Notebooks

Neural Control of Action, Especially of Rapid Motor Sequences

23 Jun 2014 09:36

I got interested in this subject because of William Calvin's "projectile theory of consciousness" or "throwing theory of thought"; his reasoning is something like this. Throwing something at a small moving target (even a large, distant moving target) needs very accurate motor control, particularly timing; doubling the distance of the throw can reduce the acceptable launch window by a factor of eight. In fact it happens too fast to be controlled by feedback from the muscles, and the precision seems to be beyond that of a single neuron, which is apt to be quite noisy. The obvious solution is to average lots of the neurons. The obvious problem is that, normally, noise will go down as the square of the number of neurons you average over, so an eight-fold reduction in noise needs a sixty-four fold increase in the number of neurons, and this is a lot of metabolically expensive brain to dedicate to lobbing rocks. Calvin's ingenious idea is that the needed neurons could be recruited as needed, and released for other tasks when not harassing the wild-life. He supposes that there are many different "sequencing tracks," each one of which carries one particular candidate sequence of muscle commands, and that these are "shaped up" by a Darwinian process, i.e. better-rated ones get copied into more tracks at the expense of the less favored, with occasional mutations, so that eventually all the tracks contain clones or near-clones of a single, highly-rated sequence, which is then executed.

Now once you have the neural machinery for this kind of sequence-generation, it's not such a stretch to imagine that your descendants will turn it to other uses. And a lot of what we do that is characteristically human --- speech, music, stories, poetry, planning, consciousness, composition of Web pages (I never said these were all admirable) ---- involves combining units into strings, with a good deal of trial and error, some of it even conscious. Calvin suggests that, in fact, Darwin Machines which originally evolved for throwing and similar ballistic motion were adapted to all these roles. I think this is a neat idea which ought to be followed up, and in particular I'd like to know if it leads to any specific predictions about the effects of brain lesions (say, damage to the planning tracks distrupting all those functions, but damage to the connections to memory of one or another leaving the rest intact), and how, if at all, it can be squared with what we know about grammar. (I'd also like to know whether it has any connection with the ideas about "strings and sequences" William James puts forth in the last chapter of his Psychology, but that's because I suspect James anticipated everything.) Of course even if Calvin's specific idea about how action is controlled turns out to be wrong, a more general notion that higher cognitive functions use the same kind of machinery as motion control could still be true.

Originally, Calvin proposed that the evaluation was done against some kind of stored memories of what had worked well in the past. More recent work in cognitive neuroscience, especially functional neuroimaging, shows that learning to use tools or perform actions creates "forward models" in the brain, circuits which can evaluate the consequences of possible lines of action. These would serve very nicely as the evaluators in his Darwin machines, at least for motion.


Notebooks: