Computational modelling of visual illusions

Illusions reveal some of the sophisticated, underlying neural mechanisms that often remain hidden in our day-to-day visual experience. Illusions have traditionally been studied using psychological methods, which quantify overall, system-level effects observable at the highest layer of the visual hierarchy. This thesis applies the relatively new technique of computational modelling to the study of visual illusions, to quantify bias and uncertainty within various levels of our visual system. The method adopted in this thesis merges statistical inferences, obtained from exposure to image subsets, with filtering operations that mimic visual neural processing from layer to layer. Previous computational models of visual illusions have considered these in isolated arrangement. This dissertation highlights the benefits of combinatorial modelling, which includes separating out the contribution of neural operations from potential statistical influences. The first study in this dissertation investigates a well-known line-length illusion in a benchmark model of the visual ventral stream, demonstrating that a model imitating the structure and function of our cortical visual system is susceptible to illusions. In the second study, we further scrutinise this line-length illusion inside each layer of the benchmark model, observing magnitudes of uncertainty and bias that propagate through each level. In the third and final study, we introduce a new model based on exponential filters inspired by contrast statistics of natural images. We apply a suite of lightness illusions to this new model and demonstrate that low-level kernel operations can account for a large set of these illusions. In summary, this thesis shows that combining filtering functions with natural image statistics not only allows for illusory bias and uncertainty to be limited in artificial neural network models. but it also provides further evidence for and against some proposed theories of visual illusions.