., decay constant)
 Optimizer type

• Other hyper-parameters may include:

 Regularization parameters (ℓ2 penalty, dropout rate)
 Batch size
 Activation functions
 Loss function
• Hyper-parameter tuning can be time-consuming for larger NNs

78
Hyper-parameter Tuning

• Grid search
 Check all values in a range with a step value

• Random search
 Randomly sample values for the parameter
 Often preferred to grid search

• Bayesian hyper-parameter optimization

 Is an active area of research

79
k-Fold Cross-Validation

• Using k-fold cross-validation for hyper-parameter tuning is common when the size of the training data is small
 It also leads to a better and less noisy estimate of the model performance by averaging the results across several folds

• E.g., 5-fold cross-validation (see the figure on the next slide)

1. Split the train data into 5 equal folds
2. First use folds 2-5 for training and fold 1 for validation
3. Repeat by using fold 2 for validation, then fold 3, fold 4, and fold 5
4. Average the results over the 5 runs (for reporting purposes)
5. Once the best hyper-parameters are determined, evaluate the model on the test data

80
k-Fold Cross-Validation
• Illustration of a 5-fold cross-validation

81
Ensemble Learning

• Ensemble learning is training multiple classifiers separately and combining their predictions
 Ensemble learning often outperforms individual classifiers
 Better results obtained with higher model variety in the ensemble
 Bagging (bootstrap aggregating)
o Randomly draw subsets from the training set (i.e., bootstrap samples)
o Train separate classifiers on each subset of the training set
o Perform classification based on the average vote of all classifiers
 Boosting
o Train a classifier, and apply weights on the training set (apply higher weights on misclassified examples, focus on “hard examples”)
o Train new classifier, reweight training set according to prediction error
o Repeat
o Perform classification based on weighted vote of the classifiers

82
Deep vs Shallow Networks

• Deeper networks perform better than shallow networks

 But only up to some limit: after a certain number of layers, the performance of deeper networks plateaus

output

Shallow Deep
NN NN

……

x1 x2 …… xN

input

83
Convolutional Neural Networks (CNNs)

• Convolutional neural networks (CNNs) were primarily designed for image data
• CNNs use a convolutional operator for extracting data features
 Allows parameter sharing
 Efficient to train
 Have less parameters than NNs with fully-connected layers
• CNNs are robust to spatial translations of objects in images
• A convolutional filter slides (i.e., convolves) across the image

Convolutional
3x3 filter
Input matrix

84
Convolutional Neural Networks (CNNs)

• When the convolutional filters are scanned over the image, they capture useful features
 E.g., edge detection by convolutions

0 1 0
Filter 1 -4 1
0 1 0
1 1 1 1 1 1 0.015686 0.015686 0.011765 0.015686 0.015686 0.015686 0.015686 0.964706 0.988235 0.964706 0.866667 0.031373 0.023529 0.007843
0.007843 0.741176 1 1 0.984314 0.023529 0.019608 0.015686 0.015686 0.015686 0.011765 0.101961 0.972549 1 1 0.996078 0.996078 0.996078 0.058824 0.015686
0.019608 0.513726 1 1 1 0.019608 0.015686 0.015686 0.015686 0.007843 0.011765 1 1 1 0.996078 0.031373 0.015686 0.019608 1 0.011765
0.015686 0.733333 1 1 0.996078 0.019608 0.019608 0.015686 0.015686 0.011765 0.984314 1 1 0.988235 0.027451 0.015686 0.007843 0.007843 1 0.352941
0.015686 0.823529 1 1 0.988235 0.019608 0.019608 0.015686 0.015686 0.019608 1 1 0.980392 0.015686 0.015686 0.015686 0.015686 0.996078 1 0.996078
0.015686 0.913726 1 1 0.996078 0.019608 0.019608 0.019608 0.019608 1 1 0.984314 0.015686 0.015686 0.015686 0.015686 0.952941 1 1 0.992157
0.019608 0.913726 1 1 0.988235 0.019608 0.019608 0.019608 0.039216 0.996078 1 0.015686 0.015686 0.015686 0.015686 0.996078 1 1 1 0.007843
0.019608 0.898039 1 1 0.988235 0.019608 0.015686 0.019608 0.968628 0.996078 0.980392 0.027451 0.015686 0.019608 0.980392 0.972549 1 1 1 0.019608
0.043137 0.905882 1 1 1 0.015686 0.035294 0.968628 1 1 0.023529 1 0.792157 0.996078 1 1 0.980392 0.992157 0.039216 0.023529
1 1 1 1 1 0.992157 0.992157 1 1 0.984314 0.015686 0.015686 0.858824 0.996078 1 0.992157 0.501961 0.019608 0.019608 0.023529
0.996078 0.992157 1 1 1 0.933333 0.003922 0.996078 1 0.988235 1 0.992157 1 1 1 0.988235 1 1 1 1
0.015686 0.74902 1 1 0.984314 0.019608 0.019608 0.031373 0.984314 0.023529 0.015686 0.015686 1 1 1 0 0.003922 0.027451 0.980392 1
0.019608 0.023529 1 1 1 0.019608 0.019608 0.564706 0.894118 0.019608 0.015686 0.015686 1 1 1 0.015686 0.015686 0.015686 0.05098 1
0.015686 0.015686 1 1 1 0.047059 0.019608 0.992157 0.007843 0.011765 0.011765 0.015686 1 1 1 0.015686 0.019608 0.996078 0.023529 0.996078
0.019608 0.015686 0.243137 1 1 0.976471 0.035294 1 0.003922 0.011765 0.011765 0.015686 1 1 1 0.988235 0.988235 1 0.003922 0.015686
0.019608 0.019608 0.027451 1 1 0.992157 0.223529 0.662745 0.011765 0.011765 0.011765 0.015686 1 1 1 0.015686 0.023529 0.996078 0.011765 0.011765
0.015686 0.015686 0.011765 1 1 1 1 0.035294 0.011765 0.011765 0.011765 0.015686 1 1 1 0.015686 0.015686 0.964706 0.003922 0.996078
0.007843 0.019608 0.011765 0.054902 1 1 0.988235 0.007843 0.011765 0.011765 0.015686 0.011765 1 1 1 0.015686 0.015686 0.015686 0.023529 1
0.007843 0.007843 0.015686 0.015686 0.960784 1 0.490196 0.015686 0.015686 0.015686 0.007843 0.027451 1 1 1 0.011765 0.011765 0.043137 1 1
0.023529 0.003922 0.007843 0.023529 0.980392 0.976471 0.039216 0.019608 0.007843 0.019608 0.015686 1 1 1 1 1 1 1 1 1

