Capacity Clinic

Predict first: If training error is already very small but validation error is much larger, will adding even more capacity usually help?

Model fit

TrainingValidationCurrent model

Error versus capacity

Training RMSEValidation RMSECurrent degree

Challenge mode

Polynomial degree

Training examples

Noise level

Training RMSE

Validation RMSE

Your diagnosis: use the two errors and the shape of the curves, not just the visual fit.

Choose a diagnosis.

Then prescribe a fix.

Choose a remedy after diagnosing.

Instructor notes

Underfitting: both training and validation errors are high. In the lecture, adding more training examples is not the main remedy. More features or a richer classifier can help.

Overfitting: training error is low while validation error is much higher. More data may help, and reducing feature/model complexity may help.

Important: the exact numerical boundary between “high” and “low” error is problem-dependent. Here the diagnosis is based on the error curves for this controlled example.

Capacity Clinic: diagnose before you prescribe

Model fit

Error versus capacity