Design and evaluation of brain-inspired predictive coding networks based on the free-energy principle for novel neuromorphic hardware

Abstract

Predictive coding (PC) based on the free-energy principle (FEP) has been extensively explored for the next-generations artificial intelligence. In this study, we constructed PC network models that can implement perception and unsupervised learning by minimizing variational free energy through neural dynamics. Furthermore, these models were applied to practical tasks, such as image discrimination and real-time prediction. For implementation as neuromorphic hardware with biological plausibility, the performance of PC networks was evaluated by applying techniques used in recent neuromorphic engineering, such as augmented direct feedback alignment and physical reservoir systems. These results pave the way for neuromorphic hardware capable of autonomous perception and learning based on the FEP.