Sheaf cohomology explains when predictive coding networks stall, NeurIPS paper shows
burny_tech · x · 2026-09-15
Jeffrey Seely's NeurIPS 2025 workshop paper reformulates linear predictive coding (PC) networks 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.
Key findings: recurrent feedback loops create internal contradictions producing prediction errors unrelated to supervision; Hodge decomposition determines when these contradictions stall learning; and the formalism doubles as a diagnostic tool and a design principle for weight initialization. In the thread, darthur praises the λi integration of the sheaf coboundary residual as a neat route to PI-controller behavior and asks about accelerating gluing dynamics with a local-rate derivative term.
More from Research
- ALPS 2027 NLP Winter School Opens Applications, March 28–April 2 in the Alps — iatitov · 2026-09-15
- Kevin Murphy unifies goal-based hierarchical RL with ACGVF and his 25-year-old HHMM work — sirbayes · 2026-09-15
- New paper: LLM projects to enhance democracy threaten democratic autonomy — sethlazar · 2026-09-15
- Nature Medicine paper: on-premise medical AI agents for reliable clinical decisions — EricTopol · 2026-09-15
- Mathematician predicts two parallel pillars of math: human understanding vs incomprehensible Lean proofs — thegautamkamath · 2026-09-15
- Consciousness giant Antonio Damasio publishes new book Natural Intelligence today — AnnaCiaunica · 2026-09-15