Applied Sheaf Theory For Multi-agent Artificial Intelligence (Reinforcement Learning) Systems: A Prospectus

Eric Schmid

REVIEW 4 major objections 5 minor 26 references

Applied Sheaf Theory For Multi-agent Artificial Intelligence (Reinforcement Learning) Systems: A Prospectus

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper argues that sheaf theory's local-to-global perspective may provide the right mathematical tools for reasoning about how local agent behaviors collectively determine emergent properties in multi-agent AI and reinforcement…

desk verdict Mostly correct tutorial plus literature review dressed as a prospectus; no new results, and Chapter 3's conclusion overclaims what was demonstrated. read the letter →

arxiv 2504.17700 v1 pith:V6AMJA33 submitted 2025-04-24 math.OC math.AT

Eric Schmid This is my paper · ORCID

classification math.OCmath.AT MSC 18F2055N3068T4290C2568T05

keywords sheaftheorycohomologycellularsheavesmulti-agentsystemsreinforcementlearningLaplaciandistributedoptimizationADMM

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper is a research prospectus, not a finished theory. It argues that sheaf theory—the branch of mathematics that tracks locally defined data and asks when it can be glued into a consistent global object—offers a natural language for multi-agent AI and reinforcement learning, where individual agents act on local information and system behavior emerges globally. The paper teaches the basics of sheaves and sheaf cohomology, then lays out a programme to encode agents, their information flows, and their coordination constraints as cells of a sheaf, so that consensus, formation control, flocking, and learning become instances of one mathematical structure. The payoff it promises is that cohomology would measure when local behaviors cannot be assembled into any globally consistent outcome, and distributed algorithms such as ADMM would solve the resulting optimization problems over the network. No such model is built here; the contribution is the proposal and the roadmap.

What carries the argument

The central object is a cellular sheaf on a graph: to every node attach a vector space (an agent's possible states), to every edge attach a vector space (the interaction's possible data), and to every node-edge incidence attach a linear map telling how node data must appear at the edge. The coboundary operator $\delta$ measures local inconsistency; its kernel is the space of global sections (consistent configurations), and its cokernel in this graph setting gives $H^1(G;\mathcal{F})$, the space of gluing obstructions. On top of this sit the sheaf Laplacian $L_{\mathcal{F}} = \delta^{\mathsf{T}}\delta$, whose gradient flow generalizes graph-Laplacian consensus, and a nonlinear version arising from edge potentials that turns coordination tasks into convex programs solved by ADMM. Unrolling the ADMM iterations yields a sheaf neural network. This machinery carries the argument because it converts a vague local-to-global intuition into equations whose solutions and obstructions are well-defined.

What would settle it

Take a small team of agents on a three-agent cycle with pairwise constraints that are deliberately inconsistent, such as relative positions that cannot be realized simultaneously. The framework predicts that $H^1(G;\mathcal{F}) \neq 0$, that no global section exists, and that the ADMM/nonlinear-sheaf solver will fail to reach exact consistency precisely in that case, converging instead to a least-squares compromise. Run the same scenario with consistent constraints: $H^1$ should vanish and the solver should reach a global section. If the nonzero cohomology class does not track which constraints are infeasible, or if the solver behaves identically in both cases, the representational claim is empty.

Watch

Extended reading notes

Core claim

The paper's central claim is that sheaf theory's local-to-global machinery can be transferred to multi-agent systems: encode each agent's state space as a stalk of a cellular sheaf on the communication graph, encode each interaction as an edge stalk with linear restriction maps, and let a global section of the sheaf be exactly a configuration in which all local agents are mutually consistent. In this translation, the coboundary operator $\delta$ measures disagreement, the sheaf Laplacian generates a distributed flow that drives the system toward consistency, and the first cohomology group $H^1(G;\mathcal{F})$ detects constraint loops that cannot be satisfied. The paper presents this translation as an 'isomorphism' to be established between specific applications and sheaf cohomology, and it reviews existing building blocks—nonlinear homological programs, ADMM, and sheaf neural networks—as evidence that the pieces are in place. Stated sympathetically, the paper is trying to establish that coordination and emergence in multi-agent systems are cohomological phenomena, and that the vocabulary of sheaves will make those phenomena visible and computable.

