Underfitting
In one sentence Underfitting is when a model is too simple to capture the real pattern, so it performs badly even on its own training data.
Updated
Underfitting is when a model is too simple for the pattern in the data, so it is wrong on the training set itself.
If overfitting is the student who memorised last year's paper, underfitting is the student who only skimmed the first chapter. Ask them anything — old questions or new — and the answers are equally poor. That is the telltale sign: an underfit model is bad everywhere, including on data it has already seen.
The classic picture is forcing a straight line through data that curves. No amount of extra training fixes it, because the model does not have the capacity to bend.
overfit : train score 99%, test score 60% ← memorised
good fit : train score 91%, test score 89%
underfit : train score 62%, test score 61% ← too simpleThe fixes all add capacity or information: a bigger or more flexible model, better features via feature-engineering, training for longer, or easing off regularization that was set too aggressively. Diagnose before treating — check the training score first. If training accuracy is already high, your problem is overfitting, and these fixes will make it worse.
Where to go next
- Full lesson: Overfitting and underfitting
- Related terms: overfitting, bias-variance-tradeoff, loss-function, feature-engineering