The cerebellar system

Thank you for the kind words, Rick—much appreciated!

Of course, the names of anatomical structures of the brain are essentially visual puns. They are named according to what they looked like (olives, worms, tufts of wool, moss, teeth, and so forth) by guys who had few or no clues as to their functions. (I say ‘guys’ advisedly: if women were involved in that, the men took the credit.) Henry has inveighed against this entrenched nomenclature, urging that at least they should be tagged neurochemically (gabaergic, etc.). So that’s one kind of ‘perceptual categories’ getting in the way.’ But besides that, the appearance of contradiction between my statement and Bill’s is an apples-and-oranges thing, where the apples of their findings are at an extremely myopic cellular and subcellular scale compared to the brain-wide scale of the oranges, their attempts to place these findings in context and paint a bigger interpretive picture.

Neuroscience knowledge of stuff going on in the nervous system has indeed advanced enormously, but that knowledge is mostly way down in the weeds of synapses and neurochemicals within specific anatomical structures and their projections. It’s when they try to integrate these findings into a more comprehensive picture that they try to fit them to theoretical concepts—not their own theoretical concepts, mind you, but rather ideas borrowed from prevalent psychological constructs (as in “the cerebellum plans movements which the motor system executes”). They recognize circuits and feedback loops, but fit them to prevailing notions in computer science and AI, such as ‘feedforward neural networks’ or an adaptive filter as e.g. proposed in

Dean, Paul, John Porrill, Carl-Fredrik Ekerot, & Henrik Jorntell (2010). The cerebellar microcircuit as an adaptive filter: experimental and computational evidence. Nat Rev Neurosci. 1.1:30-43. doi: 10.1038/nrn2756. Epub 2009 Dec 9. PMID: 19997115.

It may be that the cerebellum is structured as a neural network; in that capacity maybe it could perform the functions that I propose in my conjectures—it could be responsible for many-one extraction of higher-level signals from lower, which are then further processed in the cerebellum; for reference values for signals at diverse levels; and for deployment of error signals at higher levels to reference input functions at lower levels, and it could be doing all of these things at once. There is a huge capacity of connectivity in the cerebellum. One would be unlikely to consider these possibilities without first discerning how the numerous cerebral loops at various levels all pass through the same deep cerebellar nuclei, subject to itsy bitsy picky inhibitory signals from the cerebellum, and how the cerebellum stands aside from those loops, and the motor and somatic loops as well. And one would be unlikely to see that big-picture circuit diagram without PCT. The conventional circuit diagrams that I borrowed from Wikipedia and various articles have the projections dangling at their margins, with no closed loop.

It is unsurprising and unblameworthy that Bill did not update neuroscience picture in Chapter 9 of B:CP. He had his hands very full, and his strengths were in modeling. He said this clearly many times, for example in 2003 he said " my modeling efforts focus on what kind of control process is done, which doesn’t depend on guessing which part of the brain does it." For context, here’s the source:

I have long said that higher systems may well act by varying the parameters of lower systems as well as their reference signals. An early demo of this effect was offered by Tom Bourbon at my suggestion, 10 or 15 years ago. He set up a model in which a higher-level system monitored the average absolute value of error signal in a control system, and varied the output gain in that system to achieve minimum error. Actually, he set this up as a reorganization task, so the gain variations were done through a random walk.

More recently, I proposed a model in which an auxiliary control system (whether you should consider it “higher” or not is debatable) changes the weightings in an output function in a way that emulates the convolution theorem. It worked pretty well when embedded in the Little Man model. I called this model the “artificial cerebellum,” because of some resemblances of the algorithm to processes known to happen in the cerebellum. Of cou[rse] this doesn’t mean that the amygdala could not do something similar. However, my modeling efforts focus on what kind of control process is done, which doesn’t depend on guessing which part of the brain does it.

I recently saw another instance, which I’m not finding now, in which he more sharply expressed low gain in control of the latest developments in neuroscience, though of course something in Nature or Science might catch his eye, or more often someone would post something about ‘mirror cells’, for example (which IIRC you immediately proposed probably carry reference signals, though the social implications of that have been more elusive of general agreement).