overfitting and underfitting

Overfitting is a model that has effectively memorized its training set rather than learned the underlying pattern: training loss keeps falling while validation loss stalls or rises, a widening gap between the two. Underfitting is the opposite failure -- the model is too simple, or hasn't trained long enough, to fit even the training data well, so both training and validation loss stay high and close together. A model that generalizes well does neither: it does on new data nearly as well as it does on the data it trained on.