home

click to place a ball · watch it roll downhill · learning rate controls step size · local minima · runs locally

learning rate
0.080
θ ← θ − α·∇L(θ) · training a neural network = rolling a ball down a loss landscape · shape determines everything
ready