Input Image Convoluted Image

85
Convolutional Neural Networks (CNNs)

• In CNNs, hidden units in a layer are only connected to a small region of the layer before it (called local receptive
field)
 The depth of each feature map corresponds to the number of convolutional filters used at each layer

w1 w2
w3 w4 w5 w6
w7 w8
Filter 1
Filter 2
Input Image
Layer 1
Feature Map Layer 2
Feature Map

86
Convolutional Neural Networks (CNNs)

• Max pooling: reports the maximum output within a rectangular neighborhood

• Average pooling: reports the average output of a rectangular neighborhood
• Pooling layers reduce the spatial size of the feature maps
 Reduce the number of parameters, prevent overfitting

MaxPool with a 2×2 filter with stride of 2

1 3 5 3
4 5
4 2 3 1
3 4
3 1 1 3

0 1 0 4
Output Matrix
Input Matrix

87
Convolutional Neural Networks (CNNs)

• Feature extraction architecture

 After 2 convolutional layers, a max-pooling layer reduces the size of the feature maps (typically by 2)
 A fully convolutional and a softmax layers are added last to perform classification

Living Room

Bedroom

Kitchen

128

256
256

512
512

512
512
128

256

512

512
64
64

Bathroom

Outdoor

Conv layer

Max Pool
Fully Connected Layer

88
Residual CNNs

• Residual networks (ResNets)

 Introduce “identity” skip connections
o Layer inputs are propagated and added to the layer output
o Mitigate the problem of vanishing gradients during training
o Allow training very deep NN (with over 1,000 layers)
 Several ResNet variants exist: 18, 34, 50, 101, 152, and 200 layers
 Are used as base models of other state-of-the-art NNs
o Other similar models: ResNeXT, DenseNet

89
Recurrent Neural Networks (RNNs)
• Recurrent NNs are used for modeling sequential data and data with varying length of inputs and outputs
 Videos, text, speech, DNA sequences, human skeletal data

• RNNs introduce recurrent connections between the neurons

 This allows processing sequential data one element at a time by selectively passing information across a sequence
 Memory of the previous inputs is stored in the model’s internal state and affect the model predictions
 Can capture correlations in sequential data

• RNNs use backpropagation-through-time for training

• RNNs are more sensitive to the vanishing gradient problem than CNNs

• RNN variants:
 Basic (Vanilla) RNN networks
o Are sensitive to the vanishing gradient problem
 Long Short-Term Memory (LSTM) networks
o LSTM mitigates the vanishing/exploding gradient problem
• Solution: a Memory Cell, updated at each step in the sequence
o Three gates control the flow of information to and from the Memory Cell
 Gated Recurrent Networks (GRU)
o Similar to LSTM, less commonly used than LSTM
90
Recurrent Neural Networks (RNNs)
• RNN use same set of weights 𝑤ℎ and 𝑤𝑥 across all time steps
 A sequence of hidden states ℎ𝑜 , ℎ𝑜 ℎ2 , ℎ3 , … is learned, which represents the memory of the network
 The hidden state at step t, ℎ 𝑡 , is calculated based on the previous hidden state ℎ 𝑡 − 1 and the input at the current step
𝑥 𝑡 , i.e., ℎ 𝑡 = 𝑓ℎ 𝑤ℎ ∗ ℎ 𝑡 − 1 + 𝑤𝑥 ∗ 𝑥 𝑡
 The function 𝑓ℎ ∙ is a nonlinear activation function, e.g., ReLU or tanh
• RNN shown rolled over time

