Name:-Ravindra

Name: -AbhishekYadav
Yadav
No.:-33
Roll No.: -06
Assignment ::-2
-02
Subject:- SOFT COMPUTING

QUES 1:- Define Artificial neuron and its

models.
Ans 1:- Definition of Artificial Neuron:
An artificial neuron is the basic unit of an
artificial neural network (ANN), inspired by
the biological neuron in the human brain. It is
a mathematical function that takes one or
more inputs, processes them, and produces
an output. The artificial neuron mimics how
biological neurons receive signals, process
them, and transmit the result to other
neurons.
Basic Structure of an Artificial Neuron:
An artificial neuron typically includes:
1. Inputs (x₁, x₂, ..., xₙ): These represent
the input signals.
2. Weights (w₁, w₂, ..., wₙ): Each input is
assigned a weight that determines its
importance.
3. Summation Function: It computes the
weighted sum of the inputs:
where b is the bias term that allows shifting
the activation function.
4. Activation Function (ϕ): It introduces
non-linearity and determines the output
based on the summation:

Models of Artificial Neuron:

1. McCulloch-Pitts Model (1943):
o One of the earliest models of a
neuron.
 o Binary model: outputs either 0 or 1.
o Activation is based on a threshold:
o No learning mechanism.
2. Perceptron Model (1958):
o Introduced by Frank Rosenblatt.
o Can learn through training (supervised
learning).
o Uses a threshold activation function or
step function.
o Capable of solving linearly separable
problems.
3. Sigmoid Neuron:
o Uses the sigmoid activation function:
o Output is continuous between 0 and
1.
 o Suitable for probabilistic
interpretations and gradient-based
optimization.
4. Tanh Neuron:
o Uses the hyperbolic tangent function:
o Output range is from -1 to 1.
o Often used in hidden layers for better
performance than sigmoid.
5. RLU (Rectified Linear Unit) Neuron:
o Activation function:
o Introduces non-linearity and helps
with vanishing gradient problems.
o Popular in deep learning models.
QUES 2:- Explain various learning techniques.
ANS 2:- Various Learning Techniques in Soft
Computing
Soft computing is a collection of
computational techniques that deal with
approximation and uncertainty to solve
complex real-life problems. In the context of
soft computing, learning techniques are
methods that enable systems (like neural
networks, fuzzy systems, or evolutionary
algorithms) to learn from data and improve
their performance over time.
Here are the main learning techniques in soft
computing:

1. Supervised Learning
• Definition: Learning from labeled data,
where the input-output pairs are known.
• Process: The system is trained using a
dataset that includes input features and
the correct output. The model makes
 predictions and is corrected by comparing
with actual outputs.
• Goal: Minimize the error between
predicted and actual outputs.
• Examples:
o Artificial Neural Networks (ANNs)
o Support Vector Machines (SVM)
• Applications: Classification, regression,
speech recognition.

2. Unsupervised Learning
• Definition: Learning from data that has no
labels.
• Process: The system tries to find patterns,
structures, or relationships within the
input data without any output
supervision.
 • Goal: Group or cluster similar data points.
• Examples:
o K-Means Clustering
o Self-Organizing Maps (SOMs)
• Applications: Market segmentation,
anomaly detection, pattern recognition.

3. Reinforcement Learning
• Definition: Learning through interaction
with an environment to achieve a goal.
• Process: An agent performs actions in an
environment, receives feedback in the
form of rewards or penalties, and learns
the optimal policy.
• Goal: Maximize cumulative reward over
time.
• Examples:
 o Q-Learning
o Temporal Difference Learning
• Applications: Robotics, game playing,
autonomous systems.

4. Evolutionary Learning
• Definition: Learning based on the
principles of natural evolution such as
selection, mutation, and crossover.
• Process: A population of solutions evolves
over time to find the best one.
• Goal: Optimize a given fitness function.
• Examples:
o Genetic Algorithms (GA)
o Genetic Programming (GP)
 • Applications: Optimization problems,
feature selection, machine learning model
design.

5. Hebbian Learning
• Definition: Based on the biological
principle: “Neurons that fire together,
wire together.”
• Process: Strengthens the connection
between two neurons if they are activated
simultaneously.
• Goal: Enhance correlation between inputs
and outputs.
• Applications: Pattern recognition,
associative memory.

6. Competitive Learning
 • Definition: Neurons compete to become
active; only one or few neurons "win" and
get updated.
• Process: Helps in clustering and
categorization.
• Goal: Group similar input patterns.
• Examples: Self-Organizing Maps (Kohonen
Networks)
• Applications: Data compression,
clustering.
QUES 3:- Compare and contrast between
ADALINE v/s MADALINE.
ANS 3:- ADALINE (Adaptive Linear
Neuron)
Aspect Description
Full Form Adaptive Linear Neuron
Developer Bernard Widrow and Marcian
Aspect Description
Hoff (1960)
Single-layer neural network
Structure
with a single output neuron
Takes multiple inputs and
Input Type
produces one output
Uses the Least Mean Square
Learning
(LMS) or delta rule for weight
Algorithm
updates
Linear in training (raw output),
Activation
threshold applied for final
Function
output
Binary classification (linearly
Use Case
separable problems)
Cannot solve non-linearly
Limitation
separable problems (like XOR)
Output Type Binary (after applying
Aspect Description
threshold)

MADALINE (Multiple ADALINE)

