Like he argues that there's a big difference between software running on hardware and our brains, in that the neurons are dynamic, bathed in chemicals etc while software doesn't interact with the hardware in any substantial way (ie doesn't physically change it).
This is true, but to me it seems much more like optimizations rather than a truly fundamental difference. Surely you could model a neuron in software with non-trivial internal state and which also responds differently based on "ambient parameters" (ala bathed in chemicals).
It would be much harder to train, at least using the methods we do now, but I fail to see a fundamental difference.
>Surely you could model a neuron in software with non-trivial internal state and which also responds differently based on "ambient parameters" (ala bathed in chemicals).
Oh you could. and it's been done. But it's much harder to train and seemingly for no real performance benefit other than "like the brain". Backpropagation is just really efficient.
This is true, but to me it seems much more like optimizations rather than a truly fundamental difference. Surely you could model a neuron in software with non-trivial internal state and which also responds differently based on "ambient parameters" (ala bathed in chemicals).
It would be much harder to train, at least using the methods we do now, but I fail to see a fundamental difference.