CS-702-B Deep and Reinforcement Learning - Nov 2023 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
   aDraw and explain the architecture of convolutional network.
   How to answer

   • Define CNN and its purpose for grid-like data
   • Draw block architecture: input -> convolution + ReLU -> pooling -> stacked blocks -> flatten -> fully connected -> softmax/output
   • Explain role of each layer: filters, feature maps, pooling, FC
   • Mention training with backpropagation and applications
2. 7 Marks
   bWhat are Recurrent Neural Networks?
   How to answer

   • Define RNN and recurrence concept
   • Explain hidden state feedback over time steps
   • Give unfolding in time and sequence processing
   • Mention applications and vanishing gradient issue
3. 7 Marks
   aExplain LSTM (Long Short Term Memory).
   How to answer

   • Expand LSTM and purpose to handle long-term dependencies
   • Explain cell state, forget/input/output gates
   • Give gate equations and information flow
   • Mention advantages over vanilla RNN
4. 7 Marks
   bDifference between Deep and Shallow Network.
   How to answer

   • Define deep vs shallow in terms of hidden layers
   • Compare on feature learning, capacity, data need
   • Compare on training complexity and vanishing gradients
   • Give examples of each
5. 7 Marks
   aDraw and explain McCulloch Pitts neuron Model.
   How to answer

   • Define McCulloch-Pitts as first artificial neuron model
   • Draw inputs, weights, summation, threshold logic
   • Explain binary activation and threshold rule
   • Give example logic gate implementation and limitations
6. 7 Marks
   bExplain sparse and contractive auto encoders.
   How to answer

   • Define autoencoders and regularization need
   • Explain sparse autoencoder with sparsity penalty
   • Explain contractive autoencoder with Jacobian penalty
   • Compare objectives and robustness
7. 7 Marks
   aExplain Better Weight Initialization Methods.
   How to answer

   • Explain importance of initialization in deep nets
   • Describe Xavier/Glorot initialization
   • Describe He initialization for ReLU
   • Mention random, zero and pretrained considerations
8. 7 Marks
   bWhat is deep learning? Explain its uses and application and history?
   How to answer

   • Define deep learning and deep neural networks
   • Outline history: perceptron, backprop, CNN, deep revival
   • Explain uses and applications: vision, NLP, RL
   • Mention frameworks and data/compute drivers
9. 7 Marks
   aWhat is semi - supervised learning ?
   How to answer

   • Define semi-supervised learning with labeled + unlabeled data
   • Explain assumptions: smoothness, cluster, manifold
   • Give typical approaches: self-training, co-training
   • Mention applications when labeling is costly
10. 7 Marks
    bWhat is PCA (Principle Component Analysis ) and RNN?
    How to answer

    • Define PCA for dimensionality reduction via principal components
    • Explain steps: covariance, eigenvectors, projection
    • Define RNN for sequential data with recurrence
    • Contrast unsupervised reduction vs supervised sequence model
11. 7 Marks
    aExplain Policy Gradient Algorithm for Full RL.
    How to answer

    • Define policy gradient for full RL / MDP
    • Explain objective J(theta) and REINFORCE update
    • Give gradient estimator and Monte-Carlo rollout steps
    • Mention baseline and variance reduction
12. 7 Marks
    bDiscuss Bandit Algorithms in details.
    How to answer

    • Define multi-armed bandit problem
    • Explain exploration vs exploitation
    • Discuss epsilon-greedy, UCB, Thompson sampling
    • Mention regret and applications
13. 7 Marks
    aExplain the Q Function and Q Learning Algorithm.
    How to answer

    • Define Q-function Q(s,a) and Bellman optimality
    • Explain Q-learning as off-policy TD control
    • Give update rule with learning rate and discount
    • Mention epsilon-greedy and convergence
14. 7 Marks
    bExplain value iteration, policy iteration and Temporal Difference Learning.
    How to answer

    • Define MDP planning vs learning setting
    • Explain value iteration with Bellman backup
    • Explain policy iteration: evaluation + improvement
    • Explain TD learning combining MC and DP
15. 14 Marks
    Write short notes:

    1. a)Face Recognition Application
    2. b)Actor-Critic Method
    3. c)Group Normalization
    How to answer

    • Face recognition: CNN pipeline, embeddings, loss
    • Actor-Critic: policy + value combination, advantage
    • Group Normalization: channel grouping, batch independence
    • Give applications and advantages for each