epoch 0 | train
A B

what to notice

The loss curve is the whole show: it climbs. Every step overshoots the bottom and lands higher than it took off, exactly as the update rule predicts once the step size crosses the valley's threshold. Nudge the learning rate down and find the edge where training comes back to life.

instruments

loss

log scale
– train -- test

gradient magnitude by layer

Gradients shrink as they travel back – the products-of-slopes fact from Maths of Learning ch3, live.

permalink

learning rate1.4