@s8n the paper for e-prop showed you can unroll the derivatives as in backpropagation through time, but then you can actually approximate over the derivatives through time, and replace them with basically a single protein marker. which enables backprop to operate online (and the same protein traces appear to exist in real brains.) it doesn't work in super deep contexts, which is fine because the brain doesn't actually have those.

its looking like the way intelligent systems work is actually nowhere near most ANN research; it seems to be mostly dumbly self organizing maps and a very thin layer that actually adapts to inputs, but

i stopped editing my reference manual when factorio space age came out. its still on the drive i just haven't been cramming the papers all night.