underfitting
/ UN-dur-FIT-ing /
Underfitting is what happens when a model is too simple to capture the real pattern in the data — so simple that it does poorly even on the examples it trained on, let alone new ones. Picture trying to summarize a winding mountain road with a single straight line: no matter how carefully you place the line, it just can't follow the curves. The model isn't memorizing the wrong things; it lacks the flexibility to learn the right things in the first place.
It is the mirror image of overfitting, and you spot it by a tell-tale sign: poor performance on the training data itself. (Overfitting, by contrast, looks great on training data and bad on new data.) Underfitting usually means the model's capacity is too low for the problem — it has too few parameters, an overly rigid form, or it was stopped before it had a chance to learn. In bias-variance language, an underfit model has high bias: it is consistently, systematically off.
The fixes pull in the opposite direction from the fixes for overfitting. Give the model more capacity (more parameters, a richer form), feed it more informative features, train it longer, or ease off heavy-handed regularization. The reason underfitting deserves its own name is that it is easy to overlook: a beginner who hears endless warnings about overfitting may strangle their model into being too simple to be useful. Both errors waste a model; good practice steers between them.
You fit a straight line to predict ice-cream sales from temperature, but real sales jump sharply once it's hot, then plateau. The line can't bend, so it under-predicts on hot days and over-predicts on warm ones — wrong on the training data and wrong on new days. Allowing a curve fixes it.
Too simple to follow the pattern — bad even on the data it trained on.
The clearest signal of underfitting is poor accuracy on the training data itself. If a model can't even do well on the examples it studied, the cure is more capacity or better features — not more data, and certainly not more regularization.