Epoch
lastmod 2026-05-13created 2022-11-259 edits
Epoch (pronounced "epok"; era)
in Deep learning
- In the Naver English dictionary, Epoch means "an era (in which important events or changes occurred)".
- The number of times every piece of data in the training dataset has passed through the model once; it means the number of times the entire training dataset is trained on.
- 1 epoch means the entire training dataset has been applied to a neural network and has passed through the network once via forward propagation and backpropagation.
- That is, 10 epochs means the model was trained on training dataset A 10 times.
- The more epochs, the more varied random weights are tried during training, so the probability of finding suitable parameters increases. (That is, the loss value goes down.)
- However, if the number of epochs is raised too much, the model may overfit the training dataset and fail to make proper predictions on other data.
Short version
In deep learning, 1 Epoch means the entire training dataset has fully passed (Forward & Backward Pass) through the neural network once.
Key points:
- Definition: going through all the data exactly once during training is called 1 epoch.
- Structure: the dataset is usually split into smaller units, each called a Batch, for training, and 1 epoch is complete once every batch has been used for training once.
- Formula:
Iterations per Epoch = Dataset Size / Batch Size
Why train for multiple epochs (Multi-epoch)?
A single epoch is not enough for a neural network to sufficiently learn (converge on) the complex patterns in the data, so the same dataset is trained on repeatedly to optimize the weights.
However, too many epochs can cause Korean: 과적합 (Overfitting), so it is important to stop training at the right point.
Batch size vs Epoch vs Iteration
See Batch size#Batch size vs Epoch vs Iteration.