Epoch

An epoch is one complete pass through the training dataset. Think of training as practicing from a deck of examples: an epoch means the model has been shown every card once, usually in a freshly shuffled order. It is a measure of data exposure, not a single learning update.

Datasets are too large, and neural networks too expensive, to process all examples at once. Training divides the data into mini-batches. For each batch, the network makes predictions, computes loss, backpropagates gradients, and the optimizer—such as Adam—updates the parameters. If 50,000 examples are trained with batches of 100, one epoch contains 500 training steps (parameter updates).

• Batch: a small group of examples processed together.
• Step / iteration: one optimizer update from one batch.
• Epoch: enough steps to cover the full training set once.

One epoch rarely gives a model enough opportunity to fit useful patterns. Many epochs allow repeated refinement, but eventually the model can begin memorizing quirks of the training data rather than learning patterns that generalize. Training code therefore commonly records both training loss and validation loss after each epoch. Falling training loss alongside rising validation loss is a classic sign of overfitting. Early stopping can halt training when validation performance has stopped improving.

Epoch boundaries are convenient points for practical controls: reshuffling data, saving checkpoints, evaluating on validation data, and changing the learning rate. A learning-rate schedule might reduce the rate after several stalled validation epochs, allowing a run that has plateaued to refine its solution. More epochs cost proportionally more compute and wall-clock time; they do not guarantee improvement. In frameworks such as Keras, epochs=20 requests twenty full passes, while the actual number completed can be smaller when early stopping intervenes.