But can we use the idea inside the training loop, as an optimizer instead of (or next to) e.g. Adam?