The document provides an overview of deep learning, focusing on its fundamentals such as neural networks, activation functions, and the process of image recognition. It explains the necessity of deep learning, highlighting its capabilities to handle large data sets and perform complex tasks, along with various applications including cancer detection, robot navigation, and autonomous driving. Additionally, it delves into the mechanics of neural networks, including the training process, backpropagation, and the challenges faced during training.

1. What is
Deep Learning?

2. Image Recognition
How does image recognition work?

3. Image Recognition
How does image recognition work?
Images of Dogs

4. Image Recognition
How does image recognition work?
Images of Dogs Artificial Neural Network

5. Image Recognition
How does image recognition work?
Images of Dogs New Images of AnimalsArtificial Neural Network

6. Image Recognition
How does image recognition work?
Images of Dogs Identifies which are dogsNew Images of AnimalsArtificial Neural Network

7. Image Recognition
Lets understand how Artificial Neural Network identifies the images

8. Image Recognition
Lets understand how Artificial Neural Network identifies the images
Training Data

9. Image Recognition
Lets understand how Artificial Neural Network identifies the images
Training Data
Artificial Neural Network

10. Image Recognition
Lets understand how Artificial Neural Network identifies the images
Training Data Feature Extraction
Artificial Neural Network

11. Image Recognition
Lets understand how Artificial Neural Network identifies the images
Training Data Feature Extraction
Artificial Neural Network
Artificial Neural Network
Test Data

12. Image Recognition
Lets understand how Artificial Neural Network identifies the images
Training Data Feature Extraction
Identifies the Dogs
Artificial Neural Network
Artificial Neural Network
Test Data

13. What’s in it for you?
What is Deep Learning?
Why do we need Deep Learning?
Applications of Deep Learning
What is a Neural Network?
Activation Functions
Working of a Neural Network

14. What is Deep Learning?

15. What is Deep Learning?
Deep Learning is a subfield of Machine Learning that deals with algorithms inspired
by the structure and function of the brain
Artificial Intelligence
Machine
Learning
Deep
Learning
Ability of a machine to imitate
intelligent human behavior
Application of AI that allows a system
to automatically learn and improve
from experience
Application of Machine Learning that
uses complex algorithms and deep
neural nets to train a model

16. Why do we need Deep Learning?
Process huge amount of data
Machine Learning algorithms work with huge amount of structured data but Deep Learning algorithms
can work with enormous amount of structured and unstructured data
Perform complex algorithms
Machine Learning algorithms cannot perform complex operations, to do that we need Deep Learning
algorithms
To achieve the best performance with large amount of data
As the amount of data increases, the performance of Machine Learning algorithms decreases, to
make sure the performance of a model is good, we need Deep Learning
Feature Extraction
Machine Learning algorithms extract patterns based on labelled sample data, while Deep Learning
algorithms take large volumes of data as input, analyze the input to extract features out of an object
and identifies similar objects

17. Applications of Deep Learning

18. Applications of Deep Learning
Cancer Detection
Deep Learning helps to detect cancerous tumors in the human body

19. Applications of Deep Learning
Robot Navigation
Deep Learning is used to train robots to perform human tasks

20. Applications of Deep Learning
Autonomous Driving Cars
Distinguishes different types of objects, people, road signs and drives
without human intervention

21. Applications of Deep Learning
Machine Translation
Given a word, phrase or a sentence in one language, automatically
translates it into another language

22. Applications of Deep Learning
Music Composition
Deep Neural Nets can be used to produce music by making computers
learn the patterns in a composition

23. Applications of Deep Learning
Colorization of images
Uses the object and their context within the photograph to color the image

24. What is a Neural Network?

25. What is a Neural Network?
Deep Learning is based on the functioning of a human brain, lets understand how does a
Biological Neural Network look like
Dendrite
Cell Nucleus
Synapse
Axon

26. What is a Neural Network?
Deep Learning is based on the functioning of a human brain, lets understand how does an
Artificial Neural Network look like
Neuron
X1
X2
Xn
Input 1
Input 2
Input n
Y Output
w1
w2
wn

27. Biological Neuron vs Artificial Neuron
Neuron
X1
X2
Xn
Input 1
Input 2
Input n
Y Output
w1
w2
wn
Biological Neuron
Cell Nucleus
Artificial Neuron
Synapse
Dendrites
Axon
Inputs
Nodes
Weights
Output
AxonCell
Nucleus
Dendrite
Synapse
Nodes

28. What is a Neural Network?
First step in the process is to calculate the weighted sum of the inputs and add a bias
X1
X2
Xn
Input 1
Input 2
Input n
w1
w2
w
Step 1
i=1
n
w x + b
i i*
n
Transfer Function

29. What is a Neural Network?
Second step in the process is to pass the calculated weighted sum as input to the activation
function to generate the output
X1
X2
Xn
Input 1
Input 2
Input n
Y
Output
w1
w2
w
Step 1
i=1
n
w x + b
i i*
n
Step 2
Activation FunctionTransfer Function

