Question: A classifier is estimated to have 12% bias and 2% variance. Describe what this suggests about its training-set and development-set error rates, and explain the likely diagnosis of the model.

Sample answer: A bias of 12% suggests the model is still making fairly large errors on the training set, so it is not fitting the available data very well. A variance of 2% means the development-set error is only slightly higher than the training-set error, so the gap between the two is small. Together, these numbers indicate high bias and low variance, which is the pattern usually associated with underfitting.

Key points:

• 12% bias implies the training-set error is still relatively high.
• 2% variance means the development-set error is only a little worse than the training-set error.
• The model shows high bias and low variance.
• This pattern is commonly described as underfitting.

Rubric: The essay should identify that the model has relatively high training error, only a small train-dev gap, and therefore is underfitting because bias is high while variance is low.