Question: Describe how a team can use training and development error rates to estimate which type of error is limiting performance, and explain how fixes aimed at bias differ from fixes aimed at variance.

Sample answer: A team can compare the model’s training error with its development error to estimate whether underfitting or overfitting is the bigger issue. If the training error is still high, that points to a bias problem; if the training error is low but the development error is much worse, that points to a variance problem. This comparison helps the team decide what to try first instead of applying changes at random. Methods aimed at lowering bias are intended to make the model fit the training data better. Methods aimed at lowering variance are intended to make the model behave more consistently on new data, improving its performance on development and test sets.

Key points:

• Training and development errors are compared to judge whether bias or variance is the larger issue.
• That comparison guides which improvement should be prioritized.
• Bias-focused changes aim to improve fit on the training set.
• Variance-focused changes aim to improve generalization beyond the training set.

Rubric: The response should explain that training and development error rates are used together to estimate bias and variance, state that this comparison helps identify the main limitation to address first, and clearly distinguish the goal of bias-reducing methods from the goal of variance-reducing methods.