Load-bearing premise

The whole proposal depends on the assumption that real multi-agent, economic, and reinforcement-learning systems can be represented faithfully and non-vacuously as sheaves—that is, that the gluing axioms and cohomology classes actually correspond to how agents exchange information and to genuine coordination failures, rather than to an arbitrary labeling. Nothing in the paper demonstrates this representation for a concrete system.

Editorial extensions

If this is right

If the representation works, consensus, formation control, flocking, and distributed sensor fusion become special cases of one homological program, solvable by a common distributed algorithm.
A nonzero $H^1(G;\mathcal{F})$ on a constraint loop gives a rigorous certificate that local demands cannot be simultaneously met, so systems can detect infeasible coordination before acting.
The sheaf Laplacian flow generalizes graph-Laplacian consensus: in the linear case it converges to a global section, and the spectral gap of $L_{\mathcal{F}}$ controls how fast local inconsistencies dissipate.
Unrolling the ADMM solver produces a sheaf-neural-network architecture, so coordination protocols can be tuned by gradient-based learning while constraints remain baked into the structure.
The same scaffold extends to hypergraphs and higher-dimensional cell complexes, allowing constraints that involve more than two agents at once.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the proposed isomorphism is ever made precise, a natural extension is to use $H^1$ as a regularizer or curriculum signal in multi-agent RL: reward local policies only when their induced edge discrepancies vanish, and penalize configurations whose cohomology class is nonzero.
The paper's framework suggests a design criterion before training: any coordination task that admits no global section in the chosen sheaf is provably unlearnable as a hard constraint, so one should either relax constraints or change the sheaf.
One testable near-term outcome, not delivered here, would be a mechanically verified proof that the ADMM updates preserve the sheaf constraints; the paper's plan for a dependently typed formalization points toward exactly that check.
Time-varying sheaves could model changing communication topologies, and the evolution of $H^1$ over time might serve as a real-time measure of coordination capacity.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a prospectus, not a research paper. It contains a standard, mostly correct tutorial on sheaf theory, a literature review of recent cellular-sheaf coordination work (mainly Hanks et al. and Riess), and a proposed research program for applying sheaf cohomology to multi-agent RL and economics. No new theorem, algorithm, or experiment appears. The abstract is upfront about this.

What it does well: the exposition in Chapter 1 is clear and accessible, with good examples. The review in Chapter 3 is a useful summary of the Hanks et al. framework and its connection to sheaf neural networks. For someone wanting a quick entry into that literature, this is a reasonable starting point.

The soft spots are real but proportional. The central claim—that multi-agent systems can be faithfully represented as sheaves with cohomology detecting emergent failures—is announced as future work and never instantiated. There is no worked example where a nonzero H^1 corresponds to a coordination failure in an RL or economic setting. The lone cohomology remark, §3.6, says H^1 reveals unsolvable constraint loops, but on a graph H^1 is just C^1/im δ, measuring edge-assignment inconsistency, not an emergent property. Chapter 3's conclusion overstates: 'we demonstrated how classical coordination tasks can be formulated' when the chapter is a review of other people's work. There are also sloppy '[?]' citation placeholders near §3.2.3 and §3.2.4, and the fine-sheaf remark in §1.6.4 omits paracompactness qualifiers. None of these are fatal for a prospectus, but they need fixing.

Bottom line: this is a decent tutorial and research proposal, not a research contribution. If the venue publishes tutorials/position papers, it could go to review after heavy revision. For a standard math or ML research journal, I'd desk reject. I would not cite it in my own work; I'd cite Hanks et al. and Hansen/Gebhart directly.

Referee Report

4 major / 5 minor

Summary. The paper is a three-part manuscript: a pedagogical introduction to sheaf theory and sheaf cohomology (Chapter 1), a PhD-style research prospectus for applying sheaves to multi-agent AI and reinforcement learning systems (Chapter 2), and a literature review of cellular sheaves for multi-agent coordination, sheaf Laplacians, homological programs, ADMM, and sheaf neural networks (Chapter 3). The abstract explicitly states that no completed model is presented; the intended contribution is described as theoretical groundwork and a roadmap for future research.

