Functional magnetic resonance imaging (fMRI) has captured the popular imagination since its introduction in the early 1990s, at least partially because of the stunningly beautiful images it generates. Although it has mostly used to identify brain regions involved in specific cognitive operations, new pattern classification techniques have been applied to fMRI data in what some have called “mind reading technology.” These techniques go beyond simply showing which brain areas are more active than others during a particular task to reveal functional relationships among multiple brain areas, while simultaneously avoiding both the spatial averaging and the low signal-to-noise ratios of traditional MRI methods. Exaggerations and speculations followed this development, including some frighteningly premature attempts at using this in the Indian justice system (and recent indications that parts of the US judicial system may be interested as well).
Although some of this fanfare is well deserved, these new techniques have at least one large shortcoming from a cognitive neuroscience perspective: we cannot know which of the spatial patterns of neural activity are intrinsically related to a certain task, and which are “epiphenomenal.” That is, some of the differentiating characteristics used by these algorithms might be merely correlational, and might fall apart if the methods were applied to a wider group of individuals or to stimuli with different characteristics.
Williams, Dang & Kanwisher addressed this concern by integrating information from correct and incorrect trials in a simple task. Typically, only correct trials are used in fMRI studies, due to the common assumption that there’s only one right way to do a task, and many wrong ways (thus, analyzing incorrect trials might statistically clutter the processes involved in good task performance). But Williams et al. reversed this logic: if pattern classification of fMRI is to focus on the real “meaningful” spatial patterns of neural activity, those extracted patterns should be precisely those which are most related to good task performance.
Williams et al. had 6 subjects categorize abstract shapes as belonging to one of three categories – spikies, smoothies and cubies (see inset image) – while inside an fMRI scanner. The authors isolated two different brain “areas of interest” as sensitive to the differences between these categories, and subtracted the correlation between categories within a single area from that within categories in a single area, and compared this between incorrect and correct trials. This calculation basically indicates whether the activity patterns correlated due to category membership are also related to task performance. In both areas of interest (lateral occipital cortex and retinotopic occipital cortex) activity was more highly correlated within than between categories, but this correlation was stronger for correct classifications only in the lateral occipital cortex.
In other words, while retinotopic occipital cortex appears to represent information that can be used by a computer algorithm to discriminate these categories of items, only “downstream” activity in the lateral areas actually correlates with performance. So although the information needed to discriminate “spikies” from “smoothies” exists in multiple parts of the brain, only some of that appears to be used in overt responding. Indeed, even when retinotopic cortex contains patterns which are reliably related to the correct answer, subjects did not appear to use this information in their judgment!
One might say that these informative representations exist at a sub- or unconscious level, since they are not used in providing a response. Williams et al preferred more subtle language: they claim that “only some spatial patterns of fMRI response are read out in task performance.” Still, the underlying claim is the same: the brain contains this information but is apparently not utilizing it in cognition.