HIDDEN STATES SEQUENCE: ℎ𝑜 , ℎ1 , ℎ2 , ℎ3 , … . OUTPUT

𝑤ℎ 𝑤ℎ 𝑤ℎ 𝑤𝑦

h0 𝑓ℎ (·) h1 𝑓ℎ (·) h2 𝑓ℎ (·) h3 𝑓𝑦 (·)

𝑤𝑥 𝑤𝑥 𝑤𝑥

x1 x2 x3

INPUT SEQUENCE: 𝑥1 , 𝑥2 , 𝑥3 , … .
91
Recurrent Neural Networks (RNNs)
• RNNs can have one of many inputs and one of many outputs

RNN Application Input Output

A person riding a
Image
motorbike on dirt
Captioning
road

Sentiment Awesome movie. Highly

Analysis recommended. Positive

Machine
Happy Diwali शुभ दीपावली
Translatio
n
Slide credit: Param Vir SIngh– Deep Learning 92
Bidirectional RNNs

• Bidirectional RNNs incorporate both forward and backward passes through sequential data
 The output may not only depend on the previous elements in the sequence, but also on future elements in the sequence
 It resembles two RNNs stacked on top of each other

ℎ𝑡 = 𝜎(𝑊 (ℎℎ) ℎ𝑡−1 + 𝑊 (ℎ𝑥) 𝑥𝑡 )

ℎ𝑡 = 𝜎(𝑊 (ℎℎ) ℎ𝑡+1 + 𝑊 (ℎ𝑥) 𝑥𝑡 )

𝑦𝑡 = 𝑓 ℎ𝑡 ; ℎ𝑡

Outputs both past and future elements

93
LSTM Networks

• Long Short-Term Memory (LSTM) networks are a variant of RNNs

• LSTM mitigates the vanishing/exploding gradient problem
 Solution: a Memory Cell, updated at each step in the sequence
• Three gates control the flow of information to and from the Memory Cell
 Input Gate: protects the current step from irrelevant inputs
 Output Gate: prevents current step from passing irrelevant information to later steps
 Forget Gate: limits information passed from one cell to the next

• Most modern RNN models use either LSTM units or other more advanced types of recurrent units (e.g., GRU
units)

94
LSTM Networks

• LSTM cell
 Input gate, output gate, forget gate, memory cell
 LSTM can learn long-term correlations within data sequences

95
Transformer Networks
• Transformer networks have been initially designed for processing test data in Large Language Models, such as
GPT-3, ChatGPT, etc.
 Later, they have been used for image tasks, and tabular data processing

• The main block of transformers is the self-attention mechanism, which uses scaled dot-product attention to force
the model to attend to portions of the data
 Several self-attention modules are combined into a multi-head attention layer

96
Generative adversarial network (GAN)
• Generative models are capable of making new plausible data.
• The term “adversarial” in this case pertains to a unique architecture
for training an effective generator network.
• Random input vectors are given to the generator to make plausible
samples that are ideally discriminated 50:50 in a fully trained model.
However, some issue in training GANs include:

• Vanishing gradient
 Best if discriminator starts in a less robust state so it is improving along with the
generator rather than being too sophisticated from the start

 Modified minimax loss can help with this, proposed in original paper, maximizing
log(D(G(z))) instead of minimizing it.

• Mode Collapse
 Local minimum in discriminator training, Wasserstein loss can help with this, uses
critic model, D(x) – D(G(z)), rather than a threshold valued discriminator

• Failure to Converge
 Regularization of discriminator can help with this

 Over training generator past random feedback from discriminator must be

https://developers.google.com/machine-learning/gan/discriminator
monitored
Generative Adversarial Networks. Goodfellow 2014 97
Slide Acknowledgements

1. Hung-yi Lee – Deep Learning Tutorial

2. Ismini Lourentzou – Introduction to Deep Learning
3. James Hays, Brown – Machine Learning Overview
4. Aleksandar (Alex) Vakanski— Deep Learning Presentation
5. Param Vir Singh, Shunyuan Zhang, Nikhil Malik – Deep Learning
6. Sebastian Ruder – An Overview of Gradient Descent Optimization Algorithms (link)

98
Other Useful Resources

99
Happy Learning

https://www.deeplearningbook.org/
https://d2l.ai/