Explaining Sensitivity to Visual Illusions by Modulating the Prior in Predictive Coding Neural Networks

Abstract

Predictive Coding describes perception as an inference process in which sensory information is combined with ‘prior’ expectations to attain the final percept. For instance, some visual illusions have been explained in the light of these predictive processes. Interestingly, some authors proposed that schizophrenia patients have a different encoding of their prior expectations compared to healthy participants and have been shown to be less sensitive to perceiving some visual illusions. The aim of this work is to test this hypothesis from a computational perspective. Previous works demonstrated that artificial neural networks implementing Predictive Coding (PC) principles can perceive some specific visual illusions as humans do.We will compare whether the visual illusions modulated by the predictive processes are identical to those that elude schizophrenia patients. This comparison will allow us to test the compelling hypothesis that schizophrenia patients have indeed different encoding of their ‘prior’ compared to healthy participants and are more robust to perceiving some illusions. Moreover, this work will have an impact also on the field of Machine Learning and more generally brain nspired AI, as it will provide a theoretical and practical ground to improve the robustness and accuracy of current neural networks implementations, in line with previous work.