About this previous year paper
SECTION A — Very Short Answers (2 Marks Each)
a. Loss Function
A loss function measures the difference between predicted output and true output and guides weight updates during training.
b. Stochastic Gradient Descent (SGD)
An optimization algorithm that updates model parameters using one or a small batch of training samples at a time.
c. AI Winter
A period of reduced funding and interest in AI due to limited computational power and unmet expectations.
d. Probabilistic Theory of Deep Learning
Views deep learning models as probabilistic systems that learn data distributions and uncertainty.
e. Linear Discriminant Analysis (LDA)
A supervised dimensionality-reduction technique that maximizes class separability.
f. Distance Metrics in ML
Used to measure similarity between data points
Examples: Euclidean, Manhattan, Cosine distance.
g. Non-Convex Optimization
Optimization problems with multiple local minima and saddle points, common in deep networks.
h. Generalization
The ability of a trained model to perform well on unseen data.
i. WaveNet
A deep generative model for raw audio waveform generation, known for high-quality speech synthesis.
j. Word2Vec
A technique that learns vector representations of words capturing semantic relationships.
SECTION B — Short Descriptive Answers (10 Marks)
a. Activation Functions (ReLU & Sigmoid)
Sigmoid: Outputs values between 0 and 1, used in binary classification
ReLU: Outputs max(0, x), faster convergence and avoids vanishing gradient
b. GANs: Generator & Discriminator
Generator: Creates fake data Discriminator: Distinguishes real vs fake data
Both trained in an adversarial manner
c. AlexNet Significance Won ImageNet 2012
Introduced ReLU, dropout, GPU training Sparked deep learning revolution
d. Spatial Transformer Networks (STNs) Allow networks to learn spatial transformations
Improve spatial invariance
e. Deep Learning in Bioinformatics Gene sequencing
Protein structure prediction Disease diagnosis
SECTION C — Optimization & Training (10 Marks)
a. Backpropagation Computes gradients using chain rule
Updates weights to minimize loss Enables deep network training
b. Role of SGD in Optimization Faster convergence
Handles large datasets Introduces noise that helps escape local minima
SECTION D — CNN & Normalization (10 Marks)
a. CNN for Image Classification Convolution layers extract features
Pooling reduces dimensionality Fully connected layers perform classification
b. Batch Normalization Normalizes layer inputs
Reduces internal covariate shift Speeds up training and improves stability
SECTION E — Initialization & Deep Architectures (10 Marks)
a. Xavier vs He Initialization Xavier: Suitable for Sigmoid/Tanh
He: Suitable for ReLU Prevents exploding/vanishing gradients
b. Residual Connections (ResNet) Skip connections allow gradient flow
Enable very deep networks Reduce degradation problem