Q1. According to Tom Mitchell’s definition of machine learning, which of the following triplet correctly describes the key elements?

1. Task (T), Time (T), Performance (P)
2. Task (T), Experience (E), Performance (P)
3. Training (T), Error (E), Probability (P)
4. Transformation (T), Encoding (E), Prediction (P)

Select one:

A
B
C
D

Q2. Which of the following is the most appropriate description of supervised learning?

1. Learning a mapping from inputs to outputs using labelled examples.
2. Identifying clusters in data without any labels.
3. Learning a policy to maximise long-term reward by interacting with an environment.
4. Generating new samples from a learned data distribution.

Select one:

A
B
C
D

Q3. In the context of generalisation, overfitting occurs when:

1. The model is too simple and underestimates the underlying structure.
2. The model fits the training data very well but performs poorly on unseen data.
3. The training and test errors are approximately equal.
4. The model has no regularisation and zero bias.

Select one:

A
B
C
D

Q4. (Short answer) Briefly distinguish regression from classification in supervised learning. Your answer should mention the nature of the output in each case.

Q5. For ordinary least squares (OLS) linear regression with design matrix X and target vector y, which expression gives the closed-form solution for the weight vector w, assuming XᵀX is invertible?

1. w = (XXᵀ)−1 X y
2. w = Xᵀ (XXᵀ)−1 y
3. w = (XᵀX)−1 Xᵀ y
4. w = X (XᵀX)−1 y

Select one:

A
B
C
D

Q6. Which statement best characterises the difference between Ridge regression and Lasso regression?

1. Ridge uses an L1 penalty, Lasso uses an L2 penalty.
2. Ridge shrinks coefficients but rarely sets them exactly to zero; Lasso encourages sparsity by driving some coefficients exactly to zero.
3. Ridge is only applicable to classification, Lasso only to regression.
4. There is no practical difference; both penalise coefficients identically.

Select one:

A
B
C
D

Q7. Logistic regression for binary classification models which of the following quantity?

1. P(y = 0 | x) as a linear function of x.
2. P(y = 1 | x) using a hinge loss.
3. The conditional mean of y as a quadratic function of x.
4. P(y = 1 | x) using a sigmoid applied to a linear combination of x.

Select one:

A
B
C
D

Q8. (Short answer) Explain, in one or two sentences, what is meant by the bias–variance trade-off in the context of linear models and regularisation.

Q9. In a classification decision tree, which of the following is a typical splitting criterion at an internal node?

1. Mean absolute error
2. Gini impurity or entropy
3. Learning rate
4. ROC–AUC

Select one:

A
B
C
D

Q10. Which statement correctly describes a Random Forest?

1. A single very deep decision tree trained on all features and all data.
2. An ensemble of decision trees where each tree is trained on a bootstrap sample and uses random subsets of features at splits.
3. A linear model that uses tree-based basis functions.
4. A gradient boosting model with a single tree.

Select one:

A
B
C
D

Q11. In gradient boosting, each successive tree is trained to:

1. Maximise the Gini impurity of the training data.
2. Randomly assign predictions to increase model diversity.
3. Approximate the negative gradient (residuals) of the loss with respect to the current model predictions.
4. Reduce the number of features by pruning.

Select one:

A
B
C
D

Q12. (Short answer) State one key advantage and one key disadvantage of using a single decision tree compared to a tree ensemble (e.g., Random Forest or Gradient Boosted Trees).

Q13. Which of the following is the primary role of a non-linear activation function in a neural network?

1. To ensure that weights remain small during training.
2. To allow the network to model non-linear relationships between inputs and outputs.
3. To normalise the input features before training.
4. To perform dimensionality reduction.

Select one:

A
B
C
D

Q14. The backpropagation algorithm is best described as:

1. A method for randomly initialising neural network weights.
2. A technique for computing gradients of the loss with respect to all parameters using the chain rule efficiently.
3. A regularisation strategy to prevent overfitting in neural networks.
4. A data augmentation technique for image classification.

Select one:

A
B
C
D

Q15. Which of the following techniques is specifically aimed at reducing co-adaptation of neurons and improving generalisation in deep networks?

1. Batch Normalisation
2. Dropout
3. Gradient Clipping
4. Learning Rate Warm-up

Select one:

A
B
C
D

Q16. (Short answer) In your own words, describe the universal approximation theorem for neural networks, and explain its practical limitation.

Q17. In the standard scaled dot-product attention mechanism, why is the dot product between queries and keys divided by √dk before applying softmax?

1. To reduce computational complexity.
2. To prevent attention scores from becoming too small.
3. To stabilise gradients by preventing excessively large values in the softmax exponent.
4. To enforce sparsity in the attention weights.

Select one:

A
B
C
D

Q18. Which of the following statements about transformers is most accurate?

1. They rely exclusively on convolutional layers for sequence modelling.
2. They use self-attention mechanisms and positional encodings instead of recurrence to model sequences.
3. They are only applicable to image data and cannot handle text.
4. They cannot be scaled beyond a few million parameters.

Select one:

A
B
C
D

Q19. Which of the following is a key risk associated with large language models (LLMs) in real-world deployment?

1. They are always perfectly calibrated and never produce incorrect outputs.
2. They may generate plausible-sounding but factually incorrect information (“hallucinations”).
3. They inherently eliminate all forms of bias present in training data.
4. They do not require any monitoring once deployed.

Select one:

A
B
C
D

Q20. (Short answer) Provide one example of an ethical or societal concern related to deploying deep learning systems (e.g., in healthcare, finance, or public policy), and briefly describe how responsible AI practice might mitigate this concern.