A step is one update cycle during model training, typically consisting of a forward pass and a backward pass over a batch of data. Training progress is often measured in steps rather than only in epochs.
A forward pass and backward pass of one batch. See backpropagation for more information on the forward pass and backward pass.