A psychophysically-tuned computational model of human primary visual cortex produces geometric optical illusions

Abstract

Geometric optical illusions (GOIs) are mismatches between physical stimuli and perception. GOIs provide an access point to study the interplay between sensation and perception, Yet, there is relatively scant quantitative investigation of the extent to which different GOIs rely on similar or distinct perceptual mechanisms, which themselves are driven by specific physical properties. We addressed this knowledge gap with a combination of psychophysics and computational modelling. First, 30 healthy adults reported quantitatively their perceptual biases with three GOIs, whose physical properties parametrically varied on a trial-by-trial basis. A given physical property, when considered in isolation, had different effects on perceptual biases depending on the GOI (e.g. the spacing of stimuli affected one GOI, but not another). For a given GOI, there were oftentimes interactions between the effects of different physical properties. Next, we used these psychophysical results to tune a computational model of primary visual cortex that combines parameters of orientation selectivity, receptive-field size, intra-cortical connectivity, and long-range interactions. We showed that similar biases generated in-silico mirror those observed in human behavior when receptive field size, bandwidth and shape (rounded or elongated) are tuned, as well as parameters encoding the strength of the long-range intra-regional interactions between receptive fields. Collectively, our results suggest that different physical properties are not operating independently, but rather synergistically, to generate a GOI. Such results provide a roadmap whereby computational modelling, informed by human psychophysics, can reveal likely mechanistic underpinnings of perception.