Significance. If the proposed framework were realized, it could connect algebraic topology with multi-agent RL and economics, and sheaf cohomology might diagnose coordination failures that graph-based methods miss. The tutorial portion is mostly standard and readable, and the literature review usefully collects recent work by Hanks et al., Riess, and the sheaf neural network community. The paper makes no testable predictions and contains no original derivation or experiment, however, and the central 'isomorphism' between applications and sheaf cohomology is deferred rather than demonstrated. The strengths are the expository clarity of Chapter 1 and the coherent organization of Chapter 3 as a survey; the weakness is the absence of an established central claim.

major comments (4)

[Section 3.2.4, Eqs. (3.2)-(3.3)] The constraint set in Eq. (3.2), namely F_{i->e}(x_i) = y_e for every endpoint of every edge, forces F_{i->e}(x_i) = F_{j->e}(x_j) for each edge, i.e., delta_F x = 0. Under that constraint the objective cannot reduce to sum_e U_e(F_{i->e}(x_i) - F_{j->e}(x_j)) as claimed in Eq. (3.3). The soft-constrained program requires the constraint delta_F x = y (one equality per edge), which is precisely the form used later in the ADMM derivation in Section 3.8.2. Without this correction, the definition of the nonlinear homological program is internally inconsistent.
[Section 3.6, third bullet] The assertion that 'if agents have cyclic dependencies that cannot all be satisfied ... H^1(G;F) != 0 will reveal that' is not a theorem of the preceding formalism. For the constant sheaf on a cycle graph, H^1(G;F) is nonzero while the consensus constraints delta x = 0 have solutions; in general H^1 measures edge-cochain obstructions modulo exact cochains, not the solvability of delta x = 0, which is governed by H^0 and by whether the target edge data lie in the image of delta_F. The manuscript should either state and prove a precise cohomological criterion or remove this claim.
[Section 3.11, first paragraph] The statement 'We demonstrated how classical coordination tasks like consensus and formation control can be formulated in this framework' overstates what the paper does. The formulations in Examples 3.2.7-3.2.8 and the algorithm in Section 3.8 are restatements of the framework of Hanks et al.; no original demonstration is supplied. The conclusion should be reframed as a review of existing results plus a proposal, or the manuscript must add an original worked example that is not borrowed from the cited literature.
[Section 2.4 and Section 2.3.3] The paper announces as its core scientific contribution an 'isomorphism' between specific applications and concepts in sheaf cohomology, but no such mapping is defined, instantiated, or tested anywhere in the manuscript. Because the abstract disclaims a completed model this is not a fatal inconsistency, but it means the central thesis is currently unsupported; the paper should explicitly separate what is established from what is proposed future work, both in Chapter 2 and in the conclusion of Chapter 3.

minor comments (5)

[Sections 3.2.3 and 3.2.4] Unresolved placeholder citations '[?]' appear for LaSalle's invariance principle and for the origin of homological programs; these must be completed before publication, as they prevent verification of cited convergence claims.
[Example 1.6.1] The circle cohomology computation is hard to follow; the role of the two overlap components and the sign convention for the 1-cocycle should be spelled out step by step, although the final result H^1(S^1;R) is correct.
[Section 1.6.4] The conditions under which Cech cohomology agrees with derived-functor cohomology should be stated once in a precise form (for example, paracompact Hausdorff spaces) rather than scattered across several paragraphs with varying qualifiers.
[Bibliography] The bibliography lists the Sheaf Neural Networks paper twice, as [6] and [18]; please consolidate the duplicate entry.
[Chapter 2] Chapter 2 is written as a first-person dissertation proposal; if the manuscript is intended for a journal, the prospectus sections should be rewritten in the standard research-paper voice.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper is an explicitly non-completed prospectus whose proposed sheaf-theoretic 'isomorphism' is future work and whose worked coordination examples reduce to cited external results, not to the paper's own inputs.

