Abstract
Categorization is crucial to how we make sense of the visual world. Using fMRI coupled with multivariate pattern analysis (MVPA) researchers have investigated the structure of categorical representations in the visual system. An open question is the role of abstraction in the structure of these representations. One possibility is complete abstraction across category members, or the representation of a prototype; another is the representation of each exemplar. Between these extremes is a continuum of intermediary models varying in level of abstraction (Vanpaemel and Storms, 2008). We investigated whether these intermediary models might reflect a variable role of abstraction in the neural representation of novel visual categories. Stimuli consisted of 16 annular square-wave gratings (1.5-8° eccentricity) varying in orientation (45, 75, 105, 135°) and spatial frequency (.25, .5, 1, 2 cyc/deg). Representational similarity analysis (RSA) was used to construct neural dissimilarity matrices (DM) for each participant (N = 10) from V1 functional data. After testing for across subject reliability of neural DMs, a group-averaged DM was found to be highly correlated with a model DM of the stimulus space (Spearman's ρ = .92, p < .001). Multidimensional scaling was then used to construct a 2D neural space. Separate groups of subjects performed distinct categorization tasks, and were trained to label (linearly and non-linearly separable) subsets of the stimuli, and generalize the learned categories to the remaining stimuli. The neural space served as input to different formal models of categorization in order to determine which best captured the choice behavior of observers in the categorization tasks. Across categorization tasks we found a variable role for abstraction, which speaks against a dichotomous approach to abstraction in the neural representation of visual categories.
Meeting abstract presented at VSS 2017