AI Glossary

Epoch

One complete pass through the entire training dataset during model training.

Overview

When people first learn about machine learning, they often assume a model only needs to see training data once in order to learn.

In reality, learning usually requires repetition.

Just as a student may need to review material multiple times before fully understanding it, a machine learning model typically needs to examine the same dataset many times before it can identify useful patterns and make reliable predictions.

This is where the concept of an epoch becomes important.

An epoch represents one complete pass through the entire training dataset. If a dataset contains 10,000 training examples, one epoch means the model has processed all 10,000 examples once.

However, one pass is rarely enough.

During the early stages of training, the model is still learning basic relationships within the data. Its predictions may be poor and its error rates relatively high. Each additional epoch provides another opportunity to learn, adjust, and improve.

A helpful way to think about epochs is studying for an exam. Reading a textbook chapter once may provide some understanding, but reviewing the material multiple times often leads to better retention and deeper comprehension.

Machine learning models work in a similar way.

The number of epochs used during training is carefully chosen because too few may result in poor learning, while too many can sometimes lead to overfitting, where the model becomes overly specialized to the training data.

Understanding epochs helps explain why model training is often described as a gradual process rather than a single event.

Why It Matters

Epochs determine how many opportunities a model has to learn from training data and improve its performance.

Real-World Example

A company training an image recognition model may run dozens of epochs so the model can repeatedly review images and improve its ability to identify objects correctly.

Related Concepts

Batch Size
Model Training
Gradient Descent
Backpropagation
Neural Network

Related Glossary Terms