AL-503-B Deep 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
   aExplain the historical progression of Deep Learning, highlighting key milestones and breakthroughs that have shaped its development.
   How to answer

   • Define deep learning and early perceptron era
   • Key milestones: backpropagation, CNNs, LSTM, AlexNet, GANs, Transformers
   • Breakthroughs in data, compute (GPU) and algorithms
   • Impact on current applications
2. 7 Marks
   bAssess the representation power of Multilayer Perceptrons (MLPs) and discuss their capacity to learn complex relationships in data.
   How to answer

   • Define Multilayer Perceptron structure
   • Universal approximation and representation power
   • Capacity to learn non-linear complex relationships
   • Limitations: data, overfitting, vanishing gradients
3. 7 Marks
   aExplain the architecture and working principles of Deep Feed forward Neural Networks. Discuss their application areas and key advantages over other types of neural networks.
   How to answer

   • Architecture: input, hidden, output layers, feedforward flow
   • Working: weighted sum, activation, forward propagation
   • Application areas: classification, regression, NLP, vision
   • Advantages: simplicity, universal approximation, scalability
4. 7 Marks
   bInvestigate the theoretical foundations and practical implications of AdaGrad, Adam, and RMSProp optimization algorithms in the context of training deep neural networks.
   How to answer

   • Theory of adaptive learning rates: AdaGrad, RMSProp, Adam
   • Update rules and moment estimates formulas
   • Practical implications: convergence speed, sparse gradients
   • Comparison and suitability for deep network training
5. 7 Marks
   aExplain the concept of Convolutional Neural Networks (CNNs) and their role in image recognition tasks.
   How to answer

   • Define CNN and its bio-inspired concept
   • Components: convolution, pooling, fully connected layers
   • Role in feature extraction for image recognition
   • Advantages in spatial hierarchy learning
6. 7 Marks
   bDefine the ReLU activation function and explain its significance in CNNs.
   How to answer

   • Definition: $ReLU(x) = max(0,x)$
   • Properties: non-linearity, sparsity, efficiency
   • Significance in avoiding vanishing gradient
   • Impact on faster training of CNNs
7. 7 Marks
   aExplain the fundamental concepts and various architectures of Deep Recurrent Neural Networks (RNNs).
   How to answer

   • Fundamentals: recurrence, hidden state, temporal unfolding
   • Architectures: Elman, stacked, bidirectional, LSTM/GRU based deep RNN
   • Training via BPTT
   • Applications in sequential data
8. 7 Marks
   bExplain the challenges of vanishing and exploding gradients during the BPTT process and their impact on the network's ability to learn long-range dependencies.
   How to answer

   • Define BPTT and gradient propagation over time
   • Causes of vanishing and exploding gradients
   • Impact on long-range dependency learning
   • Mitigations: clipping, initialization, LSTM/GRU
9. 7 Marks
   aExplain the concept of value iteration in dynamic programming for solving Markov Decision Processes (MDPs).
   How to answer

   • Define MDP and Bellman optimality equation
   • Value iteration algorithm steps
   • Convergence concept and stopping criteria
   • Role in solving MDPs via dynamic programming
10. 7 Marks
    bCompare and contrast policy iteration with value iteration in terms of convergence and computational complexity.
    How to answer

    • Outline policy iteration: evaluation + improvement
    • Outline value iteration approach
    • Compare convergence rate and iterations
    • Compare computational complexity per iteration
11. 7 Marks
    aDiscuss the Backpropagation algorithm and its role in optimizing the weights of neural networks during the training process.
    How to answer

    • Define backpropagation and chain rule basis
    • Forward pass, loss computation, backward gradient flow
    • Weight update via gradient descent
    • Role in optimizing deep network training
12. 7 Marks
    bDiscuss the role of regularization in auto-encoders and its impact on the training process.
    How to answer

    • Define auto-encoders and overfitting risk
    • Regularization types: sparse, denoising, contractive, L1/L2
    • Impact on latent representation and generalization
    • Effect on training stability and reconstruction
13. 7 Marks
    aExplain the concept of Deep Dream and its role in generating surreal images based on neural network activations.
    How to answer

    • Define Deep Dream: maximizing layer activations
    • Process: iterative gradient ascent on input image
    • Role in visualizing learned features
    • Generation of surreal dream-like images
14. 7 Marks
    bDiscuss the challenges associated with encoding and decoding sequential data and how RNNs handle these challenges through their inherent architecture.
    How to answer

    • Challenges: variable length, context, vanishing gradients
    • Encoding-decoding sequential data concept
    • RNN inherent memory and hidden state mechanism
    • Seq2seq, attention need and limitations
15. 7 Marks
    aCompare and contrast different least squares methods, such as LSPI (Least Squares Policy Iteration).
    How to answer

    • Define least squares methods in RL context
    • LSPI algorithm: policy evaluation and improvement
    • Comparison with other least squares variants
    • Sample efficiency and function approximation benefits
16. 7 Marks
    bExplore advanced Q-learning algorithms, such as Double DQN and Dueling DQN.
    How to answer

    • Basics of Q-learning and DQN overestimation
    • Double DQN: decoupled selection and evaluation
    • Dueling DQN: value and advantage streams
    • Performance improvements and use cases