DEEP LEARNING

UNIT - I

By,
Dr. Himani Deshpande 1
UNIT - I

1.1

History of Deep Learning,

1.2
Deep Learning Success Stories,
Multilayer Perceptrons (MLPs),
Deep Networks: Three Classes of Deep Learning,
Representation Power of MLPs,
Basic Terminologies of Deep Learning
Sigmoid Neurons,
Gradient Descent,
Feedforward Neural Networks,
2
Representation Power of Feedforward Neural Networks
DEEP LEARNING

¡ Deep learning is a branch of machine

learning that deploys algorithms for
data processing and imitates the
thinking process and even
develops abstractions.

3
HISTORY OF DEEP LEARNING

¡ Most of us know Deep Learning to be a 21st Century invention, but it has

been around since the 1940s.

4
 History of
Deep
Learning

BIOLOGICAL NEURON

Output

Input

This myelin sheath allows electrical impulses to transmit quickly and efficiently along the
nerve cells. It gets thicker and thicker as you generate memories stronger and stronger 5
 History of
Deep
Learning

NEURONS

Biological Neuron Artificial Neuron

6
History of Deep Learning

7
1943 History of
Deep
EVOLUTION OF DEEP LEARNING Learning

MCCULLOCH-PITTS NEURAL MODEL

The McCulloch-Pitts neural model was
the earliest NN discovered in 1943.

McCulloch and Pitts's paper provided a

way to describe brain functions in abstract
terms, and showed that simple elements
connected in a neural network can have
immense computational power.

The first mathematical model of a neural network.

8
1943 History of
Deep
MCCULLOCH PITTS(MP MODEL) IN NEURAL NETWORK Learning
OR
LINEAR THRESHOLD GATE MODEL

The diagrammatic representation of the McCulloch Pitts Model 9

1943 History of
Deep
Learning

AND GATE USING MP MODEL

x1
W1

Y Y
W2
x2

10
1943 History of
Deep
Learning

LIMITATIONS OF MP

¡ One of the difficulties with the McCulloch-Pitts neuron was its simplicity.

¡ It only allowed for binary inputs and outputs.

¡ It only used the threshold step activation function and it did not
incorporate weighting the different inputs.

¡ All the inputs to the model have equal weights. 11

1957 History of
Deep
Learning

EVOLUTION OF DEEP LEARNING- PERCEPTRON

¡ An American Scientist Rosenblatt was very much inspired by the
biological neuron and its ability to learn and the term perceptron was
introduced by him around 1957.

¡ Rosenblatt’s perceptron has one or more inputs with only one output
and a processor.

12
1957 History of
Deep
Learning

EVOLUTION OF DEEP LEARNING

13
1957 History of
Deep

Perceptron neural model Learning

14
1957 History of
Deep
Learning

COMPONENTS OF PERCEPTRON NEURAL MODEL

15
1960 History of
Deep
Learning
DELTA LEARNING RULE OR WIDROW HOFF RULE

It was developed by
Bernard Widrow and
Marcian Hoff and It
depends on supervised
learning and has a
continuous activation
function. 16
1960 History of
Deep

Delta Learning Rule Learning

view of the Sum Squared Error Function

¡ It is also known as the Least Mean Square method and it minimizes error over all the training patterns.
17

¡ It is based on a gradient descent approach which continues forever.

PERCEPTRON VS ADALINE

19
 History of
Deep
Learning

MLP

21
 History of
Deep
Learning

PERCEPTRON LIMITATION

22
History of
Deep
Learning

23
 History of
Deep
Learning

BACKPROPOGATION

24
 History of
Deep
Learning

GRADIENT DESCENT

Gradient descent is generally attributed to Augustin-Louis

Cauchy, who first suggested it in 1847

In mathematics, gradient descent (also often

called steepest descent) is a first-
order iterative optimization algorithm for finding a local
minimum of a differentiable function.

