This first set of Machine Learning MCQs builds the foundation for everything that follows. It covers the three main learning paradigms, features and labels, train/validation/test splits, data leakage, overfitting and underfitting, and the bias-variance trade-off. The questions suit GATE Data Science and AI aspirants, university semester exams, UGC NET and placement or interview rounds where the basics of model generalisation are tested before any algorithm.
How to practise: read each question and tap the option you think is correct. You will see right away whether you got it, along with a short explanation, and your score is tracked at the bottom of the screen. This set has 18 questions.
Topics covered in this set
- Supervised, unsupervised and reinforcement learning
- Features, labels and the hypothesis space
- Training, validation and test splits
- Overfitting, underfitting and generalisation
- Bias-variance decomposition of error
- Data leakage and the no free lunch idea
Machine Learning MCQs with answers and explanations
Q1. A bank wants to group its customers into segments based on spending patterns, without any predefined segment labels. Which type of learning fits this task?
a) Supervised classification
b) Unsupervised learning
c) Reinforcement learning
d) Supervised regression
Q2. Which of the following tasks is best formulated as a regression problem?
a) Detecting whether an email is spam
b) Recognising handwritten digits 0-9
c) Grouping news articles by topic without labels
d) Predicting tomorrow's maximum temperature in degrees Celsius
Q3. In reinforcement learning, what does the agent try to maximise?
a) The expected cumulative (discounted) reward over time
b) The accuracy on a labelled test set
c) The number of states it visits
d) The immediate reward of the current action only
Q4. A model achieves 99% accuracy on the training data but only 62% on unseen test data. What is the most likely diagnosis?
a) Underfitting due to high bias
b) Irreducible noise dominates the error
c) Overfitting due to high variance
d) The learning rate is too small
Q5. A dataset contains 2,500 labelled samples. If 80% is used for training and the remainder is split equally into validation and test sets, how many samples are in the test set?
a) 500
b) 250
c) 200
d) 125
Q6. For squared-error loss, the expected test error of a model at a point can be decomposed as which of the following?
a) Bias^2 + Variance + Irreducible noise
b) Bias + Variance - Noise
c) Bias x Variance + Noise
d) Variance^2 + Bias + Noise
Q7. A regression model has bias 0.3, variance 0.05 and irreducible noise variance 0.02 at a point. What is its expected squared error there?
a) 0.37
b) 0.35
c) 0.07
d) 0.16
Q8. Which change would most likely reduce underfitting in a model?
a) Increasing the strength of regularisation
b) Removing informative features
c) Using a more expressive model or adding relevant features
d) Reducing the number of training epochs drastically
Q9. As model complexity increases steadily, how do bias and variance typically behave?
a) Both bias and variance increase
b) Bias decreases while variance increases
c) Bias increases while variance decreases
d) Both remain constant if data is fixed
Q10. What is the main purpose of a separate validation set?
a) To tune hyperparameters and select models without touching the test set
b) To increase the size of the training data
c) To compute the irreducible error exactly
d) To label the unlabelled training examples
Q11. A data scientist standardises the entire dataset using its overall mean and standard deviation, and only then splits it into train and test sets. What problem does this introduce?
a) Class imbalance
b) Vanishing gradients
c) Data leakage from the test set into training
d) The curse of dimensionality
Q12. In supervised learning, the input variables used by the model to make predictions are called what?
a) Labels
b) Rewards
c) Hyperparameters
d) Features
Q13. Which scenario is a typical example of semi-supervised learning?
a) Training a chess engine through self-play
b) Using a small set of labelled images together with a large set of unlabelled images
c) Clustering documents with k-means
d) Fitting a linear regression on fully labelled data
Q14. For a fixed model class that is overfitting, which action generally reduces variance most directly?
a) Collecting more training data
b) Adding more polynomial features
c) Removing the regularisation term
d) Training for many more epochs
Q15. Which component of error cannot be reduced by any choice of model?
a) Bias
b) Variance
c) Irreducible error caused by noise in the data
d) Optimisation error
Q16. What does the 'no free lunch' theorem imply for machine learning?
a) Deep neural networks are always the best choice
b) No single algorithm performs best across all possible problems
c) More data never improves performance
d) Every model has zero bias given infinite data
Q17. Training and validation error are both high and close to each other. What is the best interpretation?
a) The model is overfitting
b) The validation set is leaking into training
c) The model generalises perfectly
d) The model is underfitting (high bias)
Q18. A spam filter is trained on emails from 2015 and deployed in 2026, where spammers use new phrases. Its accuracy drops sharply. What is this phenomenon called?
a) Dataset shift (concept or covariate drift)
b) Underfitting
c) Label smoothing
d) Feature scaling error
Quick revision notes
- Supervised: labelled (x, y) pairs; unsupervised: only x; reinforcement: agent, states, actions, rewards.
- Regression predicts continuous values; classification predicts discrete classes.
- Expected squared error = Bias^2 + Variance + Irreducible noise.
- Overfitting: low training error, high test error (high variance).
- Underfitting: high training and test error (high bias).
- Use train set to fit, validation set to tune, test set only once for final evaluation.
- Fit preprocessing (scaling, imputation) on training data only to avoid data leakage.
- More data reduces variance; more model capacity reduces bias.
Frequently asked questions
What is the difference between bias and variance in machine learning?
Bias is the error from overly simple assumptions that make a model miss the true pattern. Variance is the error from sensitivity to the particular training sample, so the model changes a lot with different data. Good models balance the two.
How can I tell if my model is overfitting?
Compare training and validation performance. If the training score is much better than the validation score, the model is memorising noise rather than learning general patterns, which is overfitting.
Why do we need both a validation set and a test set?
The validation set is used repeatedly to tune hyperparameters, so performance on it becomes optimistically biased. The untouched test set gives an honest final estimate of how the model will perform on new data.
Comments
Post a Comment