ML

02 / GRADIENT DESCENT

Small steps. Less error.

Adjust the learning rate. Watch the loss change.

LOSS LANDSCAPEDRAG TO ROTATE

Steps0
Mean squared error—
θ ← θ − η ∇L

Backpropagation computes the gradients. Gradient descent updates the weights and biases.
The learning rate η controls the step size. Large steps can overshoot.