Local learning is all you need: brain-aligned predictive coding networks

Abstract

Predictive Coding (PC) is an influential account of how brain representations emerge through hierarchical, local, self-supervised prediction-based learning, without requiring globally propagated error signals. Despite extensive evidence supporting PC, a striking paradox persists: computational models built on these biologically grounded principles have consistently failed to match state-of-the-art Deep Neural Networks in performance and brain-alignment. Here, we show that this paradox can be overcome. We introduce Predictive Brain Networks (PBNs), a hierarchical recurrent framework that implements local PC-based learning and achieves representational alignment with the human ventral visual stream matching or exceeding established vision models. Importantly, PBN top-down predictions encode unique, brain-aligned information not captured during bottom-up feature extraction, with this contribution increasing in locally trained variants. Further, PBNs substantially outperform existing PC-inspired networks on downstream classification tasks, reproduce key signatures of PC (with higher-level predictions shaping lower-level visual responses as observed in the primate cortex), and support uncertainty estimation. PBNs therefore close a longstanding gap in computational neuroscience by combining two properties that have been largely separated: the brain-inspired learning mechanisms of predictive coding and the representational power of modern deep networks. They provide a scalable framework for studying biologically grounded learning without sacrificing the capacity to produce rich neural representations.