This paper, and its "haphazard" J.Neurophysiol predecessor, is interesting and beautifully illustrated, but perhaps somewhat misleading, in that it revives the rather unsatisfactory notion that much of the detail of cortical circuitry might be random. Though the paper is careful to say that it only proposes a statistical view of the initial "seeding" of wiring that precedes the critical period, the title, and some of the presentation, lay claim to a
more general view of cortical architecture.
The most obvious, but not the most important, problem with the paper is that, as it admits in the Discussion, it assumes that a rather precise retinotopic map is achieved in the
subplate prior to the establishment of the thalamocortical connections proper. In essence, all the machinery (such as Hebbian adjustment) that was previously proposed to directly achieve thalamocortical wiring is swept under the carpet in the statistical model. It's still there, but no longer out in the open.
Almost equally obvious is that the core idea of the paper is Soodak's, but here it is very nicely explained, illustrated and elaborated.
But the more important point is that the paper distinctly, and wrongly, conveys the impression that the statistical seeding process does most of the "heavy lifting" in the
wiring, with activity-dependent learning, or possibly molecular guidance, merely "refining"
the initial blueprint. In particular, the obvious woeful defect of the statistical model, that the predicted "aspect-ratio" of the RFs is only half that seen in mature simple cells
(see the "haphazard" paper) is lost in a welter of claimed minor successes. But the point of
models is not to understand map formation etc, but how the cortex analyses the world. A cat that cannot quickly detect mice (because a mouse's edges are nearly invisible) will die, but one that does not have orientation maps will probably survive (as do mice, squirrels etc). So by focussing on all the results that neuroscientists have amassed, and not on the task that the cat's cortex must solve, the paper seems more impressive than it really is.
This point is really vital: "refining" the initial map (whether this is achieved by activity-dependent mechanisms in the subplate and/or the cortex itself) is not a detail to
be filled in by minor adjustment after the major blueprint has been implemented (like a new homeowner painting the walls) but the essence of the problem itself.
Let me give a simple example: suppose one were to propose a "method" for finding primes: just pick a number at random and it will be close to a prime. In other words, finding primes is easy, since one gets very close just by guessing (eg 10), and a prime is trivially close
(11). Obviously, while this works approximately for numbers under 20, it is not merely bad in the general case, but, by definition, as bad as any method can possibly be. "Statistical wiring" (more earthily but less seductively called "haphazard" in the earlier paper) may be the way the initial wiring is done (though at the cost of implausibly accurate wiring in the subplate), but this is about as relevent to the final outcome as the stavelines of Beethoven's blank music paper was to the symphony that emerged.
Specifically, the problem is that a mature simple cell may be connected to 30 out of several hundred geniculate cells to which it could in principle be connected (i.e. whose axons and dendrites intermingle). There are superastronomical numbers of ways in which this selection
could be made, but (as Alonso et al have shown) the final connections are precise,
resulting in the mature aspect ratio etc. While at the beginning, when only a few geniculate
cells are connected, some degree of orientation tuning is almost trivially easy to synthesise (like picking near primes below 20), it is an almost impossibly difficult task to
wire up 30 inputs selectively. It's the old "curse of dimensionality" again: what is trivial
at low dimensions becomes rapidly impossible at high dimensions. So the last bit of the wiring, the "refinement", is actually almost all of the problem, and the first bit, the
"seeding", can be done by almost any method. Of course Ringach might argue that the last "refinements"are not done precisely, and the outcome is still dominated by the initial seeding, but this probably reflects merely our current ignorance about what exactly 'simple" cells encode.
If anyone really feels that the major architecture of the cortex is set "statistically" and is merely "refined" by learning, I suggest that they ask a baby to read, and comment on, Ringach's paper!