CS-601 Machine Learning - May 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 various types of machine learning used for continuous data and non-continuous data.
   How to answer

   • Define supervised, unsupervised, semi-supervised, reinforcement learning
   • Distinguish continuous data - regression, time-series vs non-continuous / discrete / categorical - classification, clustering
   • Give example algorithms for each type
2. 7 Marks
   bGive an explanation of the One-hot encoding as well as the Label encoding. In what ways do they change the dimensionality of the data that has been provided?
   How to answer

   • Define One-hot encoding with example vector
   • Define Label encoding with example integers
   • Compare effect on dimensionality - expansion vs same dimension and implications
3. 7 Marks
   aBreak down how CNN actually operates. The image is downscaled, and the number of filters is increased as we approach the model's output, but why?
   How to answer

   • Outline convolution, activation, pooling, fully connected stages
   • Explain spatial downsampling and hierarchical feature learning
   • Explain why filters increase - more abstract features, compensate spatial loss, control parameters
4. 7 Marks
   bReplace the old weights (not the bias) in the network depicted in the following figure using back-propagation algorithms. A [0,1] input pattern is given to the network, and the desired output is 1. Use the Sigmoid activation function and a learning rate of = 0.3.

   

   How to answer

   • Compute forward pass with sigmoid for z1, z2, y for input [0,1]
   • Compute output error and backpropagate deltas to hidden and input layers
   • Update weights with learning rate 0.3 using gradient descent rule
5. 7 Marks
   aDescribe the benefits of transfer learning features that can be transferred. Explain Inception net architecture in detail
   How to answer

   • Define transfer learning and transferable features - early layers, embeddings
   • List benefits - less data, faster training, better generalization
   • Explain Inception net - inception modules, 1x1 convolutions, GoogLeNet structure
6. 7 Marks
   bWhat are the structural and operational differences between a feed-forward network and a recurrent neural network? Identify the differences between LSTM, GRU, and vanilla RNNs.
   How to answer

   • Compare feed-forward acyclic vs recurrent cyclic with feedback and memory
   • Explain operation on sequential data vs static input
   • Differentiate LSTM gates, GRU simplified gates, vanilla RNN vanishing gradient
7. 14 Marks
   Differentiate:

   1. i)Various loss functions
   2. ii)Types of Gradient Descent Optimizers
   How to answer

   • Define loss functions - MSE, cross-entropy, hinge etc. with use cases
   • Define gradient descent variants - batch, stochastic, mini-batch, Adam, RMSprop etc.
   • Tabulate differences on computation, convergence, memory
8. 7 Marks
   aClarify the meaning of Confusion metrics in the context of machine learning. What other metrics might be derived from the metric of confusion?
   How to answer

   • Define confusion matrix - TP, TN, FP, FN
   • Explain accuracy, precision, recall, F1 derived from it
   • Give example of imbalanced data use
9. 7 Marks
   b"Model performance can be greatly improved through careful hyperparameter tuning". Justify the statement.
   How to answer

   • Define hyperparameters vs parameters with examples
   • Explain tuning methods - grid, random, Bayesian search
   • Justify impact on performance with overfitting/underfitting example
10. 7 Marks
    aExplain how the KNN method is implemented. Below is information regarding a player's speed and agility that will be used to determine whether he will be drafted into the team. Predict the likelihood that a player with "speed = 6.75" and "agility = 3" will make the team using the KNN machine learning model assuming "k = 3".

    

    How to answer

    • Outline KNN steps - distance metric, choose k, majority vote
    • Compute Euclidean distances for speed=6.75, agility=3 to all IDs
    • Identify 3 nearest neighbours and predict Draft yes/no
11. 7 Marks
    bHow can we make the model non-linear? If we only use linearity, how will that affect Gradient Descent?
    How to answer

    • Explain role of non-linear activation functions - ReLU, sigmoid, tanh
    • Show linear-only network collapses to linear model
    • Explain effect on gradient descent - convex vs non-convex, representational power
12. 7 Marks
    aBased on figure, which depicts the Bayesian belief network for the data set shown in the table below, respond to the following questions. (Assume that each attribute has a binary value.) https://www.rgpvonline.com

    1. i)Draw the probability table for each node in the network.
    2. ii)Use the Bayesian network to predict the 'car value' for the following. P(Mileage = Lo, Engine = Bad, Air Conditioner = Broken)

    

    How to answer

    • Compute prior and conditional probabilities from counts table for Mileage, Engine, Air Conditioner, Car Value
    • Draw CPTs per Bayesian belief network structure
    • Apply chain rule to predict P for Mileage=Lo, Engine=Bad, Air Conditioner=Broken
13. 7 Marks
    bDescribe the procedure that can be used to identify overfitting and underfitting in a CNN model. Include some potential solutions to this problem as well.
    How to answer

    • Define overfitting vs underfitting via training vs validation loss curves
    • Describe detection procedure - split, learning curves, metrics
    • List solutions - regularization, dropout, augmentation, early stopping, more data
14. 14 Marks
    Explain the following with an appropriate example:

    1. i)Computer Vision
    2. ii)Reinforcement learning
    How to answer

    • Define Computer Vision with CNN example - image classification, detection
    • Define Reinforcement learning - agent, environment, reward with game/robot example
    • Give appropriate real-world example for each