Topological reuse · Figure 3

A lens on the clock learns a + b + 3 in 100 steps. A new network needs 4,400, and from 96 examples it never learns it.

What you are looking at. Across: training steps, on a log scale, so the first hundred steps and the last ten thousand both fit. Up: how many of the 401 pairs the network never trained on it gets right. Every line is the same new task, learned from the same examples, by a different route. Seed 10; the dotted line at 95% is where this project calls a task learned.