epoch 0 | train –
● A ● B ○ –

what to notice

One hidden layer is the whole story here. Strip it away (take the layer count to zero) and watch the boundary stay a straight line at roughly half wrong forever - the collapse argument from Maths of Learning ch2, running live. Put it back and the four corners separate in seconds.

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 rate0.05