AL-503-B Deep Learning - Dec 2025 Question Paper

Under each question, How to answer lists the points an answer should cover, outlined by CampusPrep from the syllabus. Check the details against your textbook.

1. 7 Marks
   aExplain the historical evolution of deep learning and McCulloch-Pitts neuron model. Discuss the representation power of multilayer perceptrons.
   How to answer

   • Timeline of deep learning: McCulloch-Pitts, perceptron, backpropagation, deep revival
   • McCulloch-Pitts neuron model: binary threshold logic, weights and activation
   • Limitations of single-layer perceptron and need for hidden layers
   • Representation power of MLPs: universal approximation and non-linear separability
2. 7 Marks
   bExplain backpropagation algorithm with suitable mathematical formulation. Discuss weight initialization and batch normalization.
   How to answer

   • Forward pass and loss definition for multilayer network
   • Chain rule derivation of gradients and backward error propagation
   • Weight update rule and steps of backpropagation algorithm
   • Weight initialization methods and batch normalization concept and benefits
3. 7 Marks
   aExplain sigmoid neurons and feed-forward neural networks.
   How to answer

   • Sigmoid neuron model and sigmoid activation function
   • Architecture and working of feed-forward neural networks
   • Forward propagation and training principle
   • Comparison with perceptron and advantages
4. 7 Marks
   bExplain PCA and SVD and their role in representation learning.
   How to answer

   • PCA concept: dimensionality reduction via principal components
   • SVD concept: matrix factorization and relation to PCA
   • Role in feature extraction and representation learning
   • Benefits for visualization, denoising and compression
5. 7 Marks
   aCompare gradient descent variants : GD, SGD and Momentum.
   How to answer

   • Update rules for GD, SGD and Momentum
   • Comparison on convergence speed, stability and computation
   • Effect of mini-batches and momentum term
   • Advantages, disadvantages and use cases
6. 7 Marks
   bExplain autoencoders and discuss regularization techniques used in autoencoders.
   How to answer

   • Autoencoder architecture: encoder, bottleneck, decoder
   • Training objective: reconstruction loss
   • Regularization need to avoid identity mapping
   • Common regularization techniques: sparse, denoising, contractive, dropout, weight decay
7. 7 Marks
   aExplain denoising, sparse, contractive and variational autoencoders.
   How to answer

   • Denoising autoencoder: learning from corrupted inputs
   • Sparse autoencoder: sparsity penalty on hidden units
   • Contractive autoencoder: Jacobian penalty for robustness
   • Variational autoencoder: probabilistic latent space and KL divergence
8. 7 Marks
   bExplain relationship of autoencoders with PCA and SVD and role of dataset augmentation.
   How to answer

   • Linear autoencoder equivalence to PCA/SVD
   • Non-linear autoencoders as generalization of PCA
   • Concept of dataset augmentation for autoencoders
   • Role in improving generalization and robustness
9. 7 Marks
   aExplain convolution operation, kernels, stride, padding and pooling in CNNs.
   How to answer

   • Convolution operation and kernels/filters
   • Stride and padding and their effect on output size
   • Pooling types: max and average pooling
   • Role in feature extraction and translation invariance
10. 7 Marks
    bExplain architectures of VGGNet and ResNet.
    How to answer

    • VGGNet: deep stack of small 3x3 convolutions
    • ResNet: residual blocks and skip connections
    • Comparison of depth, parameters and training
    • Advantages for image classification and vanishing gradient
11. 7 Marks
    aExplain regularization techniques in deep networks such as dropout, weight decay and early stopping.
    How to answer

    • Dropout: random neuron deactivation during training
    • Weight decay: L1/L2 penalty on weights
    • Early stopping: monitoring validation loss
    • Effect on overfitting and generalization
12. 7 Marks
    bExplain visualization of CNNs, Deep Dream and Deep Art.
    How to answer

    • Methods to visualize CNN filters and feature maps
    • Deep Dream: amplifying learned features
    • Deep Art / neural style transfer concept
    • Applications in interpretability and art generation
13. 7 Marks
    aExplain BPTT, vanishing and exploding gradient problems and truncated BPTT.
    How to answer

    • BPTT algorithm for training RNNs over time steps
    • Vanishing and exploding gradient causes
    • Truncated BPTT procedure and trade-offs
    • Effect on long-term dependency learning
14. 7 Marks
    bExplain LSTM and GRU architectures and their role in solving vanishing gradient problem.
    How to answer

    • LSTM gates: input, forget, output and cell state
    • GRU gates: update and reset gates
    • How gating preserves long-term gradients
    • Comparison of LSTM vs GRU complexity and performance
15. 7 Marks
    aExplain RBMs, Gibbs sampling and deep belief networks.
    How to answer

    • RBM structure: visible and hidden layers with energy model
    • Gibbs sampling for inference and learning
    • Deep belief networks as stack of RBMs
    • Layer-wise pretraining principle
16. 7 Marks
    bExplain GANs and applications of deep learning in vision, speech and medical domains.
    How to answer

    • GAN framework: generator vs discriminator game
    • Training objective and minimax loss
    • Applications in vision, speech and medical domains
    • Examples: image synthesis, enhancement, diagnosis