Lovely Professional University

Department of Computer Science & Engineering

Presentation on
INT254-Foundations of Machine Learning

Presented By:
Dr. Premananda Sahu
Department of Machine Learning 1
Topics to be covered…
• Soft Computing: What is Soft Computing?
• Comparison with Hard Computing
• Components of Soft Computing
• Neural Network: Biological Motivation and History
• Applications of Neural Networks
• Structure and Working of Artificial Neurons
• Perceptron Learning Algorithm
• Architecture and Working of Feedforward Neural
Networks
• Activation functions
• Case study: Practical Problem-solving using Neural
2
Networks
What is Soft Computing

• Fusion of Fuzzy Logic, Neuro Computing & Evolutionary

Computing

• Designed to enable solutions to real world problems

• Solutions which are difficult to model mathematically

SC
Soft = EC + NN
Evolutionary +
Neural FLFuzzy
Computi Computing Network Logic
ng

3
Soft Computing Vs Hard Computing
Hard Computing Soft Computing

Precisely analytical model Approximate model

Exact algorithms Clever algorithms (Heuristic)

Exact input data Ambiguous/noisy input data

Precision and categoricity Approximation

Rigidly written programs Trainable and adaptive

Precise answer Approximate answer

4
Components of Soft Computing

5
Neural Network: Biological Motivation

6
 Neural Network: Biological Neuron

•Inputs (red arrow) represent signals received from other neurons.

•Dendrites: Branch-like structures that receive input signals.

•Cell Body (Soma): The central part of the neuron that processes inputs.

•Nucleus: The core of the neuron, containing genetic material.

•Axon: The extended structure that transmits signals to other neurons.

•Outputs (red arrow) are signals sent to other neurons

7
•Inputs
Neural
(x₀, x₁,Network: Artificial
x₂, ..., xₙ): Represent Neuron
numerical values fed(perceptron)
into the artificial neuron.
•Weights (w₁, w₂, ..., wₙ): Each input is multiplied by a corresponding weight, which
determines its importance.
•Bias (b): A constant value added to the weighted sum to adjust the output.
•Bias: type of error occurs due to wrong assumptions
•Activation Function (f): allowing network to learn complex patterns.
•Activation Function (f): Applied to the linear output to introduce non-linearity
•Output (y): The final processed value, determining the neuron’s response.

8
Comparison of Biological Neuron and Artificial Neuron

9
Neural Network
•Neural networks mimic biological brain functions.
•Consist of layers, neurons, and connections.
•Each neuron applies weights and biases.
•Activation functions introduce non-linear properties.
•Outputs help in decision-making processes.
•Process input data using mathematical transformations.

10
Neural Network: History

11
Applications of Neural Networks

12
Structure and Working of Artificial
Neurons

Consist of 3 main parts: input processing, activation function, and weight update (learning 13
Structure and Working of Artificial
Neurons

14
Structure and Working of Artificial
Neurons

15
Perceptron Learning Algorithm
• Fundamental supervised learning algorithm used for binary classification problems.
• Based on adjusting weights iteratively to correctly classify data points.

16
Perceptron Learning Algorithm

• Hyperparameters are variables that control the training of machine learning models.
• Learning rate (η) is crucial hyperparameter in machine learning that controls how much the model's
weights are updated during training.
• η is too high → The model converges (produce accurate predictions) too fast
• η is too low → The model learns too slowly 17
Perceptron Learning Algorithm-: Example
Calculation

18
Perceptron Learning Algorithm-: Example
Calculation

19
Perceptron Learning Algorithm-: Example
Calculation

20
Perceptron Learning Algorithm-: Example
Calculation

21
Architecture and Working of Feedforward Neural Networks

• ANN is mathematical model that can learn and used in ML.

• Feedforward Neural Network (FNN) is the simplest type of ANN where information flows only in one

direction.

• Flows from I/P to O/P without looping.

• FNN consists of 3 layers:

• I/P layer, Hidden layer, O/P layer

• loss function measures how well the neural network's predictions match the actual target values.
22
Architecture and Working of Feedforward Neural Networks

• Input Layer: Receives raw data or features (𝑥1,𝑥2,...,𝑥𝑛​).

• No computations are performed in this layer.

• Hidden layer : Contains multiple neurons that process input data.

• Each neuron applies weighted sum of inputs followed by an activation function.

• Output Layer : The number of neurons depends on the type of problem

• Binary classification: 1 neuron (sigmoid activation).

• Multi-class classification: n neurons (softmax activation).

• Regression: 1 neuron (linear activation).

23
Architecture and Working of Feedforward Neural Networks

24
Activation Functions

25
Case study: Practical Problem-solving using Neural
Networks

26
Case study: Practical Problem-solving using Neural
Networks

(a) Neural Network Structure for Digit Recognition

• Represents a feedforward neural network used for handwritten digit recognition.
•Input layer consists of pixel values from a grayscale (shades of gray, without any color)
image of the digit "3."
•Hidden layer performs weighted summation and applies an activation function (σ).
•The output layer has 10 neurons (0-9), where the neuron corresponding to "3" has the highest
activation.
•Key Concept: The weighted sum of inputs and weights is computed using a dot product,
followed by an activation function to introduce non-linearity. 27
Case study: Practical Problem-solving using Neural
Networks
(b) Optical Computation of Neural Network Operations
•Explains how optical components can perform vector-matrix multiplication, a key operation in neural
networks.
•Step 1: Element-wise multiplication
• Each input value xi​is multiplied by the corresponding weight wij.
•Step 2: Summation of products
• The results are summed to obtain the neuron activation y 1​.
•Optical Interpretation
• A light source represents input data.
• A modulator applies weights to the input data in a spatially parallel manner.
• A bucket detector aggregates the weighted signals, simulating a neuron summation.
28
Case study: Practical Problem-solving using Neural
Networks

29