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.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.