full rationale

The manuscript does not present a derivation chain from assumptions to a prediction. Its abstract states: 'This paper does not present a completed model but rather lays theoretical groundwork and identifies promising research directions.' The central proposed contribution, 'The core scientific contribution will be establishing an "isomorphism" between specific applications and concepts in sheaf cohomology' (§2.4), is explicitly future work, so there is no fitted input called a prediction and no self-definitional reduction. The technical content in Chapter 3 is presented as a literature review with external attribution: 'Recent work by Hanks et al. introduced a general framework for heterogeneous multi-agent coordination using cellular sheaves and a new concept called the nonlinear sheaf Laplacian' (§3.1), and the ADMM and convergence claims are cited to Boyd et al. and Hanks et al. The paper's own consensus and formation examples (Examples 3.2.7 and 3.2.8) are explicit, self-contained reductions to homological programs, but they are illustrations of a known framework, not predictions derived from its own assumptions. No self-citations are load-bearing, and no uniqueness theorem is imported from the author's prior work. The manuscript does contain missing citation placeholders ('[?]' at §3.2.3, §3.2.4, and §3.2.5) and some overstatement in §3.11 ('We demonstrated...' relative to the review nature of Chapter 3), but these are completeness and framing issues, not circularity. The proposed sheaf-'isomorphism' for multi-agent and RL systems is asserted as a research goal, not obtained by construction from its own definitions.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper contains no fitted parameters and invents no new mathematical entities. Its expository portions rely on standard mathematics; its review and prospectus depend on the external results of Hanks et al., Riess, and the sheaf neural network literature, plus the untested assumption that multi-agent systems admit faithful sheaf representations.

assumptions (5)

standard math Standard sheaf-theoretic definitions and theorems (presheaves, sheaves, stalks, sheafification, Cech and derived-functor cohomology) are assumed as background.
Chapter 1 is a pedagogical exposition; the paper relies on this background without proving it.
domain assumption Cech cohomology equals derived functor sheaf cohomology for the spaces considered.
§1.6.4 states 'We will assume henceforth that we are working in a context where this is true (e.g. X is paracompact)'; this restricts the applicability of the exposition.
domain assumption Hanks et al.'s convergence results hold: the nonlinear sheaf Laplacian dynamics converge to a global section when desired edge values lie in the image of the coboundary, and the ADMM-based solver converges to the optimum under convexity.
§3.2.7, footnote 8, and §3.8.5 invoke theorems from Hanks et al. arXiv:2504.02049 without proof; the literature review and prospectus depend on these results.
domain assumption The optimization problems are convex and the edge potentials are strongly convex where convergence guarantees are stated.
§3.8.1 lists convexity, strong convexity of potentials, and feasibility/Slater's condition as assumptions for the distributed algorithm.
domain assumption A faithful representation of multi-agent RL and economic systems as cellular sheaves exists and preserves the phenomena of interest (the 'isomorphism' hypothesis).
§2.3.3 and §2.4 propose establishing this mapping as the core contribution; it is an untested research hypothesis, load-bearing for the prospectus.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Applied Sheaf Theory For Multi-agent Artificial Intelligence (Reinforcement Learning) Systems: A Prospectus." pith.science (2026). https://pith.science/paper/V6AMJA33

@misc{pith2026250417700,
  author       = {Pith},
  title        = {Pith review of: Applied Sheaf Theory For Multi-agent Artificial Intelligence (Reinforcement Learning) Systems: A Prospectus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6AMJA33}},
  note         = {Machine review of arXiv:2504.17700}
}

read the original abstract

This paper provides a pedagogical introduction to classical sheaf theory and sheaf cohomology, followed by a research prospectus exploring potential applications to multi-agent artificial intelligence systems. The first section offers a comprehensive overview of fundamental sheaf-theoretic concepts-presheaves, sheaves, stalks, and cohomology-aimed at researchers in computer science and AI who may not have extensive background in algebraic topology. The second section presents a detailed research prospectus that outlines a roadmap for developing sheaf-theoretic approaches to model and analyze complex systems of interacting agents. We propose that sheaf theory's inherent local-to-global perspective may provide valuable mathematical tools for reasoning about how local agent behaviors collectively determine emergent system properties. The third section contains a literature review connecting sheaf theory with existing research in multi-agent systems, reinforcement learning, and economic modeling. This paper does not present a completed model but rather lays theoretical groundwork and identifies promising research directions that could bridge abstract mathematics with practical AI applications, potentially revealing new approaches to coordination and emergence in multi-agent systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages

