Question: Write a concise analytical response comparing reducing model capacity with adding regularization when extra computation is not a constraint.

Sample answer: Making the network smaller by removing layers or reducing units can lower variance, but it is usually a fallback rather than the first choice. If compute is not a limiting factor, regularization is often the better option because it tends to improve generalization more reliably. Regularization can reduce variance, though it may do so at the cost of increasing bias.

Key points:

• A smaller model can reduce variance.
• Reducing model capacity should be used carefully.
• Regularization is often the preferred remedy when compute is available.
• Regularization can improve generalization.
• Regularization may lower variance while raising bias.

Rubric: A strong response explains that reducing model size can help with variance, notes that it should be used carefully, identifies regularization as the preferred approach when computation is not a concern, and mentions the bias-variance tradeoff.