Nafnlaus 🇮🇸 🇺🇦 on Nostr: npub1vypq6…xaghx ... differently. We seem to be something more like non-gaussian ...

npub1vypq6jrs2melhg7y3ug2jn6dguqs6kzxhmtcd3a3qezelautzyvquxaghx (npub1vyp…aghx) ... differently. We seem to be something more like non-gaussian PCNs than DNNs, with neurons being like autoencoders, trying to adjust their weights so that their pulses match the weighted firing rates of their downstream connections. And PCNs need not be organized into neat layers and can learn in realtime. We don't use them in ANNs because backpropagation is somewhat more efficient, but if we go to neuromorphic hardware, it's probably going to have to be PCN-based.