Gradient
Also: gradients, gradient vector
For f(x₁, …, xₙ), ∇f collects ∂f/∂xᵢ for every input. Its direction is where f rises fastest and its length is how fast; training moves parameters against the gradient of the loss.
∇f = (∂f/∂x₁, …, ∂f/∂xₙ)
Example
For f(x, y) = x² + y² at (3, 4), ∇f = (2x, 2y) = (6, 8): it points straight away from the minimum at (0, 0), so stepping the opposite way goes downhill.