>>3
The problem with mammalian brains is that there's A LOT of duplication and redundancy for error correction, which is no longer necessary when you transition to digital electronic computation. Simulating neural networks for all subsystems is overkill, you need merely implement the higher-order function of various subsystems and modalities. We know how the eyes and visual cortex work to some degree, they integrate the input signal and perform second-order Lagrangian differentiation spatially and temporally. You don't need a neural network to implement that, just a stream-lined DSP pipeline or multiple parallel pipelines.
We're starting to see applications of this approach showing up in research vehicles that can drive themselves, simple object recognizers, and UAV visual intelligence systems.