Case context: A team trains a speech-recognition network with aggressive dropout to control overfitting. The model finishes training with a training error of 22%, which is far above the team’s target of 8%. The validation error is 24%, so the gap between training and validation performance is small.

Question: Should the team reduce or remove dropout in this situation? Explain the reasoning and the likely effect on bias and variance.

Sample answer: Yes. The model appears to have high avoidable bias because its training error is still much worse than the desired level. At the same time, the small gap between training and validation error suggests low variance, so the model is not currently overfitting badly. Reducing dropout would make the network more flexible and should help it fit the training data better, which can lower bias. The trade-off is that variance may rise, so the team should check validation performance after changing the regularization.

Key points:

• Training error far above the target indicates high avoidable bias
• A small training-validation gap indicates low variance
• Reducing dropout is reasonable when more flexibility is needed
• Lower bias may come with a rise in variance
• Validation performance should be monitored after the change

Rubric: Full credit requires identifying high avoidable bias and low variance from the case, recommending less dropout, and stating that variance may increase while validation performance should be checked afterward.