Saturday, November 7, 2009

Gray matter talks

New work on how the brain computes speech production.

A recent issue of Science had a few interesting articles on neuroscience, including one on how Broca's area in the brain processes language. But another article on functional brain imaging brought up a factoid that makes for an interesting introduction:
"Questions about functional segregation are constrained by the resolution of fMRI. For example, a voxel (volume element of several mm cubed) contains on average 5.5 million neurons, 10e10 synapses, 22 km of dendrites, and 220 km of axons."
Wow- I didn't know the brain was quite so dense. So why aren't we smarter? Honestly, one gets the impression that the brain is not very efficiently designed. No wonder vast regions of the brain can be destroyed in dementia before much of a deficit is noticed.

The current work (accompanied by a review) has its origin in rare people who have epilepsy and who get the unnerving procedure of having electrodes stuck into their brains. I have no idea why such an invasive test is done, but for Sahin et al., it was a godsend, allowing them to do the kind of electrophysiology normally restricted to other animals such as monkeys, rats, and cats. Such work has told us huge amounts about the visual system and other brain circuits, but cats can't talk, so such work can't tell us much about that human ability.

Image of Broca's area and fMRI activation by a speech task, in patient A.

Broca's area and a few others are well-known to be involved in speech production, by way of strokes and other lesions that specifically affect those abilities. But how does it work? Is processing sequential, like the visual system, where a hierarchy of processing takes signals from raw retinal input to various color, edge, motion, and shape detection, up to object recognition? Or is language different, capable of being processed in a more parallel fashion, with all elements (word choice, grammar, phoneme production) coming together at once?

Patient A, with the surface of Broca's area exposed and electrode paths indicated.

This paper supports the former model, finding sub-locations in Broca's area that activate during specific stages of speech production, indicating that, while the full region involved is far smaller than the huge areas devoted to visual scene interpretation, it is also hierarchically arranged, with computations happening sequentially.

Electrophysiology like this is still a pretty quaint and brutal way to look at the brain. Electrodes, resembling dipsticks, are stuck right into the grey matter, with the hope that not too many cells are killed, no blood vessels are blown, and that whatever neurons are near the active surfaces of the electrode give off enough electrical buzz for detection (called a local field potential). Each electrode has several channels (i.e. openings) along its length, so you can listen in to several discrete depths once it is inserted.

Illustration of probe location, with channels indicated in yellow.

The experiment was to cue the patient with a fill-in-the-blank sentence, such as "Yesterday, they ____" plus a generic word to fill in, such as "to walk", or "to think". The patient had the task of computing the right form of the verb (walked, thought) and speaking it silently. (How speaking silently actually works in these experiments is a little hard to understand.) The observation was that reading the cue sentence correlated with a small activation near the Broca area electrodes, while producing the requisite word correlated with much larger activations.

Mix of electrode traces from a few channels. The sentence with blank is presented at the cue time, and the word to be filled in is presented at the target time zero milliseconds (ms). The colored arrows point to the segments discussed below at 200, 320, and 450 ms.

The second observation was that the activations related to speech production were complex, taking place over roughly 600 milliseconds (ms), with distinct peaks and troughs at 200, 320, and 450 ms after presentation of the fill-in word, depending on the channel and electrode location. The point of the experiment was to vary the fill-in words such that more or less complex processing demands could be correlated with more or less complex electrophysiology during these periods of proccessing in Broca's area.

The strategy is much like trying to figure out how a computer works by holding a few electrodes to a computer chip while it is working- a ludicrously difficult and primitive approach to reverse engineering. But it is all we've got for the moment, until functional imaging and non-invasive EEG technologies reach higher resolution. The observations are thus correspondingly crude- that the processing of incipient speech can be broken down into sequential phases (three in this paper) of word identification, inflection processing, and phonology processing. Here is the variation they observe in word identification:

Here, the brown curve around 200 ms is higher for rare words (no examples given) than for common words (but not for short vs long, or multi-syllable vs monosyllable), indicating that this activity is related to word identification.

These curves at 320 ms indicate variation in response to verbal inflection processing- past tenses, irregular forms, etc. "Read" is the control, with no fill-in work. The patient just reads the sentence. "Null-inflect" is when a fill-in is asked for, but the proper form happens to be the same as what is cued, so there is no phonological processing, only implicit grammatical/inflection processing (Every day they ____ [walk]). And "overt" is when the fill-in demands both grammatical processing and changes to the word (suffix or change in form). (Yesterday they ____ [walk]). The experiments were run repeatedly, with flashed cards and randomized orders, with the curves reflecting averages, and curve differences given P-values of 0.01 or less.

Curves at 450 ms (note the different channel (depth) used) correlate with processing for sound construction- related to number of syllables, changes in word form, etc.

The author's conclusions are that they have dissected some aspects of speech production based on where their electrodes penetrated the patient's brains and the timing of observed electrical events, coupled with experimental variation of tasks given to the patients. While most of this was surely known indirectly, based on the many patients with known defects of Broca's and other areas from strokes and other ailments, seeing this activity and its variations in real time is certainly unprecedented, and will contribute to ongoing refinement of the functional mapping of the brain, which, as noted above, has so very far yet to go.

As a stutterer, this is fascinating, and I hope that far more is learned about the nature of speech construction, to the point that the miswiring involved in stuttering might be diagnosed (if not fixed). And of course it also indicates yet again that our behaviors are not magical products of souls, but are computational products of brains.


  1. I am killing a few brain cells with beer currently, but I enjoyed your entry! Fascinating stuff.

    The brain sounds pretty magical to me!


  2. "I'm just another ten percenter. My mind is like an ocean. My mind is like an ocean and I'm hanging in the harbor...."

    - Frank Black from "Ten Percenter"