Search Papers | Poster Sessions | All Posters
Poster A132 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Context-dependent efficient sensory coding underlying the tilt-illusion in human visual cortex and artificial neural networks
Ling-Qi Zhang1 (), Geoffrey K. Aguirre2, Alan A. Stocker2; 1Janelia Research Campus, Howard Hughes Medical Institute, 2University of Pennsylvania
Human perception of basic visual attributes is often influenced by spatial context. A canonical example is the tilt illusion, in which the perceived orientation of a stimulus is altered by the presence of a spatially oriented surround. We hypothesize that surround effects originate from changes in neural representation that maximize coding efficiency based on spatial context. We simultaneously collected perceptual and fMRI data from human participants in a tilt-illusion experiment. We found that orientation encoding in the absence of a surround reflects natural scene statistics both in behavior and in the neural responses of the visual cortex. In the presence of an orientated surround, encoding accuracy was significantly increased at the surround orientation. The pattern of change in coding accuracy is consistent with the surround-conditioned orientation statistics of spatially adjacent regions in natural images. Furthermore, we found the same orientation encoding characteristics and contextual modulation in convolutional neural networks trained on natural images. Our results suggest that efficient coding based on spatial context is a general mechanism in visual processing of natural images.
Keywords: efficient coding tilt illusion Fisher information convolutional neural network