How a network computes, why gradients vanish, and what makes depth trainable
optim.14 hands you three instruments: the loss curve, the gradient norm, and the first NaN. They are good at telling you that a run is broken and roughly what kind of broken. None of them can tell you where, because each of them is a single number summed over the whole model.
🔒 This is a Pro lesson — the interactive figure, worked examples, quiz and practice open with Pro access.
▶ Diagnosing a Broken Net: Per-Layer Statistics