Although the anatomical and functional columnarity of the neocortex has never been in doubt, the size, cell composition, synaptic organization, expression of signaling molecules, and function of various types of ‘‘columns’’ are dramatically different. Columns could be defined by cell constellation, pattern of connectivity, myelin content, staining property, magnitude of gene expression, or functional properties.
The cortical column: a structure without a function
I remember reading these papers years ago. I can’t recall the details, but I do recall being unconvinced. Again, not recalling these papers in particular, there are two main arguments against Mountcastle’s proposal.
We haven’t figured out what a column does, so maybe it doesn’t do anything.
We find anatomical, physiological, etc. differences between columns, suggesting they don’t share common function.
That may be a bit of a simplification, but not too much. The second point is the only one worth debating. The amount of commonality between columns, imo, overwhelms the differences. Nothing I have read suggests otherwise. Mountcastle, made his argument on multiple lines of evidence. Two of my favorites are…The rapid expansion of the human neocortex vs. our hominid relatives occurred very rapidly, tens of thousands of years. Not enough time to evolve substantial new function. Also, the amazing flexibility we have to learn things for which there were no evolutionary survival pressures is another.
is perhaps more relevant than it first seems if we consider it points out the assumption of a one-one mapping between structure and function. There seems to be a strong case that living systems do not have this simple mapping (unlike machines). But that is more a question of whether reverse engineering is the best approach and Numenta has placed its bet. I admire the commitment.
The mainstream research on canonical circuits attempts to elaborate the following sort of explanation:
P1 the cortex is remarkably uniform;
P2 the cortex is the main site of a bewilderingly variety of functions.
EC there must be a canonical circuit common all over the cortex, able to perform many different functions.
Explanations of the sort EC, with suffix C for “circuit”, are doomed to failure, as I will argue in §3.3. If the set of premises is enforced with plasticity, a different sort of explanation can be offered:
P1 the cortex is remarkably uniform;
P2 the cortex is the main site of a bewilderingly variety of functions.
P3 the cortex is characterized by a remarkable plasticity.
ED there must be a strategy common all over the cortex, which enables a basic circuit to gradually change and develop a wide variety of functions, depending on the input patterns.
The explanatory limits of EC can be well interpreted in the light of timescales, following Marom (2010). Canonical circuits are abstracted over a highly simplified temporal manifold, which takes care of one or just few short timescales, neglecting slower timescales at which important circuital adaptations take place.
…all canonical solutions proposed so far have overlooked the dimension of cortical development due to plasticity, which is the main source of its computational flexibility, as supported from the reviewed evidences. Thus, a successful road towards a canonical explanation of the cortex paradox should be better construed as a mixed explanation of both the constituents essential for its computational power, and the developmental account of how cortical maps achieve their mature functions.
I find these arguments compelling. I take ‘basic circuit’ to mean some kind of computational unit, say of the same order of complexity as a Raspberry Pi. 100 million of those at 1000 MIPS each gives you the computational power of a mammal, approaching that of a human brain. A ‘circuit’ could be a ‘column’.
This ability for a fixed unit of ‘hardware’ to ‘change and develop a wide variety of functions’ is my definition of software.
I think you imply that this is passive, reactive, perhaps akin to adjusting weights in response to ‘input patterns’. I think it likely to be more a matter of active selection between pre-coded functions stored in some kind of inherited library of such functions.
So there we have two competing hypotheses. It should be possible to devise experiments to choose between them.