Overfitting
When a model memorizes the training data’s noise instead of the pattern, so it aces training but fails on new data.
Overfitting is when a model fits the training set too closely — capturing its noise and quirks rather than the underlying signal — so training accuracy keeps rising while performance on unseen data gets worse.
Worked example: the tell is a widening gap between training and validation loss — training loss falls while validation loss bottoms out then climbs; that turning point is where you should have stopped. Gotcha: the fixes are more data, regularization (dropout, weight decay), and early stopping — not a bigger model, which usually overfits harder unless it is regularized.