25
UAT

26
 History of
Deep
Learning

DEEP LEARNING TIME LINE

27
 History of
Deep
Learning

MILESTONES IN THE DEVELOPMENT OF DEEP

28
THE HISTORY OF DEEP LEARNING

¡ The history of deep learning dates back to 1943 when Warren McCulloch and Walter
Pitts created a computer model based on the neural networks of the human brain.
¡ Warren McCulloch and Walter Pitts used a combination of mathematics and
algorithms they called threshold logic to mimic the thought process.
¡ Since then, deep learning has evolved steadily, over the years with two significant
breaks in its development.
¡ The development of the basics of a continuous Back Propagation Model is credited to
Henry J. Kelley in 1960. Stuart Dreyfus came up with a simpler version based only on
the chain rule in 1962. The concept of back propagation existed in the early 1960s
but only became useful until 1985. 29
DEEP LEARNING SUCCESS STORIES

¡ AI-Powered Healthcare at Scale

¡ AI-Powered Weather Forecasting
¡ AI Accelerated Cyber Defense
¡ Defending the Planet with AI

32
APPLICATIONS OF NEURAL NETWORK

ALEXA

33
DEEP LEARNING SUCCESS STORIES

34
DEEP LEARNING SUCCESS STORY
¡ Self Driving Cars or Autonomous Vehicles
¡ News Accumulation and Fake News Detection
¡ Natural Language Processing NLP
¡ Virtual Assistants
¡ Visual Recognition
¡ Deep Learning Applications in Healthcare
¡ Personalization
¡ Colourization of Black and White Images
¡ Adding Sounds to Silent Movies
35

¡ Automatic Machine Translation

36
37
MULTILAYER PERCEPTRON

38
MULTILAYER NEURAL NETWORK

Perceptron Multilayer Neural Network

After Rosenblatt perceptron was developed in the 1950s, there

was a lack of interest in neural networks until 1986, when
Dr. Hinton and his colleagues developed the backpropagation
algorithm to train a multilayer neural network.
39
SLP & MLP

40
MLP

¡ Multi-Layer perceptron defines the most complex architecture of artificial neural

networks. It is substantially formed from multiple layers of the perceptron.

A multilayer perceptron (MLP) is a feed-forward artificial neural

network that generates a set of outputs from a set of inputs.
An MLP is characterized by several layers of input nodes connected as
a directed graph between the input and output layers.

MLP uses backpropagation for training the network. MLP is a deep

learning method. 41
42
REPRESENTATION POWER

¡ Representation power is related to the ability of a neural network to

assign proper labels to a particular instance and create well-defined
accurate decision boundaries for that class.

43
MLP REPRESENTATION POWER

Multilayer perceptron (MLP), has a high representation power

44
MLP REPRESENTATION POWER

2 INPUTS

45
MLP REPRESENTATION POWER

2 INPUTS

First perceptron fires for -1 , -1

First perceptron fires for -1 , 1
First perceptron fires for 1, -1
First perceptron fires for 1, 1

46
MLP REPRESENTATION POWER

2 INPUTS

47
MLP REPRESENTATION POWER

2 INPUTS

We have made sure that every

neuron in middle layer fires to one
specific input

So, weights in final layer can be

adjusted.

48
MLP REPRESENTATION POWER

3 INPUTS

49
MLP REPRESENTATION POWER

Multilayer perceptron (MLP), has a high representation power

50
LP REPRESENTATION POWER

51
MLP REPRESENTATION POWER

MORAL OF THE STORY

MLP with a single hidden layer can represent any Boolean

function

52
53
54
SIGMOID FUNCTION

55
SIGMOID NEURON

The building block of the deep neural

networks is called the sigmoid neuron.

Sigmoid neurons are similar to

perceptrons, but they are slightly
modified such that the output from
the sigmoid neuron is much smoother
than the step functional output from
perceptron. 56
WHY SIGMOID NEURON ?

Sigmoid neurons don't just output 0 or 1.

Perceptron suffers from a major disadvantage of
being very harsh at the boundaries when it
comes to deciding the output.

57
 Perceptron suffers from a major disadvantage of
being very harsh at the boundaries when it
WHY SIGMOID NEURON ?
comes to deciding the output.

58
SIGMOID NEURONS

¡ Sigmoid neurons can be used for both binary classification and regression problems.
The output will be a value between ‘0’ and ‘1’ .

¡ In contrast to Perceptron 0/1 function, ‘S-shaped’ Sigmoid family of functions gives

us a smoother curve — closer to how humans make decisions.

