Sample questions
Ml Model Selection TuningDifficulty 1
In machine learning, which of the following best distinguishes a 'hyperparameter' from a model parameter?
- aA hyperparameter is learned automatically from the training data during the fitting process.
- bA hyperparameter only exists for models that use gradient-based optimization internally.
- cIt is always set before training begins (never learned from data) and controls how the learning process behaves.✓
- dA hyperparameter is a value stored only for logging purposes and has no effect on training.
Explanation:Hyperparameters (e.g. regularization strength, number of estimators, tree depth) are chosen before training and shape how the model learns; parameters (weights, coefficients, splits) are learned from data during training.
Ml Model Selection TuningDifficulty 1
What is the primary goal of hyperparameter tuning?
- aTo find hyperparameter values that generalize well to unseen data, not just ones that minimize training error.✓
- bTo make the model memorize the training set as precisely as possible.
- cTo reduce the number of features the model uses regardless of accuracy.
- dTo guarantee the model will never make an incorrect prediction in production.
Explanation:Tuning searches for hyperparameter settings that improve performance on data the model hasn't seen, estimated via a validation set; optimizing purely for training-set fit would encourage overfitting.
Ml Model Selection TuningDifficulty 2
A team defines a small set of candidate values for each hyperparameter and evaluates every possible combination of those values. Which search strategy are they using?
- aSequential fine-tuning of a single hyperparameter at a time using production feedback.
- bA strategy that samples hyperparameter combinations uniformly at random from a defined range.
- cA strategy that uses the outcome of previous trials to decide which combination to try next.
- dAn exhaustive search over every combination in a predefined grid of candidate values.✓
Explanation:Evaluating every combination from a predefined set of candidate values per hyperparameter is the exhaustive ('grid') search idea; it's simple but its cost grows multiplicatively with the number of hyperparameters and values.
Ml Model Selection TuningDifficulty 2
Compared to exhaustively trying every combination on a fixed grid, what is a commonly cited advantage of sampling hyperparameter combinations at random from a defined range?
- aIt always finds the mathematically optimal combination faster than any other method.
- bIt covers more distinct values per hyperparameter under a fixed budget.✓
- cIt removes the need for a validation set entirely during the search.
- dIt guarantees that every possible combination will eventually be tried.
Explanation:Random sampling spreads a fixed budget of trials across more distinct values per hyperparameter, which tends to help because in practice only a subset of hyperparameters strongly affect performance.
Ml Model Selection TuningDifficulty 3
A tuning approach builds a probabilistic model of how hyperparameter values relate to validation performance, and uses it to decide which combination is most promising to try next. What idea does this describe?
- aExhaustively evaluating every combination on a predefined grid.
- bUniformly sampling combinations at random, independent of past results.
- cUsing the outcome of earlier trials to guide which combination to try next, rather than sampling blindly.✓
- dManually adjusting one hyperparameter at a time based on intuition alone.
Explanation:This describes the idea behind sequential model-based search (often called Bayesian optimization): it uses past trial outcomes to model the search space and prioritize promising regions, which can reduce the number of trials needed versus grid or pure random search.
Ml Model Selection TuningDifficulty 2
A team wants a linear model where many coefficients become exactly zero, effectively performing feature selection during training. Which regularization approach is best known for this behavior?
- aL2 regularization, because it shrinks large coefficients much more aggressively than small ones.
- bNo regularization, since removing the penalty term is what causes coefficients to reach exactly zero.
- cIncreasing the number of training epochs without changing the loss function.
- dL1 regularization, because its penalty can push less useful coefficients all the way to zero.✓
Explanation:L1 (absolute-value) regularization tends to produce sparse solutions, driving the coefficients of less useful features exactly to zero, which acts like built-in feature selection. L2 shrinks coefficients but rarely zeroes them out.