This is crazy from the perspective of "old fashioned binaries", but less so in the context of neural networks. You can do all sorts of splicing and dicing of th

This is crazy from the perspective of "old fashioned binaries", but less so in the context of neural networks. You can do all sorts of splicing and dicing of the bits in neural networks (their weights) and end up with useful networks. Dropout, for example, specifically trains a network to be resilient to having swaths of the network removed, and makes individual features in the network resilient to having a random selection of other features present or not-present. If I remember right, the original dropout paper even analogizes this to how genes have evolved to be resilient to this type of random pairing.