Sigmoid neurons are similar to perceptrons, but modified so that small changes
in their weights and bias cause only a small change in their output. 59
PERCEPTRON FUNCTION

60
SIGMOID FUNCTION

61
PERCEPTRON VS SIGMOID

62
PERCEPTRON VS SIGMOID

63
64
GUESS THE GRE SCORE

¡ Guess your friend's GRE score

320
¡ 300
¡ 310
¡ 307

65
GOAL

¡ Find a better way of traversing the error surface so that we can reach
the minimum value quickly without resorting to brute force search

GDà .
An Algorithm
To
Minimize a function
By
Optimizing parameters 66
VALUES OF WEIGHTS AND BIAS

Network starts with the random guess for values of w, b

Then changes the value by some factor

¡ Θ = [w, b] à values of w and b

¡ ! Θ = [! w, ! b ] à change in values of w and b
67
CHANGING THE WEIGHTS

68
CHANGING THE WEIGHTS

69
GRADIENT DESCENT

¡ Gradient Descent is defined as one of the most commonly used iterative

optimization algorithms of machine learning to train the machine learning
and deep learning models.
¡ It helps in finding the local minimum of a function.

70
GRADIENT DESCENT

Ø Gradient descent is an optimization algorithm for minimizing the cost.

Ø
Ø To think of it intuitively, while climbing down a hill you should take
small steps and walk down instead of just jumping down at once.

Ø Consider the bottom to be the minimum cost point.

71
GRADIENT DESCENT

¡ First-Order Derivative
¡ The first order derivatives tell about the direction of the function whether the
function is increasing or decreasing.
¡ The first derivative math or first-order derivative can be interpreted as an
instantaneous rate of change.
¡ It can also be predicted from the slope of the tangent line.

72
 The main objective of using a gradient descent
algorithm is to minimize the cost function using
GRADIENT DESCENT iteration.

¡ Calculates the first-order derivative of the function to compute the gradient or slope
of that function.

¡ Move away from the direction of the gradient, which means the slope increased
from the current point by alpha times, where Alpha is defined as the Learning Rate.

¡ It is a tuning parameter in the optimization process which helps to decide the length
of the steps.

73
GRADIENT DESCENT

It is the learning algorithm for

Sigmoid Neuron

74
GRADIENT DESCENT

75
ERROR SURFACE

76
ERROR SURFACE

77
GRADIENT DESCENT

Max error

Min error
78
GRADIENT DESCENT

Gradient Descent Algorithm: How small should ! be?

¡ Pick an initial point "!

¡ Iterate until convergence

!!"# = !! − $! %&(!! )
where #" is the $ "# step size (also called learning rate)
Possible Stopping Criteria: iterate until 79

∇&("" ) ≤ * for some * > 0

 New value = Old value - learning rate * gradient
GRADIENT DESCENT

" # = #!
Step size: 0.8
# (#) = −4

80
 New value = Old value - learning rate * gradient
GRADIENT DESCENT

" # = #!

81

Derivative of curve with respect to x-axis

 New value = Old value - learning rate * gradient
GRADIENT DESCENT

" # = #!
Step size: 0.8
# (#) = −4

Derivative of curve with

respect to x-axis

82
 New value = Old value - learning rate * gradient
GRADIENT DESCENT

" # = #!
Step size: 0.8

# (#) = −4
Derivative à 2x

83
 New value = Old value - learning rate * gradient
GRADIENT DESCENT
" # = #!

Step size: 0.8

)(&) = −*
)(%) = −* − (,. - ∗ / ∗ (−*))
Derivative à 2x

84
GRADIENT DESCENT
" # = #!

Step size: 0.8

' ($) = −4
' (") = 2.4
Derivative à 2x

85
GRADIENT DESCENT " # = #!

Step size: 0.8

' ($) = −4
' (") = 2.4
' (%) = 2.4 − 0.8 ∗ 2 ∗ 2.4
Derivative à 2x

' (") = 0.4

86
GRADIENT DESCENT " # = #!

Step size: .8
' ($) = −4
' (") = 2.4
' (%) = −1.44

Derivative à 2x

87
GRADIENT DESCENT " # = #!

Step size: .8
' ($) = −4
' (") = 2.4
' (%) = −1.44

' (&) = −1.44 − 0.8 ∗ 2 ∗ (−1.44) Derivative à 2x

88
GRADIENT DESCENT
" # = #!

Step size: .8
' ($) = −4
' (") = 2.4
' (%) = −1.44
' (&) = .864
' (() = −0.5184
' (') = 0.31104

' (&$) = −8.842965 − 07

89
GRADIENT DESCENT

90

Step size: 0.9

GRADIENT DESCENT

91

Step size: .2
GRADIENT DESCENT

92

Step size matters!

