[Nitpicker]

I know there are creatures with very few very large neurons such as a particular sea snail. And these days we have better and better ways to see what is going on inside living cells.

But do we have any clue what specific reactions and metabolic pathways are involved in normal neuronal behavior? Do we know anything about the specific molecules, including enzymes, which are involved in responding to activity in the synapses? Do we have a clue what differs between an excitatory and inhibitory synaptic events?

We might actually have to know something about what neurons do in order to make our models mimic living brains. What basis to we have to consider such questions? Numenta-Phil suggests that an overview of HTM would notice that it paid more attention to the connections and arrangement of neurons than to the actual operation of them. Might we need to have both?

[joel]

>>Do we have a clue

Not sure about you, but neurophysiologists know very much, doctors know a great deal, and lay people who've done a little reading and watched a few Discovery-channel shows have at least some knowledge about neuron cell metabolism, membrane electrochemical dynamics, and synaptic mechanisms.

If I understand it at all, I think Numenta's technology does not attempt to mimic individual neurons or even cortical microcolumns, but rather cortical columns. A cortical column is a relatively large-scale structure in the context of brain organization, with many thousands of neurons. So they feel they can abstract away from the details of individual cells.

Mr. Nitpicker, aside from the books already mentioned, you might enjoy Churchland's books:

"Neurophylosophy" Churchland MIT press 1986
"The computational brain" Churchland, Sejnowski MIT press 1992

Personally, I enjoy trying to work a bit closer to the cellular details then the HTM model. Probably the Numenta engineers would as well, but one very quickly runs into performance limitations of the computers doing the grinding. Remember they're trying to code up something usefull rather than something merely fascinating. The architecture I'm playing with has 3 layers of linear-summing threshold units (a very simple neuron model). I'm running on a modern 3GHz linux box, and the sims bog down when I make the layers much larger than 25x25 units. Obviously I won't be solving real-world problems with a network that size.

Google "blue brain" if you want to hear about a project that attempts to include as much anatomical realism as is currently possible, using big-iron computing.