Saturday, January 5, 2013

The music of working memory

An analysis of brain oscillations binding content between the frontal and parietal cortices.

I'll admit it- I am fascinated by brain waves. They seem to provide dynamic structure to the activity coursing through the static anatomy of the brain, organizing all those things we love about it, like consciousness. There is also a deep connection to music, as the the various "bands" of the neural oscillations, going in very rough terms (delta ~2Hz, theta ~4Hz, alpha ~10Hz, beta ~20Hz, gamma ~40Hz) can be seen as octaves in an internal, ongoing symphony.

But knowledge about how these oscillations work has been very slow in coming. It is obviously a very complicated system, and our scientific tools are still woefully unequal to the vast scale of tens of billions of neurons and trillions of synapses. A recent paper describes at a very gross level the function of oscillations between the frontal cortex and the posterior parietal cortex as carrying some kinds of working memory.

"We tested the hypothesis that neuronal synchronization across the fronto-parietal network carries content-specific information that contributes directly to visual working memory. The pattern of fronto-parietal synchronization should thus vary as a function of the object held in memory."

Their memory test (done in macaque monkeys) was: first, an image is displayed, then turned off, and then after a pause of one to three seconds, two images are displayed, one of which matches the first one. The pause is the important part, during which the researchers hunt through their electrode recordings for the ghosts of working memory. The monkeys finally get a reward for correctly tracking their eyes to the matching image of the pair that are displayed later. It's an extremely simple test, and one could imagine that it hardly touches the capabilities of these monkeys for remembering things.

They have inserted a series of electrodes into the macaque's brains, in the areas putatively involved in keeping the first image "in mind". These locations are the side of the prefrontal cortex, and the posterior of the parietal cortex:
Macaque brain, marked for the areas studied- lateral frontal cortex (right), and posterior parietal cortex (left).
The idea is that the first image needs to be kept in mind to rapidly match it in the test, so the researches are hunting for neural correlates of this mental content. They know it is visual content, which narrows things down slightly. They also know about prior work which has observed some brain wave synchronization between these two brain areas during focussed attention and task performance, and which involves the parietal cortex closely with visual working memory.

Their observations take the form of correlations between the oscillations detected in one location compared with those in the other. An example is shown below, where a sample electrode trace from each area is shown in B, and from that data, a graph of correlation (coherence) with respect to time and frequency (C) is developed. The correlations are detectable at 20Hz

The visual test is diagrammed in A, while a pair of local field electrode recordings is in B. dPFC = dorsal prefrontal cortex, PEC = posterior parietal cortex, caudal section of arbitrary area E.

Fine, so correlation is not so hard to establish. But causality between the two brain areas and content of what these oscillations are encoding? That is a taller order.

To start, they sift through many trials on one monkey, choosing electrode pairs that show the kind of correlation seen above, to see whether the identity of the displayed shape or its location had additional effects on the oscillation correlations. This would indicate that the relationship was related to the visual content, not just part of the more global attention system or something similar. They develop a metric called coherence selectivity index (CSI), related to mutual information theory, that can be graphed:

Coherence selectivity index showing the degree to which correlations (of electrical oscillations) between the anatomical locations are also correlated with the content of the visual image, either of location or of image identity.

This supports the idea that the oscillations being observed as related between the two brain locations also are related to the tested content- the images of the memory test, during the time when the images are not visible. I have to say that this is pretty meagre stuff, both in terms of the statistics being used to ferret signals out of messy data, and conceptually, in that the connection between this kind of wave correlation and something called "working memory" is pretty faint. I am not saying that they are misrepresenting anything, just that the technical means for figuring out what is on someone's "mind" by electrodes is still awfully primitive.

Lastly they ask questions about which parts of the brain lead these synchronizations, and which follow. They use another statistical method, of Granger causality, which allows them to classify sub-regions where they have electrodes- paired and showing image-identity correlations as discussed above- as either senders or receivers of signals. That is to say, the oscillations (and perhaps detailed spike trains) are reliably offset in one or the other direction.

The conclusions are mixed, that while the parietal areas they measure are more often "senders" than "receivers", causality goes in both directions. This follows a growing theme in brain science, where most higher-level activity happens in recurrent networks, not in linear processing streams. For example, attention is believed to be an interaction between top-down selection and bottom-up processing streams to form coalitions of synchrony. But of course, the result could also just reflect muddy data.

They also note that the relative phases of the presumptively correlated oscillations are not firmly related, which is quite different from the highly structured phase relations of neural signalling and oscillations in the memory storage system of the hippocampus. Again, this may reflect their muddy data, or perhaps suggests that such detail can not be retained in this much longer-range, yet still functional, inter-cortical synchronization.

WGC is Wiener-Granger causality, PPC is posterior parietal cortex,  PFC is prefrontal cortex, and time is as above, relative to the image memory task. Inthe lower panels, various sub-areas of the parietal cortex area are colored blue, and sub-areas of the prefrontal cortex are red. Parietal are

"Our findings demonstrate that fronto-parietal synchronization during visual working memory is widespread, task-dependent, and content-specific during the delay period."

This seems a bit strong, though perhaps if one adds in all the prior work that has been done in this network and how visual memories are handled in the parietal cortex, it makes more sense. From what I see in this paper, however, the conclusions are more preliminary than final. At any rate, it is clear that scientists are gaining (slowly) on this significant problem, and pinning down just what constitutes our various memories and thoughts.