GRADIENT DESCENT

93

Step size matters!

GRADIENT DESCENT WITH TWO VARIABLES

(m,c) c

Loss function
¡ Y = mx +c
¡ Loss function (m,c)

c m

94
GRADIENT DESCENT WITH TWO VARIABLES

c
(m,c) X Y
1 2
3 4
¡ Y = mx +C
Training Data
¡ L(m,c)

¡ Initial assumptions à c = 0, m = 1
¡ L(m,c)
= [2-(c+m*1)] ^2 + [4-(c+3m)] ^2
95
GRADIENT DESCENT WITH TWO VARIABLES

¡ Y = mx +C
¡ L(m,c) c
(m,c) X Y
1 2
¡ Initial assumptions à c = 0, m = 1 3 4
¡ L(m,c) = [2-(c+m*1)] ^2 + [4-(c+3m)] ^2 Training Data

dL/dc= -2 [2-(c+m)] + [-2(4-(c+3m)]

= -2 [2-(1)] + [-2(4-(4+3)]
= -2[1] + [-2] = -4

C new = C old - L R* (-4)

96
= 0 - (0.001 *-4)
= 0.004
VARIANTS OF GRADIENT DESCENT ALGORITHM

97
 NEURONS

Sigmoid Neuron
Smooth output

MLP
Harsh Threshhold

PERCEPTRON
Real Input & Binary Output

FEED FORWARD
MP NEURON 98

Binary Input & Output

Multilayered Network of Neurons

99
MUTI LAYERED

100
FEED FORWARD NN

101
REPRESENTATION POWER

102
103
BASIC TERMINOLOGIES OF NN

ü Neuron ü Gradient Descent

NN
ü Weights ü Forward Propagation
TERMINOLOGIES
ü Bias ü Back Propagation
ü Activation Function ü Epochs
ü Layers in NN ü Dropouts
ü Learning Rate ü Cost Function
etc……
104
IMAGE REPRESENTATION IN PIXELS

105
 NN
WORKING

Hidden layers Output layer 106

Input layer
 NN
WORKING

Hidden layers Output layer 107

Input layer
 NN
WORKING
28 pixel

…
28 pixel

28 X 28
= 784 pixel

Hidden layers Output layer 108

Input layer
 NN
WORKING x1
0.8

x2 0.2

x3
784 inputs/nodes 0.1

x4
0.3

…
28 pixel

x781 0.8

28 pixel x782
0.2

x783
0.7

x784
109
 NN
WORKING x1
0.8

x2 0.2

x3
784 inputs/nodes 0.1

x4
0.3

…
28 pixel

x781 0.8

28 pixel x782
0.2

x783
0.7

x784
Weights 110
 NN
WORKING x1
0.8
B1
x2 0.2
B2
x3
784 inputs/nodes 0.1
B3
x4
0.3

…
28 pixel

B4
x781 0.8

28 pixel x782 B5
0.2

x783
B6
0.7

x784
Weights 111
 NN Activation
WORKING (x1*0.8 + x3*0.2)+B1
x1
0.8
Function
B1
x2 0.2

B2
x3
0.1
784 inputs/nodes
B3
x4
0.3

…
x781 0.8
B4

x782 B5
0.2

x783
0.7 B6

x784
Weights 112
 NN
WORKING x1
0.8
B1
0.6
x2 0.2

B2 0.2
x3 0.2
0.1
784 inputs/nodes
B3 0.4 0.5
x4
0.3

…
28 pixel

B4 0.3
x781 0.8 0.3

0.3 0.2
28 pixel x782 B5
0.2

x783 0.5
0.7 B6

x784
Weights 113
 NN
WORKING x1
0.8
B1
0.6
x2 0.2

B2 0.2
x3 0.2
0.1
784 inputs/nodes
B3 0.4 0.5
x4
0.3

…
28 pixel

B4 0.3
x781 0.8 0.3

0.3
28 pixel x782 B5
0.2 0.2

x783 0.5
0.7 B6

x784
114
 NN
WORKING n
x1 i o
0.8
e d ict
B1
p r
0.6
x2 0.2

B2 0.2
x3 0.2
0.1 0.5 Actual Error
784 inputs/nodes
B3 0.4 Output
0.5
x4
0.3 0 -0.5

