{"id":"arxiv:2511.11092","kind":"preprint","title":"Sheaf Cohomology of Linear Predictive Coding Networks","authors":["Jeffrey Seely"],"author_count":1,"year":2025,"published":"2025-11-14","venue":null,"cited_by":0,"doi":null,"arxiv_id":"2511.11092","nct_id":null,"source_url":"https://arxiv.org/abs/2511.11092","archive_url":"https://space-kitty.com/papers/arxiv:2511.11092","updated":"2025-11-14","pmid":null,"pmcid":null,"openalex_id":"W4416343319","pdf_url":"https://arxiv.org/pdf/2511.11092","abstract":"Predictive coding (PC) replaces global backpropagation with local optimization over weights and activations. We show that linear PC networks admit a natural formulation as cellular sheaves: the sheaf coboundary maps activations to edge-wise prediction errors, and PC inference is diffusion under the sheaf Laplacian. Sheaf cohomology then characterizes irreducible error patterns that inference cannot remove. We analyze recurrent topologies where feedback loops create internal contradictions, introducing prediction errors unrelated to supervision. Using a Hodge decomposition, we determine when these contradictions cause learning to stall. The sheaf formalism provides both diagnostic tools for identifying problematic network configurations and design principles for effective weight initialization for recurrent PC networks.","abstract_license":"CC0-1.0 (arXiv descriptive metadata)","license":null,"is_open_access":false,"retracted":false,"peer_reviewed":false,"categories":["cs.LG"],"details":null,"provenance":"arxiv","attribution":"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.","topics":[{"slug":"neuroai","name":"NeuroAI & Computational Neuroscience","domain_slug":"neuroscience"}],"cited_in_dossiers":[]}