tag:blogger.com,1999:blog-5737874959005852552.post6148305216688650646..comments2023-10-18T16:31:41.080-05:00Comments on MVPA Meanderings: A toolbox for representational similarity analysis (and some RSA musings)Jo Etzelhttp://www.blogger.com/profile/04277620767760987432noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-5737874959005852552.post-9314194041885136162014-07-18T11:00:51.626-05:002014-07-18T11:00:51.626-05:00Thanks for the comment! I totally agree that termi...Thanks for the comment! I totally agree that terminology is part of the confusion, and that different methods (averaging, etc) are proper for addressing different questions.<br /><br />I often think of "classes" as the "discrete units we want to make conclusions about". In classification terms, classes are usually the targets - what the algorithm is learning to distinguish ("w and f can be significantly classified in ROI A").<br /><br />If classes are the "discrete units we want to make conclusions about", then I wouldn't call your example of a 12-faces RSA a single-class analysis, but rather 12 (the unique faces). <br /><br />Obviously, the 12 faces all belong to a group - faces! But the12x12 RDM is intrinsically also about the individual stimuli (as you phrased it, "the discriminability between each pair of individual faces") - using 12 different faces would result in different RDMs. Hopefully, the relationship between the RDMs for stimulus super-classes would be the same after changing the stimulus images (eg the ROI has information about faces but not artificial objects).<br /><br />This is a bit like the mass-univariate fMRI "first-level" and "second-level" analysis concept - some of the details that matter in the first level hopefully don't in the second. But we certainly need better terminology (and I'm not recommending "super-class"); something that makes the "level" of analysis clearer.Jo Etzelhttps://www.blogger.com/profile/04277620767760987432noreply@blogger.comtag:blogger.com,1999:blog-5737874959005852552.post-83174688898686113052014-07-18T08:20:40.883-05:002014-07-18T08:20:40.883-05:00I have to say this is a very nice post about RSA a...I have to say this is a very nice post about RSA and the toolbox. Thanks! <br />Also thanks to the authors who put a lot of efforts in making it available. <br />There is just a few aspects here which I think deserve a bit of elaboration. I have struggled a lot with some of the concepts that you describe, and I just want to add my own two cents. <br />What you say about the classes 'w' and 'f' is true, however I think it would be good to agree on what is a ‘class’. If a ‘class’ is something like, say faces, then you can investigate this class with different exemplars of faces, with trial repetitions for each exemplar. In this example case of a single ‘class’ RSA investigation, the trial repetitions of each particular individual face will enable estimating the ld-t. If you had say 12 different faces to make this class, then you would end up with a 12 x 12 RDM, and the elements of the RDM would tell you what is the discriminability between each pair of individual faces in the region you are currently looking at. This would be an example of a single-‘class’ RSA study. This particular example with faces has been particularly challenging by the way, but that is a different story. <br />One of the powerful aspects about RSA in general, is that it enables you to investigate ‘classes’ of stimuli, without assuming said ‘classes’ a priori. This is the case for a condition-rich design like the one described in Kriegeskorte et al. 2008. We could say that this study is composed of roughly speaking 6 ‘classes’ with stimuli drawn from human bodies, human faces, animal bodies, animal faces, natural inanimate objects and artificial inanimate objects. But what the report from Kriegeskorte et al. showed is that human inferior temporal cortex (and monkey IT!) emphasised particular class distinctions in the context of this condition rich design. Just by eye-balling the RDM, you can see that there is information in hIT about these ‘classes’, where exemplars within the ‘classes’ lead to more similar activity patterns, and exemplars that belong to different classes, like say animate and inanimate objects, lead to dissimilar activity patterns. This is different from many studies in which the experimenters assumed such ‘classes’, and averaged the activity patterns across different exemplars of the ‘classes’ before comparing them. This comment is not a rant per se about averaging, I would say that the two methods enable answering different questions. <br />But in the case of RSA, if there is information related to a particular ‘class’, then the RDM should show it and, with the RSA-toolbox that you very well describe in your post, hypothesis about ‘classes’ can be tested with models and statistical inference.<br />Thank you again for the wonderful blog post!<br />Ian Charest<br />Ian Charesthttp://www.mrc-cbu.cam.ac.uk/people/ian.charest/noreply@blogger.com