THEORY EXAMINATION (SEM–VI) 2016-17 ARTIFICIAL NEURAL NETWORK

ARTIFICIAL NEURAL NETWORK (NEC013)

Time: 3 Hours  Max Marks: 100

SECTION – A (Short Answer Questions)

(10 × 2 = 20 Marks)

(a) Neural Computing

Neural computing is a computing paradigm inspired by the human brain, where information is processed using interconnected artificial neurons that learn from data.

(b) BNN and ANN

BNN (Biological Neural Network): Natural network of neurons in the human brain.

ANN (Artificial Neural Network): Mathematical model inspired by BNN, implemented using algorithms and software.

(c) ADALINE model

ADALINE (Adaptive Linear Neuron) is a single-layer neural network that uses a linear activation function and updates weights using the least mean square (LMS) rule.

(d) ART models

ART (Adaptive Resonance Theory) models are neural networks used for pattern clustering that maintain stability while learning new patterns without forgetting old ones.

(e) Boltzmann learning

Boltzmann learning is a stochastic learning rule where neurons update states probabilistically to minimize network energy.

(f) Linear Associative Network (LAN)

LAN is a network that associates input patterns with output patterns using linear mapping, commonly used in memory models.

(g) Number of hidden nodes

The number of hidden nodes depends on problem complexity. Too few cause underfitting; too many cause overfitting. There is no fixed rule, usually decided experimentally.

(h) Pattern association

Pattern association is the task of mapping an input pattern to a corresponding output pattern, such as input–output recall.

(i) Network inversion

Network inversion is the process of determining the input that produces a given output in a neural network.

(j) Components of CL network

CL (Competitive Learning) network components include:

Input layer                            Competitive layer                     Winner-take-all mechanism

Used for pattern clustering and feature extraction.

SECTION – B (Long Answer Questions)

(Attempt any FIVE – 5 × 10 = 50 Marks)

2(a) Full Counter Propagation Network (Full CPN)

Architecture:                                                   Full CPN consists of:

Input layer

Kohonen (competitive) layer                            Grossberg (output) layer

Training Phases:

Unsupervised learning in Kohonen layer (clustering)

Supervised learning in Grossberg layer (association)

Used in pattern classification and mapping.

2(b) Biological neuron and neuron models

Biological neuron parts:                                Dendrites

Cell body                                                          Axon

Synapse

Neuron models:                                             McCulloch–Pitts neuron

ADALINE                                                          Sigmoid neuron

Each model simplifies biological behavior mathematically.

2(c) Types of learning – Hebbian and Boltzmann

Types of learning:

Supervised                                                       Unsupervised

Reinforcement

Hebbian learning:
“Neurons that fire together wire together.” Weight increases if both input and output are active.

Boltzmann learning:
Uses probabilistic neuron activation and simulated annealing to reach minimum energy state.

2(d) RBF network for pattern classification

In pattern classification, data clusters around centers. RBF networks use radial basis functions (usually Gaussian) centered at these clusters.

Basis functions are decided by:              Data distribution

Distance measure                                     Cluster centers

2(e) MLP architecture and backpropagation

MLP Architecture:                                  Input layer

One or more hidden layers                      Output layer

Backpropagation:
Error is propagated backward to update weights using gradient descent:

Δw=−η∂E∂w\Delta w = -\eta \frac{\partial E}{\partial w}Δw=−η∂w∂E​

Used for nonlinear classification problems.

2(f) Recognition of consonant–vowel (CV) segments

ANNs recognize CV segments by extracting spectral and temporal features.

Texture classification:
Classifies patterns based on texture properties.

Segmentation:
Divides image or signal into meaningful regions.

2(g) Pattern association, classification, and mapping

Pattern association: Input → output recall

Pattern classification: Assign input to a class

Pattern mapping: Transform input pattern to output pattern

Example: Speech recognition systems.

2(h) Feed-forward vs Feed-back networks

Stochastic networks: Use randomness            Simulated annealing: Gradual reduction of randomness
Boltzmann machine: Energy-based stochastic network

SECTION – C (Very Long Answer Questions)

(Attempt any TWO – 2 × 15 = 30 Marks)

3(a) Hopfield network – storage and recall algorithm

Storage algorithm:                                         Weights are computed using:

wij=∑xixjw_{ij} = \sum x_i x_jwij​=∑xi​xj​

Recall algorithm:                                             Initialize with input pattern

Update neurons iteratively                                Network converges to stored pattern

Hopfield network acts as content-addressable memory.

3(b) AND, OR, XOR using MP neurons

AND & OR can be implemented using single-layer perceptrons.

XOR problem:
XOR is not linearly separable, so it cannot be solved by a single-layer perceptron.

Solution:
Use Multilayer Perceptron (MLP) with hidden layer.

4(a) Self-Organizing Maps (SOM)

SOM maps high-dimensional data to low-dimensional grids.

Training steps:

Initialize weights                          Find Best Matching Unit (BMU)             Update neighborhood weights

Applications:

Data compression                        Visualization                                           Clustering

4(b) ART networks

ART networks perform stable pattern clustering.

Features:

Plasticity                                       Stability                                                  Vigilance parameter

Advantages:

No catastrophic forgetting          Online learning

5(a–c) Short Notes

(a) Principal Component Analysis (PCA)

PCA reduces dimensionality by transforming data into uncorrelated components.

(b) Vector Quantization (VQ)

VQ represents large datasets using a small set of representative vectors.

(c) Mexican Hat Networks

These networks use excitation at center and inhibition around, useful in feature detection.