How a network computes, why gradients vanish, and what makes depth trainable
Softmax has a property no elementwise activation has: its outputs are tied to each other. Every probability is divided by the same total, so raising one logit does not only raise its own probability. It lowers all the others, because the shares have to keep adding to 1.
🔒 This is a Pro lesson — the interactive figure, worked examples, quiz and practice open with Pro access.
▶ The Softmax Jacobian