Questions: Structural Risk Minimization

4 questions to test your understanding

Score: 0 / 4
Question 1 Multiple Choice

SRM chooses among hypothesis classes H_1, H_2, H_3 with VC dimensions 2, 5, 20. On 100 training examples, H_1 achieves 15% training error, H_2 achieves 4% training error, and H_3 achieves 0.5% training error. Which class would SRM likely select?

AH_3, because it has the lowest training error and the penalty term is negligible with 100 samples
BH_2, because it balances moderate training error with a manageable complexity penalty — the bound for H_3 adds a large penalty of roughly sqrt(20/100) ≈ 0.45, wiping out its training error advantage
CH_1, because SRM always prefers the simplest model regardless of training error
DIt depends entirely on the test set, which SRM does not use
Question 2 True / False

SRM is essentially equivalent to L2 regularization — both add a penalty to the training objective to prevent overfitting.

TTrue
FFalse
Question 3 True / False

If you had unlimited training data, SRM would always select the most complex hypothesis class in the hierarchy.

TTrue
FFalse
Question 4 Short Answer

Explain how SRM operationalizes the bias-complexity tradeoff into an algorithm, and what theoretical guarantee it provides.

Think about your answer, then reveal below.