Labs · Training

Forward and backward

A small MLP run forward, its loss sent back through the chain rule, every gradient checked against the definition of a derivative, and one step taken.

\delta_{l-1} = (\delta_l W_l^\top) \odot \sigma'(z_{l-1})

3 → 4 → 4 → 3 · a batch of 4

Smooth everywhere: every gradient non-zero, the check agrees to 1e-6.

Forward, backward, checked, stepped

Reads /api/labs/backprop — loading

Computing…