arXiv: Authors present predictive coding networks aligned with human vision

arXiv: Authors present predictive coding networks aligned with human vision

Dirk Gütlin, Ryszard Auksztulewicz and Adrien Doerig introduce Predictive Brain Networks (PBNs), a recurrent framework using local predictive coding (PC)-based learning. They argue that PBNs match or exceed established vision models in representational alignment with the human ventral visual stream, and outperform existing PC-inspired networks on downstream classification tasks. The authors also report that top-down predictions carry brain-aligned information not captured by bottom-up feature extraction, and that these predictions shape lower-level visual responses. They say PBNs support uncertainty estimation. The proposed contribution is a way to study biologically grounded learning while retaining rich neural representations; the abstract does not establish how well these results generalize beyond the reported evaluations. What datasets and evaluation settings underlie the reported alignment and classification results, and how do the authors test generalization?

Publication: b8f34cc9-ca3e-487a-8671-7d323c349d32

attribution — pass: The reading correctly names authors Dirk Gütlin, Ryszard Auksztulewicz and Adrien Doerig. It frames the results as their assertions with phrasing such as 'they argue', 'the authors report' and 'they say'.

Authors:Dirk Gütlin, Ryszard Auksztulewicz, Adrien Doerig

summary — uncertain: Central terms are never explained in plain language: predictive coding, local learning, ventral visual stream, representational alignment and top-down vs bottom-up. A non-specialist cannot tell what changed. The reading also omits the motivating paradox: PC-based models learn without globally propagated error signals but have trailed deep networks. That drops the authors' central claim to close a longstanding gap. Smallest repair: briefly define these terms (or tooltip them) and add one attributed sentence on the paradox the authors say PBNs overcome.

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

date — pass: The displayed originalPublished date of 2026-10-27 matches the arXiv v1 submission date. The later feed date was correctly not used.

[Submitted on 27 Oct 2026]

support — pass: Each claim is directly supported by the abstract. The limitations accurately note that only the abstract is available and that generalization is not established. The question follows from the evidence and introduces no unsupported benefits or costs.

PBNs substantially outperform existing PC-inspired networks on downstream classification tasks

media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.