By my rough estimate the brain has 1e11 neurons running at a cycle time of around 1e3, executing instructions that would take (say) 1e2 CPU cycles to emulate. That comes out at a ‘power’ of 16.
A desktop CPU (i7 8 core) runs at about 300K MIPS (3e9), a gamer GPU at (say) 2 TFLOPS (2e12), not even close. High end super computers run in the order of 100 PFLOPS (1e17), which is roughly similar.
The biological brain is far better at parallelisation than any computer we can build. We may still be a bit short of the necessary computer power if that’s what matters.
Seems likely there are two or more different mechanisms. HTM relates directly to storing sequences as SDRs represented by synapses, but these are easily overwritten by new inputs. A longer term mechanism would more likely use DNA, since mechanisms already exist for reading and writing (and do not for proteins). Plenty of work going on, but no solid breakthrough that I’ve heard of.
Absolutely agree. Why not to stick to the very well established facts first (D.Hebb’s principle, V.Mountcastle one algorithm and etc) and then try to build on them as axioms.
Sequences should ideally be stored without overwriting older ones. I think J.Hawkins knows that and the Numenta team is implementing exactly that. Otherwise there’s no way to memorize significant (emotionally strong) sequences even in the presence of similar ones (eg specific car/cat vs their category “images”).
When people say “practice makes perfect” (or practice makes permanent), long term memory building seems to solidify gradually, at a place beyond the neocortex/frontal lobe/consciousness. In Cerebellum? most likely. Gradually establishing permant synaptic connections? I would believe so, like foot traffic gradually shaping a clear pathway across a lawn.
When people remember things for the short term, it’s mostly (with obvious exceptions – we do retain certain short term memories subconsciouly, without noticing/awareness) happening in the frontal lobe/consciousness and/or hippocampus … neuroscientists have already discovered short term synapses forming and dissolving, without yet a comprehensive theory/conjecture on how it could relate to short term memory forming & erasing … but the connection (pun intended) seems quite straightforward if indeed that’s how STM is accomplished in the brain, and Numenta’s SDR way seems way more complicated, making one wonder whether Occam’s razor principle could apply here.
I’m very very interested in seeing a breakthrough in this field (hopefully in my lifetime), so I am keeping my finger crossed …