Conference on Cognitive Computational Neuroscience

Novelty detection (ND) is a crucial capacity that underpins adaptive decision-making and flexible cognition. This capacity to detect patterns as novel or familiar depending on the subject's past experience is empirically matched by numerous experimental verification of the existence of neurons that have their activity modulated by novelty across many different brain areas. Despite its functional significance and the invariable existence of putative neural substrates, the computational mechanisms that underlie ND have received limited attention from neuroscience in recent years. We propose an energy-based approach to ND through predictive coding, which is more biologically plausible and inherently connects ND and memory. In this work, we (a) propose a robust, integrated energy-based ND model through Predictive Coding Network (PCN), (b) compare its performance to ND models using Hopfield Network (HN) or Modern Continuous Hopfield Network (MCHN) and provide mathematical analysis for the differences in performance, and overall (c) propose a candidate computational mechanism that underlies the existence of novelty neurons and the brain's flexible capacity to perform ND across many areas.