Take an optimizer step

Update weights by moving a small amount against the gradient, controlled by the learning rate. You’ll compare plain stochastic gradient descent (SGD) with Adam, which adapts step sizes using recent gradient history.