Micah Blumberg wrote:
I am also making lists of things in my notes, to research later, there are a lot of questions, I wonder if his neurology has been fact checked by professional neurologists. His whole book needs to be fact checked, there are a lot of assumptions I think, assumption that may not based on hard evidence, or that have only anecdotal evidence. One of the big flawed ideas in neuroscience is the assumption that if something (like long term memory) stops working because something in a specific region (like the hypothalamus) is broken or damaged then that broken part is responsible for creating whatever feature has stopped working (memory) but this is correlation mixed with causation, and the result is flawed neuroscience. He may not be aware of it. I also have misgivings about the notion that "connections in pattern modules" are genetically predetermined, he cites Markram from Blue Brain as supporting this idea, but Markram released a video this year showing that neurons form connections randomly. Kurzweil also says in his book that there are quadrillions of connections, and the genome just isn't complex enough to have a genetic map for all those connections. So I really think Kurzweil is not connecting the dots right in front of his face when he claims that the connections are genetically predetermined. It's confusing, because I wonder if I'm missing the point, but I'm pretty sure he is not making the connection, I will keep reading and researching. Let me know your opinions on this please.
It's late here, and I've been stressed out because my card bounced when I tried to buy groceries today, so take this as a preliminary posting.
Markram and kurzweil's reports do seem to be reconcilable. Markram showed rules for how the networks are patterned, kurzweil claimed that there were discreet subunits that behaved as computing elements and didn't change much with learning. I don't see much conflict between those views because I don't think Kurzweil claimed that the networks were stereotypical in organization, only that they weren't very adaptive.
If there are 10,000 neurons in each column, then there should be about 100 kurzweil-nets in each column. Now the Markram patterning would pre-dispose the kurzweil-net to learn a different pattern from it's neighbors. For the same reason, conventional neural network simulations have to seed the networks with random data otherwise different sub-circuits might train on the same pattern, which would be bad.
A fair amount of work remains to precisely characterize what these assemblies do and how their learned interconnections leads to pattern-based processing...
Maybe there are more answers in the other 200 pages of the book. -- E T F N H E D E D Powers are not rights. ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
