To find the bottom of a surface you can only feel underfoot, step downhill: the gradient ∇f points the way f rises fastest, so move a little against it, x ← x − η·∇f, and repeat until the ground is flat. Training a neural network is this, on a surface with millions of dimensions.