Abstract
Many research questions in visual perception–particularly those dealing with notions such as "independence," "invariance," and "holism" of visual representations–are special cases of the problem of perceptual separability. In visual neuroscience, there is great interest in determining to what extent important object dimensions are represented separately in the brain. However, progress in the study of separability has been hindered by inadequate methods for its detection. In particular, the definitions of independent or separable representations used in most published research are operational, lacking a theory to guide the interpretation of results. Here we describe a new test of perceptual separability for fMRI data, based on a theoretical definition from multidimensional signal detection theory. The test is essentially an extension to multi-voxel pattern analyses: a linear classifier is used to classify activity patterns related to individual stimulus presentations on the basis of a specific stimulus dimension (e.g., "sad" vs. "neutral" faces). The data points are then projected to a line orthogonal to the classification bound (i.e., the "emotion" dimension) and used to estimate a probabilistic perceptual distribution for each stimulus, as proposed by signal detection theory. These estimated perceptual distributions can be used to directly assess perceptual separability. This test has a strong theoretical basis and can be related to behavioral tests of separability that have been widely applied in the past. We apply the test to the study of separability of human face identity and emotional expression. Twenty-one participants completed a face identification task with faces varying in identity and emotional expression (neutral/sad). fMRI data was analyzed using the new separability test. Violations of separability were present for both emotional expression and identity, they were widespread across areas in the face network, and they were more prevalent in the left hemisphere.
Meeting abstract presented at VSS 2016