Underfitting happens when a model is too simple to capture the real pattern.
It performs poorly even on the data it was trained on.
Fitting a straight line to a strongly curved relationship is the classic case.
Giving the model more capacity or better features usually fixes it.
The goal sits between the two: a model that generalises rather than memorises.
Underfitting is the quieter failure, and it often goes unnoticed.
The model is bad, and it is bad everywhere, including on training data.
Overfitting at least looks impressive somewhere first.
The usual cause is a model too simple for the pattern in front of it.
A straight line will never follow a curve, however long you train it.
Another cause is features that simply do not carry the information.
No model predicts house price from the colour of the front door.
Too much regularization does it too, by forcing the model to stay simple.
Stopping training too early gives you the same result.
The fix is more capacity, better features, or less restriction.
Diagnose it by looking at the training error first.
If the model cannot fit data it has already seen, more data will not help.
That is the key difference from overfitting, where more data usually does help.
Try a deliberately oversized model as a test.
If that also fails, your problem is the features, not the algorithm.
One question on this card. Every option explains itself once you pick, right or wrong.
55 concepts in this module, then a graded test with every answer explained.