Deep Learning (DL) SPPU Endsem Question Paper May 2025

Questions

1. Q1

   1. a)

      Explain the basic architecture of a Convolutional Neural Network (CNN) and describe the role of each key component within the architecture.

      [9]
   2. b)

      Define padding and strides in CNNs, and explain their impact on the output size of feature maps in convolutional layers.

      [8]
   OR

2. Q2

   1. a)

      Discuss the role of the ReLU layer in CNNs. How does ReLU differ from other activation functions, and why is it preferred in convolutional layers?

      [9]
   2. b)

      Describe the concept of pooling in CNNs and explain how it contributes to spatial dimensionality reduction. What are some common pooling methods, and how do they differ?

      [8]

3. Q3

   1. a)

      Explain the structure of Recurrent Neural Networks (RNNs) and the concept of unfolding computational graphs. How do RNNs handle sequential data, and what challenges arise with long-term dependencies?

      [9]
   2. b)

      Discuss the Bidirectional RNN and Encoder-Decoder Sequence-to- Sequence Architecture. How do these architectures enhance the ability of RNNs to handle complex sequential tasks? 2

      [9]
   OR

4. Q4

   1. a)

      Outline practical methodologies for evaluating the performance of recurrent neural networks (RNNs). Describe commonly used performance metrics, baseline models, and strategies for selecting hyperparameters in RNN- based models.

      [9]
   2. b)

      Describe the role of Long Short-Term Memory (LSTM) networks and other gated RNNs in addressing the limitations of standard RNNs. How do LSTM networks manage long-term dependencies, and what are some alternative strategies for handling multiple time scales?

      [9]

5. Q5

   1. a)

      What is a deep generative model, and how does it differ from discriminative models? Provide examples of deep generative models and their applications.

      [9]
   2. b)

      Describe the structure and functioning of a Boltzmann Machine (BM). What is the concept of energy in a Boltzmann Machine, and how does it guide the learning

      [8]
   OR

6. Q6

   1. a)

      What is a Generative Adversarial Network (GAN), and how do the generator and discriminator networks interact in the GAN framework? Describe their roles and training process.

      [9]
   2. b)

      Discuss the various types of GAN architectures and provide examples of applications where GANs have proven effective.

      [8]

7. Q7

   1. a)

      What is deep reinforcement learning, and how does it extend traditional reinforcement learning techniques? Discuss its significance in AI and real- world applications.

      [6]
   2. b)

      Explain the concept of a Markov Decision Process (MDP) and describe its components. How does an MDP serve as a foundation for reinforcement learning?

      [6]
   3. c)

      Describe the basic framework of reinforcement learning, including key concepts such as agents, environments, actions, states, rewards, and policies. 3

      [6]
   OR

8. Q8

   1. a)

      What are the main challenges faced in reinforcement learning, and what techniques are commonly used to address them?

      [6]
   2. b)

      Explain Q-Learning and how Deep Q-Networks (DQN) improve upon traditional Q-Learning for complex environments.

      [6]
   3. c)

      This part has equations that don’t show as text. Open the PDF to read it.

      [6]