Question: A wildfire-smoke detector has an estimated human-level error of about 1%. On the training set, it makes 2% error. On a training-dev set drawn from the same distribution as training, it makes 8% error. Explain why this pattern points to high variance rather than high bias.

Sample answer: This pattern suggests high variance because the training error is low relative to the human-level error, so the model is already fitting the training examples reasonably well and does not appear to have a major bias problem. The much larger jump from 2% error on training data to 8% error on training-dev data shows that the model performs noticeably worse on new examples from the same distribution. That kind of gap indicates poor generalization, which is the signature of high variance.

Key points:

• Training error is close to the human-level error, so the model is not showing strong high bias.
• Training-dev error is much higher than training error, revealing a generalization gap.
• A large gap between training and training-dev performance indicates high variance.

Rubric: A strong response should state that the low training error relative to the benchmark rules out high bias and that the much worse training-dev error indicates the model does not generalize well, which is evidence of high variance.