Perceptual Distortions in PredNet and Quantification of Top-down / Bottom-up Flow

Perceptual distortions are widely observed in various psychiatric diseases, including Autism Spectrum Disorder (ASD). Recent Bayesian models of psychiatric disorders and learning disabilities propose a general theory grounded on the concept of "aberrant precision." However, these models have yet to be used as phenomenological models for visual distortions because previous models usually only deal with a low-dimensional input with Laplace approximation. Such assumptions are necessary for precision to be well defined; otherwise, the explanation based on aberrant precision was hardly applicable. This study addresses these limitations using the predictive coding-based deep neural network PredNet and a new analysis method inspired by the precision account using the Hilbert-Schmidt Independence Criterion (HSIC). We found that the visual distortion happened when trained with more extended temporal contexts, which was mitigated when we increased the weight of the prediction error of the top layer. From HSIC analysis, we showed that this weight increase enhanced the top-down information flow in prediction, which led to an enhanced ability to capture global features of the visual input, such as rotation and average brightness and hue.