[1]

Distributed Multi-Agent Coordination over Cellular Sheaves

Tyler Hanks, Hans Riess, Samuel Cohen, Trevor Gross, Mat thew Hale, and James Fairbanks. Distributed Multi-Agent Coordination over Cellular Sheaves . arXiv:2504.02049, 2025 (under review)

arXiv 2025
[2]

Lattice Theory in Multi-Agent Systems

Hans Riess. Lattice Theory in Multi-Agent Systems . Ph.D. thesis, University of Pennsylva- nia, 2023. Also available as arXiv:2304.02568

work page Pith review arXiv 2023
[3]

A Gentle Introduction to Sheaves on Graphs

Jakob Hansen. A Gentle Introduction to Sheaves on Graphs . Preprint, 2020. Available at http://www.jakobhansen.org/publications/gentleintroduction.pdf

work page 2020
[4]

Justin M. Curry. Sheaves, Cosheaves and Applications . Ph.D. thesis, University of Pennsylva- nia, 2014. Available as arXiv:1303.3255

arXiv 2014
[5]

Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers

Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and J onathan Eckstein. “Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers.” Foundations and Trends in Machine Learning 3(1):1–122, 2011

work page 2011
[7]

Sheaf Neural Networks with Connection Laplacians

Federico Barbero, Cristian Bodnar, Haitz S´ aez de Oc´ ar iz Borde, Michael M. Bronstein, Petar Veliˇ ckovi´ c, and Pietro Li` o. “Sheaf Neural Networks with Connection Laplacians.” In Proceedings of the ICML 2022 Workshop on Topology, Algebra, an d Geometry in Machine Learning, PMLR 196:1–13, 2022. arXiv:2206.08702

arXiv 2022
[9]

Learning Sheaf Laplaci ans from Smooth Signals

Jakob Hansen and Robert Ghrist. “Learning Sheaf Laplaci ans from Smooth Signals.” In Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2019) , pp. 5446–5450. DOI: 10.1109/ICASSP.2019.8683709

arXiv 2019
[10]

Constructive Sheaf Semantics

Sterling, J. Constructive Sheaf Semantics. https://github.com/jonsterling/constructive-sheaf-s emantics

work page

Show all 26 references

[11]

Type-Topology in Univalent Foundations

Escardo, M. Type-Topology in Univalent Foundations. https://www.cs.bham.ac.uk/~mhe/agda-new/
[12]

Homotopy Type Theo ry: Univalent Foundations of Mathematics

The Univalent Foundations Program. Homotopy Type Theo ry: Univalent Foundations of Mathematics. Institute for Advanced Study, 2013

2013
[13]

HVM2: A Parallel Evaluator for Interaction C ombinators

Taelin, V. HVM2: A Parallel Evaluator for Interaction C ombinators. Work in progress, 2024. 63 64 BIBLIOGRAPHY

2024
[14]

Sheaves in Geometry and Logi c: A First Introduction to Topos Theory

Mac Lane, S., & Moerdijk, I. Sheaves in Geometry and Logi c: A First Introduction to Topos Theory. Springer, 1994

1994
[15]

Interaction Nets

Lafont, Y. Interaction Nets. POPL ’90, 1990

1990
[16]

Interaction Combinators

Lafont, Y. Interaction Combinators. Information and C omputation, 137(1):69–101, 1997

1997
[17]

A Denotational Semantics for the Symmetric In teraction Combinators

Mazza, D. A Denotational Semantics for the Symmetric In teraction Combinators. Mathemat- ical Structures in Computer Science, 17:527–62, 2007

2007
[18]

Sheaf Neural Networks

Hansen, J., & Gebhart, T. Sheaf Neural Networks. arXiv: 2012.06333, 2020

2012 arXiv
[19]

Topology and data

Carlsson, G. Topology and data. Bulletin of the America n Mathematical Society, 46(2):255- 308, 2009

2009
[20]

Vigneaux, J. P. Topology of Statistical Systems: A Coho mological Approach to Information Theory. Springer, 2019

2019
[21]

