During that SfN session last week, Hans Op de Beeck also asked a question about McGugin et al. 2012. At the time, I actually misunderstood and thought that him and Nancy Kanwisher shared the same concern (see previous post). In fact, it turns out that Hans, unlike Nancy, agrees that the effects reported in Rankin’s paper are expertise effects – but he had another concern, which he was nice enough to clarify in an email.
Hi Isabel,
I do not question the existence of expertise effects in the FFA and in many other areas. Indeed the PNAS paper shows expertise effects.
however, I do question the relevance of these effects for explaining face selectivity as there is up to now no good evidence that the expertise effects are specific to face selective voxels. To the contrary, the PNAS paper shows the expertise effects are the same in face selective and non face selective voxels. IF expertise would explain the difference in stimulus (face) selectivity between these two sets of voxels, then effects of expertise should also be different between these two sets of voxels.
Your response/blog is not addressing this most critical issue, nor do most of your papers as they are too narrowly focussing upon face selective voxels. The PNAS paper addresses the issue and congratulations to rankin and all of the lab for this work. However, in this paper you find evidence AGAINST your expertise hypothesis: expertise effects are the same in face selective and non face selective voxels. Thus expertise cannot explain the difference in face selectivity among those voxels.
Maybe you can add this comment to your blog?
Cheers,
Hans
You just have to love it when you think you have published a result that means A, and someone else believes it means not-A! Science is never boring.
I believe our results actually support our claims, and let me try to explain why.
Below is Figure 3 from McGugin et al. (2012). Here is the sequence of steps that produce this analysis, which is conducted in flattened cortex 2D space: We used a standard-resolution localizer, faces – objects, to localize the peak of the right FFA in each subject. We then vary the threshold on that localizer to define a very small ROI around that peak (25 mm2) and larger. We then look at the high resolution data for faces, animals, cars and planes within these regions. Half of the high-res dataset is used to classify voxels based on their maximal responses: voxels that respond most to faces, voxels that respond most to animals, and voxels that respond most to cars or planes (we call those non-sensitive). Car and plane voxels are lumped together because this step occurs before we look at expertise, and for the subjects as a group regardless of expertise, those non-selective voxels do not clearly replicate that maximal response to objects. Note that if we separate them into car or plane voxels, the results are the same.
Using the other half of the high-res data, we then calculate the response to cars vs. animals in each of these 3 types of voxels – and the bar graph represents the magnitude of the correlation of this response with car expertise. For the purpose of this discussion, I am going to ignore “animal voxels”, and focus on the face and non-selective voxels.
Let’s consider possible results:
– If we had found that car expertise effects were obtained ONLY in voxels maximally response to faces, Hans suggests that this would support the expertise hypothesis. This is in fact what we obtained in the left FFA, as shown in Supp Figure 4.
– if we had found that car expertise effects were only obtained in non-sensitive voxels, this would suggest that the car expertise response is found outside of the most face-selective voxels, that they are just very close to each other, but do not spatially overlap. This is NOT what we found, because voxels that show the strongest reliable response to faces also show a car expertise effect, on in both left and right FFA.
– What we did find in the right FFA is this. First, highly face selective voxels, even in the very middle of the FFA, show a car expertise effect. Non-sensitive voxels also show a car expertise effect. But most importantly, these car expertise effects drop out when you move away from the peak of face selectivity, in a ROI that is 300 mm2-200mm2 around the FFA. In the paper we discuss how the signal in the larger ring region is as good as it is in the center ROI. Moreover, the mean response to cars over all subjects is the same in ROI-25mm2 and in ROI-300-200mm2, both in “face” voxels (p=.85) or in non-sensitive voxels (p=.94). So the only difference between the voxels within 200 mm and those passed 200mm2 appears to be that the response across subjects within 200mm2, and not outside of this ring, is predicted by behavioral car expertise.
So interestingly, the results are a bit different in the two hemispheres. I don’t particularly like to think of these voxels as “face” or “car” voxels, I believe things at the voxel (and single-cell) level may be better described in more probalistic terms. But since this is how many of us have been describing our results, I’ll stick with it here.
We are often encouraged to think of face-selective areas surrounded by non-face selective areas. This model however does not apply here at all. When it comes to predict behavioral performance, what matters is NOT what the maximal response of a voxel is, but whether is it near or far the peak of face selectivity. Note that the voxels within the 200mm2 border that do not respond reliably more to one category (therefore, they respond to faces and objects) respond more like face-voxels then they respond like non-selective voxels outside of that region. I made this summary figure to illustrate that the only model with “two kinds” of voxels (model A) that can explain these data is one where one kind of voxel shows both face selectivity and car expertise effects. Even then, this really fits with the pattern on the left FFA. In this Figure, the colors on the right just refers to “kinds” of cells, and the discussion on the right refers to whether you can use a “two kinds of voxels” or a “three-kinds of voxels” (model B) to explain the results. Note that given our design, “face response” means a stronger response to faces than objects, but “car response” means a correlation between responses to cars and car expertise.
In our view, these results are compatible with a spatial distribution of face and car expertise responses that overlap to a large extent, and likely overlap more as a function of expertise, since it likely takes a lot of expertise with cars to match that which we have with faces and to recruit the same number of cells. If we had a single scale on which to compare degree of expertise for faces and cars, the comparison of the spatial distribution of selectivity for the two categories in people with matched expertise would be interesting.
In sum, are there voxels that show an expertise effect for cars and that are not categorized here as “face voxels”?: sure there are. This of course doesn’t mean that these voxels do not respond to faces. Their selectivity relative to other categories, for one reason or another, was not as strong and reliable. Also note that the location of these voxels remains constrained by face-selectivity: they are found near the peak of face selectivity, defined independently. In a way, faces appear to recruit a subset of an area that responds to expertise more generally. It would be very interesting to test subjects with enough of a range in expertise for faces to actually compare apples to apples here: expertise effect for cars and expertise effect (not just mean selectivity) for faces. Perhaps the overlap is even greater then.
Most importantly, we do not find “face” voxels that do not show a car expertise effect. By an extension to humans of the results that Doris Tsao and her colleagues have published, whereby in the center of a face-selective patch, 97% of the cells are face-selective. If this result holds true in human FFA, then we’d have to conclude that much of the car expertise effects found in the heart of FFA are coming from face cells!