Abstract
The ventral stream is known to host low-level representations of visual objects in early visual areas and high-level representations emphasizing categorical divisions in more anterior regions. Here we compare similarity spaces among key functional regions to elucidate representational transformations across stages of processing. We performed high-resolution fMRI in 24 subjects to measure response patterns elicited by 62 images of isolated real-world objects (24 animate, 24 inanimate, 14 well-controlled faces) in V1, V2, V3, LOC, OFA, FFA, and PPA. For each pair of stimuli, we computed an unbiased estimate of their representational discriminability (linear discriminant t values) and assembled these estimates in a representational dissimilarity matrix (RDM) for each brain region. To quantify the relatedness and distinctness of regional representations, we computed a novel measure of noise-corrected shared and non-shared RDM variance. Early visual regions did not exhibit strong categorical structure, while representations became highly categorical in higher-level regions. Categoricality did not arise gradually along the ventral stream, but appeared suddenly at the stage of LOC and OFA, where faces, bodies, and inanimate objects formed distinct clusters. The same clusters were present in FFA, where the representation appeared to be of low dimensionality, and broke down in PPA, where scenes were distinct and all other images formed a single cluster. V1-3 shared most of their dissimilarity variance, as did LOC, OFA, and FFA, but V1-3 were significantly distinct from the category-selective regions. PPA was significantly distinct from V1-3 as well as LOC, OFA, and FFA. The ventral stream hosts object representations that are both related to and distinct from one another to different degrees. Our novel methods for inference on shared and non-shared dissimilarity variance contribute an important tool for understanding brain computations as transformations of representational spaces.
Meeting abstract presented at VSS 2016