al-503-deep-system-dec-2024 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 difference between supervised and unsupervised learning. Provide examples where deep learning can be used for each type.
   How to answer

   • Define supervised learning with labelled data and unsupervised learning with unlabelled data
   • Compare on objective, data, feedback, examples
   • Deep learning example for supervised: image classification, speech recognition
   • Deep learning example for unsupervised: autoencoders, clustering, GANs
2. 7 Marks
   bWhat are activation functions? Why are they important in neural networks? List and explain different types of activation functions used in deep learning.
   How to answer

   • Define activation function and its role in introducing non-linearity
   • Explain importance for learning complex mappings and backpropagation
   • List Sigmoid, Tanh, ReLU, Leaky ReLU, Softmax
   • Explain each with formula, range and use-case
3. 7 Marks
   aBriefly explain the architecture of a simple neural network. What are activation functions and why are they used? Discuss the role of backpropagation in training neural networks.
   How to answer

   • Describe simple neural network layers: input, hidden, output, weights, bias
   • Define activation functions and need for non-linearity
   • List common activations briefly
   • Explain backpropagation steps: forward pass, loss, gradient, weight update
4. 7 Marks
   bDefine the terms "Overfitting" and "Underfitting" in the context of deep learning models. What techniques can be used to prevent overfitting?
   How to answer

   • Define overfitting: low training error but poor generalization
   • Define underfitting: poor performance on both train and test
   • Explain causes: model complexity, data size
   • List prevention techniques: regularization, dropout, early stopping, augmentation, cross-validation
5. 7 Marks
   aWhat is the Vanishing Gradient Problem? How does it affect the training of deep neural networks and what strategies can be used to mitigate it?
   How to answer

   • Define vanishing gradient: gradients shrink in early layers
   • Explain cause: saturating activations, deep chain rule multiplication
   • Effect: slow/no learning, poor convergence
   • Mitigations: ReLU, initialization, batch norm, residual connections, LSTM/GRU
6. 7 Marks
   bWhat is a Recurrent Neural Network (RNN), and how is it useful for sequence data? Compare it with a traditional feed-forward neural network.
   How to answer

   • Define RNN with recurrent connections and hidden state
   • Explain suitability for sequential data: memory of past inputs
   • Describe feed-forward network with no memory
   • Compare on architecture, memory, applications
7. 7 Marks
   aDiscuss the role of dropout in training deep neural networks. How does it help improve generalization?
   How to answer

   • Define dropout as random deactivation of neurons during training
   • Explain ensemble effect and preventing co-adaptation
   • Explain dropout rate and train vs test behaviour
   • Explain improved generalization and reduced overfitting
8. 7 Marks
   bExplain the backpropagation algorithm. How is it used to update the weights in a neural network during training?
   How to answer

   • Define backpropagation as gradient-based weight update
   • Steps: forward pass, loss computation, backward chain rule
   • Weight update using gradient descent / learning rate
   • Explain iterative error minimization over epochs
9. 7 Marks
   aDevelop A table with examples of different formats of data that can be used with convolutional network system.
   How to answer

   • Define CNN input requirements: grid-like data
   • Tabulate formats: images, video, spectrograms, text embeddings
   • Give example for each format
   • Mention channels, dimensions
10. 7 Marks
    bWhat is batch normalization and why is it used during the training of deep networks?
    How to answer

    • Define batch normalization: normalizing layer inputs per mini-batch
    • Explain parameters gamma, beta and inference statistics
    • Reasons: reduce internal covariate shift, faster training
    • Benefits: higher learning rates, regularization effect
11. 7 Marks
    aWhat is the difference between normalization and standardization in data preprocessing?
    How to answer

    • Define normalization: rescaling to [0,1] range
    • Define standardization: zero mean unit variance
    • Compare formulas and when each is used
    • Give examples of algorithms preferring each
12. 7 Marks
    bDiscuss unsupervised learning with examples.
    How to answer

    • Define unsupervised learning without labels
    • Explain clustering, association, dimensionality reduction
    • Give examples: k-means, autoencoders, PCA
    • Mention deep learning uses
13. 7 Marks
    aExplain how reinforcement learning differs from supervised learning and provide an example of how deep learning can be applied in reinforcement learning.
    How to answer

    • Define reinforcement vs supervised: agent-reward vs labelled mapping
    • Compare feedback, objective, data needs
    • Explain deep RL: neural network as policy/value approximator
    • Example: game playing, robotics, DQN
14. 7 Marks
    bDiscuss the differences between Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). When would you choose one over the other?
    How to answer

    • Define GANs: generator-discriminator adversarial game
    • Define VAEs: probabilistic encoder-decoder with latent distribution
    • Compare on training, output quality, stability, interpretability
    • Guidance on choosing: sharp generation vs structured latent modelling
15. 14 Marks
    Explain the working of any two of the following machine learning algorithms in detail : a) Long Short-Term Memory Networks (LSTM) b) Logistic Regression c) Recursive Neural Networks
    How to answer

    • Select any two algorithms from LSTM, Logistic Regression, Recursive NN
    • Explain architecture and key equations/gates
    • Explain training and data flow
    • Mention applications and advantages