Mini-Batch

A small subset of the training data used in a single iteration of gradient descent. Training on mini-batches rather than the full dataset at once makes training faster and more memory-efficient, while still providing a good enough estimate of the true gradient.