Prediction of fMRI activation patterns

| Biology
Predicting Human Brain Activity Associated with the Meanings of Nouns
Tom M. Mitchell, Svetlana V. Shinkareva, Andrew Carlson, Kai-Min Chang, Vicente L. Malave, Robert A. Mason, Marcel Adam Just

The question of how the human brain represents conceptual knowledge has been debated in many scientific fields. Brain imaging studies have shown that different spatial patterns of neural activation are associated with thinking about different semantic categories of pictures and words (for example, tools, buildings, and animals). We present a computational model that predicts the functional magnetic resonance imaging (fMRI) neural activation associated with words for which fMRI data are not yet available. This model is trained with a combination of data from a trillion-word text corpus and observed fMRI data associated with viewing several dozen concrete nouns. Once trained, the model predicts fMRI activation for thousands of other concrete nouns in the text corpus, with highly significant accuracies over the 60 nouns for which we currently have fMRI data.

This paper is pretty interesting. Basically, they have measured fMRI activation data for 60 words. So, voxel v has activation level A_v_w for word w. For each word, they have measurements of a bunch of linguistic parameters P_1, P_2, P_3..., etc. They then fit a predictive model for the effect of each parameter on the activation level of each voxel, so for instance you could say that if a word is associated with "sight" (i.e., it appears near "sight" in text corpii) that increases the activation of voxel v by .1 units. This is fairly straightforward regression modelling stuff.

Once you have the model fitted, you can then predict the activation of each voxel for a novel word by taking its linguistic parameter values and plugging them into the model. Their results are actually pretty good. They have a corpus of 60 word/fMRI pairs and they use 58 as a training set and 2 as a test set. They then try to differentiate the two test words by asking which predicted activation pattern is closer. The results are significantly better than chance: mean=.77 for what appears to be arbitrary words and mean=.62 when the words are from the same semantic category (e.g., "celery" and "corn"). Moreover, a significant amount of the error appears to come from head motion by the subjects.

I'm not sure how to interpret this from a scientific perspective. It's a long way from knowing which brain cells are used in processing certain words to knowing how the brain actually processes those words. On the other hand, it's not clear you ned that deep an understanding to build a brain-scanning fMRI gizmo that does something useful. Though we're a long way from that too. Even ignoring the fact that we don't understand the brain well enough, hanging out with your head in a noisy magnet probably isn't a lot of fun.