Aspect Description
Many ADALINE or Multiple
Full Form
ADALINE
Bernard Widrow and colleagues
Developer
(1962)
Multi-layer neural network
Structure composed of multiple ADALINE
units in layers
Takes inputs and passes through
Input Type
multiple layers of ADALINEs
Learning Uses MADALINE Rule I, II, or III
Algorithm (a heuristic method) for weight
Aspect Description
updates
Activation
Step function (sign function)
Function
Training Data Supervised learning
Can handle non-linearly
Use Case
separable problems
Complex training process
Limitation
compared to ADALINE
Output Type Binary (±1 typically)

Key Differences:
Feature ADALINE MADALINE
Architecture Single-layer Multi-layer
Complexity Simple More complex
Problem Linearly Can solve non-
Feature ADALINE MADALINE
Solving separable linearly
problems only separable
problems
Learning Delta rule MADALINE rules
Algorithm (LMS) (heuristic)
Output
Single Multiple
Neurons

Analogy:
• ADALINE is like a single opinion based on a
few facts.
• MADALINE is like a team of experts
collaborating to form a better decision.
QUES4:- Discuss about error back propagation
algorithm (EBPA) and its characteristic.
ANS 4:- Error Backpropagation Algorithm
(EBPA)
The Error Backpropagation Algorithm is a
supervised learning algorithm used for
training multilayer feedforward neural
networks (especially Multilayer MLPs). It’s
one of the most widely used algorithms in
artificial neural networks.

Working of Error Backpropagation

Algorithm:
The learning process occurs in two main
phases:
1. Forward Pass:
• Input data is passed through the network
layer by layer.
• Each neuron computes a weighted sum of
its inputs and passes the result through an
 activation function (like sigmoid, tanh,
RLU).
• The output is compared with the target
output to compute the error.
2. Backward Pass (Backpropagation):
• The error is propagated backwards from
the output layer to the input layer.
• Gradients of the error with respect to the
weights are calculated using the chain rule
of calculus.
• Weights are updated to minimize the
error using gradient descent:

Characteristics of EBPA:
Characteristic Description
Type Supervised Learning
Architecture Requires multilayer
Characteristic Description
Requirement feedforward network
(minimum 1 hidden layer)
Activation Must be differentiable (e.g.,
Function sigmoid, tanh, RLU)
Typically Mean Squared Error
Error Function
(MSE)
Learning Rule Gradient Descent
Training
Iterative (epoch-based)
Method
Weight Update Based on gradient of the error
May take many iterations;
Convergence
sensitive to learning rate
Can occur with too many
Overfitting
layers or training for too long
Suitable for deep networks
Scalability
but computationally
Characteristic Description
expensive
Can be improved using
techniques like momentum,
Optimization learning rate decay, batch
training, or adaptive methods
(e.g., Adam, RMSprop)

Steps in EBPA:
1. Initialize weights and biases randomly.
2. Feedforward the input to compute the
output.
3. Compute the error at the output layer.
4. Backpropagate the error to calculate
gradients.
5. Update weights and biases using
gradient descent.
6. Repeat steps 2–5 for many iterations
(epochs) until the error is minimized.

Advantages:
• Can learn complex non-linear mappings.
• Widely used in practical applications like
image recognition, speech processing, etc.
• Forms the backbone of deep learning.
Limitations:
• Prone to local minima.
• Slow convergence if not optimized
properly.
• Requires large datasets for good
generalization.
• Sensitive to initial weight values and
hyperparameters.
QUES 5:- Explain architecture of Adaptive
Resonance Theory (ART).
ANS 5:- Architecture of Adaptive
Resonance Theory (ART)
Adaptive Resonance Theory (ART) is a neural
network model developed by Stephen
Grossberg and Gail Carpenter in the 1980s. It
was designed to address the stability-
plasticity dilemma — the challenge of
learning new information (plasticity) without
forgetting previously learned knowledge
(stability).
ART networks are mainly used for pattern
recognition and clustering, especially when
the data is noisy or presented in an
online/real-time manner.

Basic Architecture of ART:

The ART architecture consists of three main
components:
1. F1 Layer (Comparison Layer):
o Divided into two sublayers: F1a and
F1b.
o F1a receives input patterns.
o F1b interacts with F2 and sends
feedback to F1a.
o It is responsible for comparing the
input with stored patterns from F2.
2. F2 Layer (Recognition Layer):
o Contains category neurons.
o Each neuron represents a learned
pattern or class.
o It competes to match the input
pattern based on similarity.
3. Gain Control and Reset Mechanism:
 o If the best match from F2 doesn’t
meet a predefined vigilance threshold
(ρ), a reset signal is sent.
o That neuron is disabled for the current
input, and the next best match is
tested.
o This continues until a neuron matches
within the acceptable vigilance limit,
ensuring stable learning.
4. Vigilance Parameter (ρ):
o Controls the degree of similarity
required to accept a new pattern into
an existing category.
o Higher ρ → more strict → more
categories formed.
o Lower ρ → more generalization →
fewer categories.
 Key Characteristics of ART Architecture:
Feature Description
Type of Unsupervised / Semi-
Learning supervised
Memory Stable – previously learned
Stability patterns are not forgotten
Can learn new patterns without
Plasticity
retraining
Matching Top-down expectations vs
Process bottom-up input
Ensures pattern similarity
Vigilance Test
before assigning to category
Performs online clustering of
Clustering
input patterns
Pattern recognition, speech
Applications recognition, classification, data
mining
 Types of ART Models:
• ART1 – for binary input patterns.
• ART2 – for continuous-valued input
patterns.
• ART3 – includes biological mechanisms
(e.g., neurotransmitter dynamics).
• Fuzzy ART – combines ART with fuzzy logic
for more flexible pattern recognition.
• ARTMAP – supervised ART for predictive
classification tasks.

THANK YOU

(RAVINDRA YADAV06)
(Abhishek YADAV 33)