…
0.4
28 pixel

B4 0.3
x781 0.3 1 +0.6
0.8
0.3 0.1 0 -0.1
28 pixel x782 B5
0.2 0.2

x783 0.5
0.7 B6

x784
115
 NN
WORKING x1
0.8
B1
0.6
x2 0.2

B2 0.2
x3 0.2
0.1
784 inputs/nodes
B3 0.4 0.5
x4
0.3

…
28 pixel

B4 0.3
x781 0.8 0.3

0.3
28 pixel x782 B5
0.2 0.2

x783 0.5
0.7 B6

x784
116
 28 pixel NN

28 pixel
WORKING

784 inputs/nodes

117
 NN
WORKING
0.4
0.4
0.3
0.2
B2 0.1
784 inputs/nodes 0.2 0.3
0.2 0.1
0.3
0.5
28 pixel

0.5
0.5
0.4
0.4
28 pixel B5 0.2
0.2 0.2

0.3
0.3

119
 NN
WORKING
0.4
0.4
0.3
0.1
B2 0.1
784 inputs/nodes 0.2
0.2 0.06
0.3
0.2
0.3
0.3
0.9
28 pixel

0.5
0.5
0.4
0.4
0.4 0.04
28 pixel
0.7
0.7 B5 0.2

0.3
0.3
0.3

120
 NN
WORKING
0.5
0.4
0.1
0.2
0.4
0.1
0.7
0.2
0.2 0.2
0.7
784 inputs/nodes
0.2
0.3
0.3
0.4 0.3
0.77
28 pixel

0.5
0.5 0.4
0.4
0.5
0.5
0.4
0.3
28 pixel
0.03
0.7
0.6 0.3
0.3
0.3
0.3
0.2

121
 NN
WORKING
0.5
0.4 Move back to
previous values
0.1
0.2
0.4
0.1
0.7
0.2
0.2 0.2
0.7
784 inputs/nodes
0.2
0.3
0.3
0.4 0.3
0.77
28 pixel

0.5
0.5 0.4
0.4
0.5
0.5
0.4
0.3
28 pixel
0.03
0.7
0.6 0.3
0.3
0.3
0.3
0.2

122
 NN
WORKING
0.4
0.4
0.3
0.1
B2 0.1
784 inputs/nodes 0.2
0.2 0.06
0.3
0.2
0.3
0.3
0.9
28 pixel

0.5
0.5
0.4
0.4
0.4 0.04
28 pixel
0.7
0.7 B5 0.2

0.3
0.3
0.3

123
NN PREDICTION EXAMPLE

124
NN IS A BLACK BOX

125
NN TERMINOLOGIES

6) MLP (Multi Layer perceptron)

A single neuron would not be able to perform highly complex tasks. Therefore, we use stacks of neurons
to generate the desired outputs.

In the simplest network we would have an input layer, a hidden layer and an output layer. Each layer has
multiple neurons and all the neurons in each layer are connected to all the neurons in the next layer. These
networks can also be called as fully connected networks.

126
 NN TERMINOLOGIES
LEARNING RATE –

The learning rate is defined as the amount of minimization in the cost function in each
iteration.
In simple terms, the rate at which we descend towards the minima of the cost function is
the learning rate.
We should choose the learning rate very carefully since it should neither be very large
that the optimal solution is missed and nor should be very low that it takes forever for
the network to converge.
127
NN TERMINOLOGIES
LEARNING RATE

• To find the best learning rate, start with a

very low value (10^-6) and slowly multiply it
by a constant until it reaches a very high
value (e.g. 10). Measure your model
performance (vs the log of your learning
rate) in your Weights and Biases to
determine which rate served you well for
your problem.
• You can then retrain your model using this
optimal learning rate. 128
FORWARD PROPAGATION

Forward Propagation refers to the movement of the input through the hidden
layers to the output layers.
In forward propagation, the information travels in a single direction FORWARD.
The input layer supplies the input to the hidden layers and then the output is
generated.
There is no backward movement.

129
BACK PROPAGATION

130
131
 NN TERMINOLOGIES
BACKPROPAGATION

¡ When we define a neural network, we assign random weights and bias values to our nodes.
¡ Once we have received the output for a single iteration, we can calculate the error of the network. This error is
then fed back to the network along with the gradient of the cost function to update the weights of the network.
¡ These weights are then updated so that the errors in the subsequent iterations is reduced. This updating of weights
using the gradient of the cost function is known as back-propagation.

