Epoch

Overview

Training a neural network means showing it the same dataset repeatedly, and an epoch is one full cycle through that data. A single pass is almost never enough: the model's weights shift only slightly with each example, so it must revisit the data many times before its internal representations settle. Epochs are how that repetition is measured, and how practitioners track where a training run stands and when it should stop.

How it works

Within an epoch, the dataset is divided into batches, and each batch produces one update to the model's weights. Once every batch has been processed, the epoch is complete and the next begins, usually with the data reshuffled so the model does not learn the order rather than the content. Progress is monitored across epochs, and training stops when performance on held-out data stops improving. Too few epochs leave a model undertrained; too many cause overfitting, where it memorizes the training data instead of learning patterns that generalize.

Why it matters

The cost of a training run is, in practice, the cost of an epoch multiplied by how many are needed, so the time each pass takes governs the economics of the whole project. That time is set by how fast the hardware can move data and compute gradients, which is why memory bandwidth, interconnect speed, and storage throughput translate directly into shorter epochs and faster iteration. On large datasets, the difference between hardware generations is measured in days per epoch.