Thoughts on The Cat is out of the Bag


I comment here as a computer designer. I justify this by analogy with the anatomist studying the eye, but with a background in classical optics. Nature found, for us and many other animals, what may well be the only(!!) good design for an eye when it discovered the lens. Perhaps digital circuits as deployed in today’s computers, are the unique good way to do information processing, such as the brain must do. The analogies between computer circuits and neurons has been remarked upon for 70 years. The uniqueness of the lens, as a design, may or may not be replicated in the realm of information processing, but I fear that too many researchers are betting against that possibility. I believe that there is evidence that the brain is not clocked. Some computer designers believe that computers would be better if they were not clocked; indeed, some are not.

Second generality: I think that accumulated research on brain anatomy of simpler animals may be a gold mine that will yield immense insights about purposes in brain features. For instance: “When did the cortex begin to fold?”, “When did hypercolumns appear”.

I am glad that they adopt statistical connections. I suspect that our DNA is devoted in large part to specifying such statistics. It is written in some language that we don’t yet understand, but have long term hope of understanding. We can now only speculate on the expressiveness of that language. Perhaps we can get a handle on the actual statistics bypassing the DNA language. I don’t know which is the better tack. Either of these tacks is to approach a computer acting like a human by simulating its neurons. This is not the only was to produce such a computer and may not be the most efficient. At least it seems plausibly possible in the next few decades.

It is perhaps not plausible for uploading, however, for that may necessarily depend on capturing specific synaptic connections.

I have heard some claim that such modeling must fail because we cannot afford to simulate various delays and nonlinear dependencies within the tree of dendrites of a neuron. This may be true of the strictest simulation, but I suspect that the computational model may actually surpass a model with such fidelity. In short I suspect that nature has largely failed to exploit the cruddy circuit quality — but I speak as an optimistic computer engineer.

Section 2.1: “Neuro Dynamics” is quite good, covering most of what I knew while clarifying a few things that I was unsure of.

I think that Hebbian learning is vital, but perhaps not the whole learning story. I am excited to think that Kanerva like memory patterns may be observed, even if not inserted.

Section 2.2 is nearly as good; I suppose that the ‘areas’ referred to are the classical areas such a Wernica’s, Broca’s, the visual, etc. I am very curious about wiring between hypercolumns. I think such broad outlines should be included here, if known.

It strikes me that this software could be installed in one of the cloud computer systems to be made available to those with hypotheses to test. The cost of such tests would would range from small company budgets to big company budgets. Perhaps IBM should build a cloud.

It is not clear to me whether the ‘thalamocortical module’ is meant to match some natural subdivision of the brain, or is merely a programming or hardware construct.

Outrageous Speculatons:

Much has been made philosophically from the homogeneity of the cortex. This projects implicitly exploits this. When humans try to understand complex things, even computer programs that they are writing, they find it necessary to ‘chunk’ the concepts so as not to have to understand everything at once. Nature had the same problem, but for different reasons. It couldn’t find a custom solution to every little thing that popped up the environment and often had to exapt other solutions—sort of like a programmer exploiting general subroutines. At least two things forced this: DNA load, and evolutionary time. I think the hypercolumns are such a chunk. Now if we could only understand the ‘function’ of a hypercolumn. It does not follow from my speculation that such a function will be comprehensible to us.

After a quick read I fail to see how connection information is recorded. Perhaps with repeatable pseudo-random message routing. Each synapse would then need to designate pre and post neurons.