Epoch
In short: In machine learning, an epoch is one complete run of the training data set through the algorithm, from start to finish.
In machine learning, an epoch is one complete run of the training data set through the algorithm, from start to finish. Training a model usually takes many epochs, with the model adjusting its internal parameters a little after each batch of examples.
Why it matters
The number of epochs is one of the main settings in training. With too few, the model has not yet learned the patterns in the data (underfitting); with too many, it starts to memorize the training examples and performs worse on new data (overfitting). Developers therefore measure the model's error on a separate validation set after each epoch and often stop training once that error stops improving, a technique called early stopping. Crypto readers meet the word in another sense too: proof-of-stake blockchains use epoch for a fixed period of blocks or slots that organizes validator duties and rewards.
Example
A model is trained on 1,000 labeled images, processed in batches of 100. One epoch therefore takes 10 training steps, and after five epochs the model has seen every image five times. If its accuracy on the validation set improves until the third epoch and then declines, the developer keeps the version saved after the third epoch.