30. Activation Functions
An Activation function takes the “weighted sum of input plus the bias” as the input to the
function and decides whether it should be fired or not
Types of Activation
Functions
Sigmoid Function
Threshold Function
ReLU Function
Hyperbolic Tangent
Function

31. Activation Functions
Sigmoid Function
Used for models where we have to predict the probability as an output. It exists between 0 and
1.
i=1
n
w x + b
i i*
0
1
Y
(X)=
1
1 + e-x

32. Threshold Function
Activation Functions
It is a threshold based activation function. If Y value is greater than a certain value, the function
is activated and fired else not.
i=1
n
w x + b
i i*
0
1
Y
(X)=
1, if x>=0
0, if x<0( )

33. ReLU Function
Activation Functions
It is the most widely used Activation function and gives an output of X if X is positive and 0
otherwise
i=1
n
w x + b
i i*
0
1
Y
(X) = max(X,0)

34. Hyperbolic Tangent
Function
Activation Functions
This function is similar to Sigmoid function and is bound to range (-1, 1)
0
1
Y
(X)=
1 - e
1 + e-2x
i=1
n
w x + b
i i*
-
1
-2x

35. Activation Functions
X
X
X
1
n
2 Y
Output
Input 1
Input 2
Input n
Input Layer Output LayerHidden Layer

36. Working of a Neural Network

37. Working of a Neural Network
Lets understand how a neural network classifies the images of Dogs and Cats
How to identify
Dogs and Cats?
Feed the images of Dogs and Cats to the
neural network as input
Artificial Neural Network

38. Working of a Neural Network
Lets understand how a neural network classifies the images of Dogs and Cats
X
X
1
n
X2
Y^
Classifies Dogs and Cats
Cats
Dogs
Hidden Layer Hidden Layer

39. Working of a Neural Network
Lets consider a simple neural network
X
X
Y
Y1
2
n
w
1
wn
i=1
m
w x
i i*( ) ^ Model Output
Actual Output
Bias
Forwardpropagation
Compare predicted
output with actual outputX2
w2

40. Working of a Neural Network
After training the Neural Network, it uses Backpropagation method to improve the performance of the
network. Cost Function helps to reduce the error rate.
X
X
X
Y
Y1
2
n
w
1
wn
i=1
m
w x
i i*( ) ^ Model Output
Actual Output
C = ½( Y – Y )^
2
Cost Function:
w2
Bias

41. Cost Function
The Cost value is actually the difference between the neural nets predicted output and the actual output from a
set of labelled training data. The least cost value is obtained by making adjustments to the weights and biases
iteratively throughout the training process.
X
X
Y
Y1
n
w
1
w
w2
n
i=1
m
w x
i i*( ) ^ Model Output
Actual Output
C = ½( Y – Y )^
2
Cost Function:
Bias
X2

42. Why are Deep Neural Nets hard to train?
Training
Gradient is the rate at which cost changes with
respect to weight and bias
Until 2006, there was no proper method to accurately train
deep neural networks due to a basic problem with the
training process:
The Vanishing Gradient

43. Why are Deep Neural Nets hard to train?
Let’s understand Gradient like a slope and the training process like Rolling a ball
Slower Faster
The ball will roll slower if the slope is gentle and will roll faster if the slope is steep. Likewise, a Neural Net will
train slowly if the Gradient is small and it will train quickly if the Gradient is large.

44. Gradient Descent
Gradient Descent is an optimization algorithm for finding the minimum of a function
C = ½( Y – Y )^ 2
C
W
C
W
Best

45. Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Cats
DogsW1
W2
W3
W4
W5
W6
W7
W8
W9
w10
W11
W12
W13
W14
W15
W16
W17
W18
W19
W20
W21
W22
W23
W24
W25
W26
Applying the weights to each interconnection

46. Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Cats
Dogs
Applying the activation functions to the hidden layers to
decide which nodes to fire and carry out feature
extraction

47. Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Cats
Dogs
Applying the activation functions to the hidden layers to
decide which nodes to fire and carry out feature
extraction

48. Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Y
Actual Output
Comparing the predicted output to actual output

49. Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Y
Actual Output
Cost Function:
2
C = ½( Y – Y )
Applying the cost function to minimize the difference
between predicted and actual output using gradient
descent algorithm

50. Neural Network Prediction
Backpropagation
2
C = ½( Y – Y )
X
X
1
3
Y^
Adjusting the weights and biases using backpropagation
method to improve the model
X2
Y
Actual Output
Cost Function:

51. W23
W24
W25
W26
Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Applying the updated weights and biases to
calculate the cost value in order to improve the
prediction rate
Cost Function:
2
C = ½( Y – Y )
W1
W2
W3
W4
W5
W6
W7
W8
W9
w10
W11
W12
W13
W14
W15
W16
W17
W18
W19
W20
W21
W22

52. Neural Network Prediction
Forwardpropagation
X
X
X
1
n
2
Y^
Cats
Dogs
Classifying the images based on the extracted features

53. Neural Network Prediction