AL-702-B Advanced Machine Learning - Dec 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. 7 Marks
   aDefine Artificial Neural Networks. What is the role of the cost function in training Artificial Neural Networks?
   How to answer

   • Definition of Artificial Neural Networks as interconnected neurons, layers, weights and activation
   • Components: input, hidden, output layers and learning from data
   • Definition of cost/loss function measuring prediction error
   • Role in training: guides optimization, weight update via minimization
2. 7 Marks
   bDescribe the role of gradient descent in the backpropagation process. How does it influence weight updates.
   How to answer

   • Overview of backpropagation: forward pass, error, backward pass
   • Gradient descent as iterative optimization of loss w.r.t. weights
   • Chain rule to compute gradients layer by layer
   • Weight update rule $w := w - \eta \nabla L$ and role of learning rate
3. 7 Marks
   aDescribe the architecture and learning process of a multi-layer perceptron.
   How to answer

   • MLP structure: input layer, one or more hidden layers, output layer, fully connected
   • Activations, weights, biases
   • Learning: forward propagation, loss, backpropagation with gradient descent
   • Non-linear separation capability
4. 7 Marks
   bWhat is recursive induction in decision trees? Explain.
   How to answer

   • Definition of recursive induction / divide-and-conquer splitting
   • Steps: select best attribute, split, recurse on subsets
   • Stopping criteria and leaf labeling
   • Example with entropy/information gain
5. 7 Marks
   aDefine entropy and information gain. Calculate the overall entropy and information gain with the following dataset.

   

   How to answer

   • Definitions: $Entropy(S) = -\sum p_i \log_2 p_i$, information gain as entropy reduction
   • Overall entropy of Loan Approved: 3 Yes, 4 No
   • Conditional entropies for Credit Score = High/Medium/Low
   • Information Gain = Overall entropy - weighted sum and interpretation
6. 7 Marks
   bHow does noisy data affect decision tree performance? Write the strategies to overcome them.
   How to answer

   • Effect of noise: overfitting, spurious splits, reduced generalization
   • Label noise vs attribute noise and impact on tree size/accuracy
   • Strategies: pruning, minimum samples, depth limit, data cleaning
   • Ensemble / validation and handling missing values
7. 7 Marks
   aCompare Bagging and Boosting techniques in ensemble learning.
   How to answer

   • Bagging: bootstrap parallel, variance reduction; Boosting: sequential error correction, bias reduction
   • Sample weighting, learner independence vs dependence
   • Comparison table on basis, robustness, examples Random Forest vs AdaBoost/XGBoost
   • Advantages and limitations
8. 7 Marks
   bDescribe the working mechanism of a Random Forest algorithm.
   How to answer

   • Bootstrap sampling and random feature subset selection
   • Building multiple decision trees in parallel and voting/averaging
   • Out-of-bag error and feature importance
   • Diagram of forest workflow
9. 7 Marks
   aWhy ensemble methods offer greater robustness compared to individual models? Discuss.
   How to answer

   • Variance reduction by averaging diverse models, bias-variance view
   • Robustness to noise, outliers and overfitting vs single model
   • Diversity through data/feature/model variation
   • Error independence and majority voting effect
10. 7 Marks
    bExplain the Bellman Equations in the context of Reinforcement Learning.
    How to answer

    • Bellman expectation and optimality equations for $V$ and $Q$
    • Relation of current value to immediate reward plus discounted future value
    • Role in policy evaluation and improvement
    • Example form $V(s)=\max_a [R+\gamma \sum P V(s')]$
11. 7 Marks
    aDifferentiate between Q-learning and SARSA algorithms.
    How to answer

    • Q-learning: off-policy, max Q update; SARSA: on-policy, actual next action
    • Update formulas comparison
    • Exploration vs exploitation behavior and convergence differences
    • Suitability and safety aspects
12. 7 Marks
    bDescribe the concept of Policy Gradient methods in Reinforcement Learning.
    How to answer

    • Policy directly parameterized $\pi_\theta(a|s)$, gradient ascent on expected return
    • REINFORCE / policy gradient theorem overview
    • Advantages for continuous and stochastic policies
    • High variance and baseline/actor-critic remedies
13. 7 Marks
    aWhat is temporal difference learning and how does it differ from Monte Carlo methods?
    How to answer

    • TD learning: bootstrapping from current estimates, online update
    • Monte Carlo: complete episode returns, no bootstrapping
    • Differences in bias-variance, online vs offline, model-free nature
    • Examples TD(0), SARSA vs MC prediction
14. 7 Marks
    bHow does Generative Adversarial Imitation learning differ from standard RL approaches?
    How to answer

    • GAIL: discriminator-generator imitation from expert demos vs reward-driven RL
    • No explicit reward function, occupancy matching
    • Architecture: generator policy, discriminator distinguishing expert vs agent
    • Advantages and challenges vs standard RL
15. 14 Marks
    Write short note on any two :

    1. a)Value Iteration and Policy Iteration
    2. b)Recent Trends in RL Architectures
    3. c)Random Forest
    4. d)Least Squares Methods
    How to answer

    • Select any two topics with definition and key steps/formulas
    • Working principle and algorithm outline
    • Advantages, limitations and applications
    • Neat sketch/flow where relevant