The signature
A model underfits when it performs poorly on both training and validation data. It lacks the capacity or features to capture the underlying pattern.
- Training error is high and barely improves.
- Validation error tracks training error closely.
- The model is too simple or under trained.
Confirming and fixing
If both curves sit high and close together, you are underfitting. The cure is to increase the model's ability to fit.
- Add capacity with more parameters or richer features.
- Reduce regularization that is too aggressive.
- Train longer or raise the learning rate within reason.
The picture
Underfitting and overfitting are opposite failure modes on the bias variance scale.
Key idea
Underfitting shows as high error on both training and validation with a small gap, meaning the model is too weak; add capacity, richer features, or training time to fix it.