Overfitting happens when a model learns its training data too well, capturing not just the real patterns but also the random noise and quirks of that specific data. As a result, it performs strongly on examples it has seen but poorly on new ones. The “Regularization for Deep Learning” chapter of Goodfellow, Bengio, and Courville’s “Deep Learning” centers on this generalization gap and the techniques used to close it.
A common remedy is regularization, methods that discourage the model from becoming overly complex. One widely used technique covered in that chapter is dropout, which randomly switches off parts of the network during training so it cannot rely too heavily on any single feature. Other defenses include collecting more data and stopping training before the model starts memorizing.
The opposite failure, underfitting, occurs when a model is too simple to capture the real patterns. Good machine learning balances the two so the model generalizes.
Why business readers should care: Overfitting is the most common reason an AI model that “looked great in testing” disappoints in production. Insisting that performance be measured on fresh, held-out data, not the data used for training, is a basic safeguard against being misled.