P., Li´ o, P ., & Bronstein, M

Bodnar, C., Di Giovanni, F., Chamberlain, B. P., Li´ o, P ., & Bronstein, M. M. Neural Sheaf Diﬀusion: A Topological Perspective on Heterophily an d Oversmoothing in GNNs. arXiv:2202.04579, 2022

2022 arXiv
[22]

Curry, J. M. Sheaves, Cosheaves, and Applications. Uni versity of Pennsylvania, 2014

2014
[23]

Vakil, The Rising Sea: Foundations of Algebraic Geometry

R. Vakil, The Rising Sea: Foundations of Algebraic Geometry . Draft (July 27, 2024), to be published by Princeton University Press, 2024

2024
[24]

J. J. Rotman, An Introduction to Homological Algebra , 2nd ed. Springer, 2009

2009
[25]

J. H. Gallier and J. Quaintance, Homology, Cohomology, and Sheaf Cohomology for Algebraic Topology, Algebraic Geometry, and Diﬀerential Geometry . Manuscript, Dept. of Computer and Information Science, Univ. of Pennsylvania, 2024

2024
[26]

Jagathese, ”Math 553: Algebraic Geometry II Lecture Notes (Spring 2023, UIC),” 2023

V. Jagathese, ”Math 553: Algebraic Geometry II Lecture Notes (Spring 2023, UIC),” 2023. Lecture notes transcribed by the author from a course taught by Kevin Tucker

2023
[27]

Hartshorne, Algebraic Geometry, vol

R. Hartshorne, Algebraic Geometry, vol. 52 of Graduate Texts in Mathematics. Springer, 1977

1977
[28]

Grothendieck, ”Sur quelques points d’alg` ebre homo logique,” Tˆ ohoku Math

A. Grothendieck, ”Sur quelques points d’alg` ebre homo logique,” Tˆ ohoku Math. J. (2), vol. 9, no. 3, pp. 119–221, 1957

1957

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Distributed Multi-Agent Coordination over Cellular Sheaves

Tyler Hanks, Hans Riess, Samuel Cohen, Trevor Gross, Mat thew Hale, and James Fairbanks. Distributed Multi-Agent Coordination over Cellular Sheaves . arXiv:2504.02049, 2025 (under review)

arXiv 2025

[2] [2]

Lattice Theory in Multi-Agent Systems

Hans Riess. Lattice Theory in Multi-Agent Systems . Ph.D. thesis, University of Pennsylva- nia, 2023. Also available as arXiv:2304.02568

work page Pith review arXiv 2023

[3] [3]

A Gentle Introduction to Sheaves on Graphs

Jakob Hansen. A Gentle Introduction to Sheaves on Graphs . Preprint, 2020. Available at http://www.jakobhansen.org/publications/gentleintroduction.pdf

work page 2020

[4] [4]

Justin M. Curry. Sheaves, Cosheaves and Applications . Ph.D. thesis, University of Pennsylva- nia, 2014. Available as arXiv:1303.3255

arXiv 2014

[5] [5]

Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers

Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and J onathan Eckstein. “Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers.” Foundations and Trends in Machine Learning 3(1):1–122, 2011

work page 2011

[6] [7]

Sheaf Neural Networks with Connection Laplacians

Federico Barbero, Cristian Bodnar, Haitz S´ aez de Oc´ ar iz Borde, Michael M. Bronstein, Petar Veliˇ ckovi´ c, and Pietro Li` o. “Sheaf Neural Networks with Connection Laplacians.” In Proceedings of the ICML 2022 Workshop on Topology, Algebra, an d Geometry in Machine Learning, PMLR 196:1–13, 2022. arXiv:2206.08702

arXiv 2022

[7] [9]

Learning Sheaf Laplaci ans from Smooth Signals

Jakob Hansen and Robert Ghrist. “Learning Sheaf Laplaci ans from Smooth Signals.” In Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2019) , pp. 5446–5450. DOI: 10.1109/ICASSP.2019.8683709

arXiv 2019

[8] [10]

Constructive Sheaf Semantics

Sterling, J. Constructive Sheaf Semantics. https://github.com/jonsterling/constructive-sheaf-s emantics

work page