To make this idea more concrete: why not use the brain of the rat (or any similar animal) in place of the NN, I.e. remove the brain in embryo stage and train it with visual stimuli relevant to the problem to be solved.

I’d wager that a rat brain would be more powerful than any state of the art AI systems in solving complex tasks such as this. How to sustain it biologically and train using optical paths are complex problems but I could be easier than training the animal and much more scalable: make NN layers with these brain-based nodes.