How would a backpropagation work in such an architecture?
How would a backpropagation work in such an architecture?
Generally it is not known to my best knowledge how or whether the brain uses backpropagation (knowing that would be a huge result). However, several papers have been written on that topic with varying approaches. I encourage you to do a literature search if you want to dig deeper.
On this topic it is worth mentioning that it is not even known why the brain uses discrete signals (spikes) for communication.
In the simplest case you can feed signals 'in one end', let it do it's thing and then take the output, which you further feed through the rest of your architecture. Rewarding can be (in the simplest case) done with neurotransmitters (mostly dopamine) as it would've worked normally. You basically have to mimic something like this[0], by rewarding based on your loss function.
(of course, in reality, you'll have to do a lot more than that)