Central Processing of Visual Information

Weill Medical Coll Of Cornell Univ, New York NY

Investigators

Linked publications & trials

Abstract

Understanding the processes by which the brain transforms sensory input into representations that support decisions and actions is a central goal of systems neuroscience. Here, we undertake inter-related streams of psychophysical, computational, and theoretical studies to advance towards this goal. We focus on visual judgments of textural similarity and material properties, two related processes that are not only good model systems for this purpose, but also important for behavior. We characterize the perceptual representations in these domains and the transformations between them, ask whether canonical neural computations can account for these findings, and develop tools to analyze the characteristics of perceptual representations in general. To do this, we build on advances of the previous project period, which developed a predictively accurate model for visual texture discrimination. Human behavior, as captured by this model, was close to normative (sensitivities to local image statistics were matched to their informativeness in natural images) and the perceptual representation had a simple, Euclidean geometry. However, when texture was used for different tasks – suprathreshold similarity and grouping -- the perceptual representation was altered dramatically, implying top-down influences on how sensory information is represented and transformed. We hypothesize that this phenomenon is general, and that task-dependent transformations have geometries that correspond to recognized canonical neural computations. Here we pursue these ideas and their implications. In Aim 1A, we delineate the transformation from threshold representations of visual texture to suprathreshold representations of similarity, and in Aim 1B, how this representation is modulated by task demands (similarity vs. dis-similarity, comparisons in working memory, and grouping). In Aim 2, we examine visual judgments of material properties. Previous work shows, perhaps surprisingly, that these affordance judgments are driven by low-level visual features at specific spatial scales, and our preliminary results indicate that local image statistics play a major role. Aim 2A will characterize the visual estimation of material properties for multiple affordances. Aim 2B will determine the computations that relate suprathreshold judgments of similarity to the representation space of affordances. These aims are pursued in a novel geometric framework, motivated by several lines of evidence that Euclidean models may fall short. We consider a wide range of model classes for the representational spaces and the transformations between them, and apply innovative, rigorous tests based on the data we acquire. As the analytical tools are likely to be widely applicable, Aim 3 expands these developments and makes easily-used tools available to the community. We anticipate that the processes that we delineate at the algorithmic and computational levels in Aims 1 and 2 will reveal generalizable insights into the functional role of canonical neural computations in visual perception, and that the tool development and promulgation in Aim 3 will enable tests of these ideas in many modalities.