CS-702-B Deep and Reinforcement Learning - Jun 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. 14 Marks
   What is the significance of gradient descent in training deep learning models? Discuss various optimization algorithms including AdaGrad, RMSProp, and Adam.
   How to answer

   • Define gradient descent and loss minimization via backpropagation
   • Explain significance: convergence, learning rate, scalability to deep networks
   • Describe AdaGrad with per-parameter adaptive rates and drawback
   • Describe RMSProp and Adam with update formulas and advantages
2. 14 Marks
   Describe the challenges of vanishing and exploding gradients in Recurrent Neural Networks (RNNs). How do GRUs and LSTMs mitigate these issues?
   How to answer

   • Explain BPTT and repeated multiplication of gradients
   • Define vanishing and exploding gradients causes and effects
   • Describe LSTM gates: input, forget, output and cell state
   • Describe GRU gates and how gating preserves long-term dependencies
3. 14 Marks
   Explain the role of regularization techniques in deep learning. How do dropout and batch normalization contribute to improved model performance?
   How to answer

   • Define overfitting and need for regularization
   • Explain dropout mechanism and ensemble effect
   • Explain batch normalization: normalization steps and internal covariate shift
   • Discuss impact on generalization, convergence and performance
4. 14 Marks
   Discuss the concept of greedy layerwise pre-training. What are some better activation functions introduced in recent years?
   How to answer

   • Define greedy layerwise pre-training and unsupervised initialization
   • Explain historical role in deep belief networks and autoencoders
   • List recent activation functions: ReLU variants, ELU, GELU, Swish/Mish
   • Compare advantages: non-saturation, faster training, vanishing gradient relief
5. 14 Marks
   How do visualization techniques like Deep Dream and Guided Backpropagation help in understanding neural networks?
   How to answer

   • Define need for interpretability in neural networks
   • Explain Deep Dream: maximizing activations to visualize features
   • Explain Guided Backpropagation: saliency and feature attribution
   • Discuss how visualizations reveal hierarchy and debugging aid
6. 14 Marks
   Define the Policy Gradient method in reinforcement learning. How does it differ from Q-learning?
   How to answer

   • Define Policy Gradient: direct policy optimization and REINFORCE
   • Give objective and gradient estimator formula
   • Define Q-learning: value-based off-policy TD learning
   • Contrast: stochastic vs deterministic, continuous actions, on-policy vs off-policy
7. 14 Marks
   Compare the Actor-Critic method with traditional reinforcement learning techniques. What advantages does it provide?
   How to answer

   • Define Actor and Critic roles in Actor-Critic
   • Compare with REINFORCE, Q-learning, DQN: variance and bias
   • Explain advantage function and TD error for updates
   • List advantages: lower variance, continuous control, stability
8. 14 Marks
   Discuss recent trends in reinforcement learning architectures, including Generative Adversarial Imitation Learning.
   How to answer

   • Outline recent RL architectures: deep RL, distributional, model-based
   • Explain Generative Adversarial Imitation Learning framework
   • Describe discriminator-generator imitation from expert demos
   • Discuss applications and challenges: sample efficiency, stability