¡ In back-propagation the movement of the network is backwards, the error along with the gradient flows back from
the out layer through the hidden layers and the weights are updated.

132
NN TERMINOLOGIES
COST FUNCTION

When we build a network, the network tries to predict the output as close as possible to the
actual value.

We measure this accuracy of the network using the cost/loss function.

The cost or loss function tries to penalize the network when it makes errors.

133
NN TERMINOLOGIES
EPOCH

One forward pass and one backward pass of all training datasets is 1 Epoch

One Epoch leads to underfitting of the curve. 134

GRADIENT DESCENT

Gradient descent was initially discovered by "Augustin-Louis Cauchy" in mid

of 18th century.

Gradient Descent is defined as one of the most commonly used iterative

optimization algorithms of machine learning to train the machine learning
and deep learning models. It helps in finding the local minimum of a
function.

135
GRADIENT DESCENT

¡ The best way to define the local minimum or local maximum of a function using
gradient descent is as follows:
à If we move towards a negative gradient or away from the gradient of the function at the
current point, it will give the local minimum of that function.
à Whenever we move towards a positive gradient or towards the gradient of the function at
the current point, we will get the local maximum of that function.

136
GRADIENT DESCENT

¡ This entire procedure is known as Gradient Ascent, which is also known as steepest descent.

¡ The main objective of using a gradient descent algorithm is to minimize the cost function using
iteration. To achieve this goal, it performs two steps iteratively:
à Calculates the first-order derivative of the function to compute the gradient or slope of that
function.
à Move away from the direction of the gradient, which means slope increased from the current
point by alpha times, where Alpha is defined as Learning Rate. It is a tuning parameter in the
optimization process which helps to decide the length of the steps.

137
 A multi-layer perceptron (MLP) is a form of feedforward neural
network that consists of multiple layers of computation nodes
FEED FORWARD NN that are connected in a feed-forward way.

¡ A feed-forward neural network, in which some routes are cycled, is the polar
opposite of a Recurrent Neural Network.

¡ The feed-forward model is the basic type of neural network because the input
is only processed in one direction. The data always flows in one direction and
never backwards/opposite.

A Feed Forward Neural Network is an artificial Neural Network in which

the nodes are connected circularly.

The process of receiving an input to produce some kind of output to make some 138

kind of prediction is known as Feed Forward.

FEED FORWARD NN

¡ Feed Forward neural network is the core of many other important neural networks
such as convolution neural network.
¡ There can be multiple hidden layers which depend on what kind of data you are
dealing with. The number of hidden layers is known as the depth of the neural
network. The deep neural network can learn from more functions. Input layer first
provides the neural network with data and the output layer then make predictions
on that data which is based on a series of functions. ReLU Function is the most
commonly used activation function in the deep neural network.
139
140
CLASSES OF DEEP LEARNING

¡ Supervised,
¡ Unsupervised and
¡ Hybrid

141
142
SUPERVISED LEARNING

¡ Supervised learning involves training a deep learning model on labelled data, where
both input data and corresponding output labels are provided.
¡ The model learns to map the input data to the correct output by minimizing the
discrepancy between the predicted and true labels.
¡ It is commonly used for tasks such as image classification, object detection, and
natural language processing, where the goal is to learn the mapping between inputs
and pre-defined outputs.

143
UNSUPERVISED LEARNING

¡ Unsupervised learning involves training a deep learning model on unlabelled data,

without explicit output labels.
¡ The model learns to find patterns, structures, or representations in the input data
without specific guidance.
¡ This approach is useful for tasks such as clustering, dimensionality reduction, and
generative modeling.
¡ Unsupervised learning enables the model to discover inherent relationships and
patterns within the data.
144
HYBRID

¡ Hybrid learning refers to a combination of supervised and unsupervised learning

approaches.
¡ In this approach, a deep learning model is trained using both labeled and unlabeled
data.
¡ The unsupervised learning component helps the model to learn useful
representations or features from the unlabeled data, which can then be utilized in
the supervised learning phase to improve the model's performance on specific tasks.
¡ Hybrid learning leverages the benefits of both paradigms and has been shown to
achieve better performance in certain scenarios.
145
146