Earlier this month, Chris Chatham over at Developing Intelligence blogged about a recent paper by Rodney J. Douglas and Kevan A.C. Martin called "Mapping the Matrix: The Ways of the Neocortex". Chris gives a good summary of the paper, which basically refutes the idea of the cortical minicolumn as the primary functional unit of the neocortex.

Since my dissertation involves modeling the neocortex, based on the assumption that the minicolumn is the primary functional unit of the neocortex, I was interested in what Douglas and Martin had to say.

The paper gives a nice overview of the history of trying to unravel the structural and functional organization of the neocortex, discussing the early work of anatomists like Golgi, Cajal, and de No, and the later breakthroughs of physiologists such as Hubel, Wiesel, and Mountcastle. Mountcastle recorded neural activity in cats in vivo, finding common activity to stimuli in the vertical extent. This suggested a vertical functional organization to the neocortex.

But Douglas and Martin object to this organization principle based on two (weak) objections:

1) The minicolumns are not neatly packaged and discrete enough.

In referring to anatomical studies in the 70's and 80's that provided more detail about cortical connectivity, they write:

What was the elementary unit of structure that formed the vertical functional column? The palisades of cell bodies still formed their neat columns, but not only did the dendrites originating from these cell bodies spread well beyond these elementary columns, so did their axonal arbors. It became impossible on anatomical grounds to define the columnar structure of a given area, let alone explain how the visible structure gave rise to the functional phenomenon of columns.

Axons and dendrites are the extensions of a neuron, a way for a small cell to increase its surface area and connect to many other neurons. You can think of axons as output cables and dendrites as input cables. What Douglas and Martin are arguing is that with even more powerful staining and imaging techniques, we can see that the neuron bodies still organize into minicolumns, but we see a lot more connectivity than we did before. How does that invalidate the concept of columnar organization?

A little later in the paper they say:

The notion of the minicolumn or module does not properly capture the granularity or the vertical and lateral interdigitation of component neurons that seems to be the essence of the cortical circuit. This means that even in a highly specialized "column" such as those evident in the rodent barrel cortex, one cannot simply clip out a cylinder of tissue that contains the whole local circuit, for later reconstruction "in silico".

So because a minicolumn is not neatly partitioned, but highly interconnected, Douglas and Martin argue that it's not a good way to conceptualize the organization of the neocortex. But this argument could be used for any functional unit in any complex network, e.g. a computer network. When you have a functional unit embedded in a highly-connected network, it's difficult to say structurally what you consider part of that unit. Is it just the unit itself? Or is it all its incoming and outgoing connections, or a subset of them?

What makes sense when trying to make sense of a highly-interconnected, networked structure is to see how it behaves. And the functional evidence is clear. We have common, coherent activity at the minicolumn level. It doesn't matter that you can't easily slice them out with a scalpel.

The second objection is also weak:

2) The minicolumn hypothesis cannot currently explain cortical function

They say:

The difficulty of defining the column has also generated the opinion that the concept has failed as a unifying principle for cerebral cortex.

I'd say that verdict is a little premature. My guess is that Douglas and Martin wouldn't think much of the ideas of Jeff Hawkins or Lee and Mumford, who base theories of neocortical function on the minicolumn as the basic functional unit. The basic idea is that the minicolumn functions as a unit situated in a hierarchy. The minicolumn learns to become strongly active via coincidental input from lower levels in the hierarchy. Once a pattern has been learned, the minicolumn acts as an inference/prediction unit, passing inferences up the hierarchy and propagating predictions down the hierarchy.

This is one of the newer theories about neocortical function, and it's built around the concept of the minicolumn as the primary functional unit. It may be flat-out wrong, but it makes some very specific predictions, and it may yield fruit. An awful lot of theoretical and modeling work remains to be done with respect to neocortical function, too much to prematurely abandon a particular approach.

Douglas and Martin then go on to describe certain principles of cortical connectivity. They also discuss functional principles they call "Just Enough," which refers to the contribution of thalamic input necessary to exceed thresholds in cortical neurons, and "Just in Time," which refers to the principle that most cortical neurons are on "stand-by" and that the timing of inputs is crucial to the function of neocortex. I don't see how any of this is mutually exclusive with a minicolumn view of cortical function, but maybe I'm missing something.

Anyway, it was an interesting paper, worth the read for the nice overview of the history of trying to reverse engineer the neocortex, even if it some of the conclusions are suspect and paper as a whole doesn't conceptually hold together very well.