Just say we have a button to tap to say "wrong output" you needent rerandomize the whole nn space, when it is being unsatisfactory due to input, just what was involved with the previous input.
Input diverging through to output, thats where the error would be.
Is this what is done already?
Also, i was thinking you could use positive reinforcement and negative reinforcement, and give each synapse a value for how "permanent" it is, then you could randomize the synapses with the least positive reinforcement over the cells which have been more correct most of the time?
Keeping the most correct functioning part of the brain stiff, is how we reuse old rules learnt.
well, you could light up all the neurons that fired in a time step, they are the area of importance, you could isolate it somehow... i dont know. maybe it doesnt work...