One pass is not training

Suppose the training collection contains 1,000 examples. The network handles them in batches of 100.

It processes the first batch and updates its parameters once. That is one iteration. Then it moves to the next batch.

After 10 iterations, the network has seen all 1,000 examples once. One complete pass through the training collection is called an .

But the parameters have only taken ten small learning steps. Training usually reshuffles the examples and begins another epoch, then another.

The examples are reused, but the parameters are slightly different each time. The scale comes from repeating the same learning loop across many batches and many epochs.