About this previous year paper
NEURAL NETWORKS AND FUZZY SYSTEM
Section-wise Solved Answers & Notes (NEE013)
SECTION – A (10 × 2 = 20 Marks)
Short & precise answers
(a) Use of bias weight in artificial neuron
Bias shifts the activation function left or right. It helps a neuron fire even when inputs are zero and improves learning flexibility.
(b) Structural organization of biological neural system
Stimulus → Receptors → Sensory Neurons → Brain (Processing) → Motor Neurons → Effectors
(c) Hebbian Learning Rule
Hebbian rule states:
If two neurons are activated simultaneously, the connection weight between them increases.
Mathematically:
Δw = ηxy
(d) Auto-associative vs Hetero-associative Memory
| Auto-Associative | Hetero-Associative |
|---|---|
| Input = Output | Input ≠ Output |
| Pattern completion | Pattern mapping |
| Example: Hopfield Net | Example: BAM |
(e) Relation between Neural Networks & Machine Learning
Neural networks are a subset of machine learning used for pattern recognition, prediction, and classification through learning from data.
(f) Fuzzy Intersection & Bounded Difference
Given:
A = {(2,1), (4,0.3), (6,0.5), (8,0.2)} B = {(2,0.5), (4,0.4), (6,0.1), (8,1)}
Fuzzy Intersection (min):
{(2,0.5), (4,0.3), (6,0.1), (8,0.2)}
Bounded Difference (A − B):
max(0, μA − μB) {(2,0.5), (4,0), (6,0.4), (8,0)}
(g) Delta Rule vs Gradient Descent
| Delta Rule | Gradient Descent |
|---|---|
| Single-layer | Multi-layer |
| Linear units | Non-linear |
| Simple update | Iterative optimization |
(h) ANN vs Conventional Computing
| ANN | Conventional |
|---|---|
| Parallel | Sequential |
| Learns from data | Programmed |
| Fault tolerant | Not tolerant |
(i) Learning & its types
Learning is adjustment of weights based on experience.
• Supervised: Target output given • Unsupervised: No target output
(j) Crisp relation vs Fuzzy logic
Crisp logic has true/false (0 or 1), while fuzzy logic allows degrees of truth (0 to 1).
SECTION – B (Any 5 × 10 = 50 Marks)
Answer outlines (write in detail in exam)
(a) Back Propagation Algorithm • Initialize weights randomly
• Forward pass → compute output • Calculate error
• Backward pass → update weights • Adjust learning rate
• Repeat till error is minimum
Error correction uses gradient descent to minimize error.
(b) Hebbian Learning & AND Gate
Hebbian learning strengthens weights when input and output are same.
For bipolar AND gate:
Weights chosen so neuron fires only when both inputs are +1.
(c) Multilayer Feed Forward Network
• Input layer → Hidden layer(s) → Output layer • No feedback loops
• Used in classification
Difference from recurrent networks: Recurrent networks have feedback and memory.
(d) Defuzzification
Converts fuzzy output to crisp value.
• Centroid Method: Center of area • Weighted Average: Weighted mean
• Center of Largest Area: Midpoint of largest membership
(e) Fuzzy Inference System (FIS)
Steps: Fuzzification
Rule evaluation Aggregation
Defuzzification Used in control systems.
(f) Linguistic Variables & Relation R Fabrics → Dirt → Detergent used
Example:
Cotton + Very dirty → High detergent Silk + Less dirty → Low detergent
(g) Learning Techniques in NN
• Supervised • Unsupervised
• Reinforcement
Momentum factor speeds convergence and avoids local minima.
(h) Activation Functions
• Step • Sigmoid
• Tanh • ReLU
They decide neuron output and non-linearity.
SECTION – C (Any 2 × 15 = 30 Marks)
Q3 Short Notes (Any three)
(i) Linear Separability Perceptron works only when data is linearly separable.
(ii) LR-Type Fuzzy Numbers Defined by left and right membership functions.
(iii) Max-Min Composition Used to combine fuzzy relations.
(iv) Rosenblatt’s Perceptron Single-layer classifier with adjustable weights.
(v) Fuzzy Entropy Theorem Measures fuzziness/uncertainty in fuzzy sets.
Q4 Fuzzy Set Operations
• Union (max) • Intersection (min)
• Complement (1 − μ)
Explain properties: commutativity, associativity, idempotency.