One training iteration

Training usually groups several examples into a small bundle called a batch.

For one batch, the network performs the complete learning loop:

Predict. Send every example forward through the layers.

Measure. Compare the predictions with the correct answers and calculate the error.

Work backward. Use backpropagation to calculate how each parameter contributed to that error.

Adjust. Use those calculations to update the parameters once.

One trip through those four steps is called a training . It is one small learning step, not the whole training process.