BrainChip Provides Details of Neural Network Architecture

What I'm trying to say is that perhaps the flexibility of spiking neurons is that the training can be done "online", in the sense of an online algorithm, i.e. the data is processed as it arrives instead of existing entirely in memory beforehand. Perhaps traditional methods used by the deep learning community don't have good properties in this instance?

Still, I believe there's a lot of future in "analog-esque" computation as we probe the frontiers of what's possible to fab and the cost of straight-jacketing the digital illusion onto tinier and tinier transistors.