From a Residual Network to a Neural ODE

A residual network updates its features one layer at a time: x next equals x plus one over L times f of x. The floor is the feature plane, and this f turns every point and pulls it inward.

Stack the layers upward, with depth running from zero to one. With four layers, each point takes four big straight steps, and its path is jagged.

The rest of the narration is in the book's Companion.