How a network computes, why gradients vanish, and what makes depth trainable
You now have four activations and no way to compare them. The honest comparison is "which one trains better", and the backward pass does not exist yet; dl.12 to dl.15 build it. What does exist is a set of measurements you can take on an untrained network, and three of the four numbers that decide the outcome are among them.
🔒 This is a Pro lesson — the interactive figure, worked examples, quiz and practice open with Pro access.
▶ Activation Lab: Reading the Gradient Before You Train