Overfitting is the phenomenon in which a machine learning model fits its training data too closely, learning noise and incidental features rather than the general patterns in the data. As a result the model performs well on the training data but worse on data it has not seen. The opposite case, in which a model is too simple to fit even the training data, is called underfitting.
It typically occurs when the data is small relative to the model's complexity or when training runs too long. To counter it, performance is measured on validation data separate from the training set, and techniques such as regularization, dropout, early stopping, and data augmentation are used. The term also describes a model or system being tuned repeatedly against the same benchmark until it fits only that benchmark.