Pith. sign in

REVIEW 3 major objections 30 references

From Self-Attention to Connection Laplacian: A Unified Operator View of Transformers

T0 review · 3 major / 0 minor · reviewed 2026-07-14 · grok-4.5

Pith's one-line read Multi-head attention is exactly one edge-dependent connection walk on the token graph.

desk verdict Correct algebraic rewrite of MHA as one edge-dependent connection walk, plus useful scale-consistent diagnostics; theorem is definitional, empirics are descriptive. read the letter →

arxiv 2607.10677 v1 pith:NDM2O3JL submitted 2026-07-12 cs.LG cs.CL

classification cs.LGcs.CL
keywords connectionLaplacianself-attentionmulti-headattentiontransformersgeometricdeeplearningwalktokengraphvectorfield
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 treats a token sequence as a vector field on the token-position graph and shows that self-attention is a connection walk: nonnegative walk weights mix messages while learned linear maps transport features along each edge. Single-head attention is exactly a constant-transport connection step; multi-head attention collapses exactly to a single scaled walk whose edge transport is the attention-gated mixture of the head value-output maps. The corresponding generator becomes a classical random-walk connection Laplacian only when the walk is stochastic and reversible and the transports are metric-compatible. On trained encoder and decoder Transformers from 124M to 8B parameters the authors then measure that effective attention graphs stabilize in deeper layers, that the effective transports approach scaled isometries, and that both signatures strengthen with scale. The result supplies a precise dictionary between ordinary attention and classical geometric operators, plus concrete operator-level diagnostics for any Transformer.

What carries the argument

The connection walk T(A, O): messages are aggregated by nonnegative walk weights A while each oriented edge transports features by a linear map O_ij. Theorem 3 proves that multi-head attention equals H · T(A_eff, O_eff), where A_eff is the average of the head attention matrices and O_eff is the A_eff-weighted mixture of the head transports M_h.

What would settle it

If the same models, measured on longer contexts, multiple domains, and full residual blocks rather than attention sublayers alone, showed persistently high adjacent-layer drift of A_eff and high scaled-isometry deviation of O_eff with no middle-layer stabilization, the claimed geometric structure would be refuted.

Watch

Extended reading notes

Core claim

Single-head attention is exactly a connection propagation step with constant transport. Multi-head attention is exactly a single scaled edge-dependent connection walk: its operator equals H times the connection walk built from the mean attention matrix and the attention-weighted mixture of the head transports.

Load-bearing premise

That short text probes and middle-layer averages of attention drift and transport orthogonality faithfully reveal a model-wide geometric regime rather than artifacts of sequence length, masking, sink handling, or layer choice.

Editorial extensions

If this is right

  • Effective attention graphs of trained Transformers stabilize across deeper layers into nearly stationary geometric operators.
  • Learned edge transports self-organize into approximate scaled isometries, more strongly as model scale increases.
  • The identification yields operator diagnostics—drift of A_eff, orthogonality deviation of O_eff, and connection energy—that can be read from any trained model.
  • Attention design can be reframed as choosing stable families of walks and edge transports (near-isometries, inverse consistency, symmetrizable walks).
  • When the walk is reversible and transports are metric-compatible, the attention generator reduces to a classical random-walk connection Laplacian with a nonnegative Dirichlet form.

Reading between the lines

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

  • Regularizing holonomy or non-isometry of O_eff may stabilize long-context training without changing the attention algebra.
  • Multi-head advantage can be read as a low-rank transport dictionary that attention selects edgewise, suggesting other operator bases could replace heads.
  • Connection-energy residuals are natural unsupervised signals for pruning heads, edges, or layers.
  • Lifting fibers from vectors to matrices or tensors would give a native operator view of vision and video Transformers on grids and space-time graphs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper models token sequences as vector fields on the token-position graph and identifies self-attention as a connection walk: nonnegative walk weights aggregate messages while edgewise linear maps transport features. Theorem 2 shows single-head attention is exactly connection propagation with constant transport M = WV WO. Theorem 3 shows multi-head attention is exactly a scaled edge-dependent connection walk TMHA = H · T(Aeff, Oeff), where Aeff is the mean of head attention matrices and Oeff_ij is the Aeff-weighted mixture of head transports Mh. Section 5 carefully states when the generator Lmix = I − T̄MHA reduces to a classical random-walk connection Laplacian (row-stochasticity, reversibility, metric-compatible transports), treating the generic directed non-isometric case as a directed connection-walk generator rather than a PSD Laplacian. Empirically, decoder-only models from 124M to 8B (plus BERT) are probed with layerwise drift of Aeff and a scaled-isometry diagnostic δorth on Oeff; middle layers often stabilize and approach approximate scaled isometries, with trends strengthening at larger scale. The paper supplies operator-level diagnostics and an ADR-style block interpretation, with an honest limitations section.

Significance. If accepted as an operator identification plus diagnostic toolkit, the work usefully bridges standard MHA and classical geometric operators (connection Laplacians, vector diffusion maps, sheaf/connection constructions) without redesigning attention. The SHA/MHA identities are elementary but exact and cleanly stated; the careful separation between the always-valid connection-walk rewrite and the optional Laplacian regime is a genuine contribution relative to prior graph/Laplacian readings of attention that mainly treat scalar mixing. Strengths include explicit block/Kronecker proofs, multi-family multi-scale empirics (GPT-2 through Llama-3/Qwen, encoder and decoder), and transparent limitations (attention-sublayer scope, descriptive not causal empirics). The paper does not ship machine-checked proofs or causal interventions, but the algebraic claims are checkable by hand and the diagnostics are falsifiable on other models/probes. The main value is conceptual unification and a reusable measurement language rather than a new architecture or performance result.

major comments (3)
  1. §4.2, Theorem 3 and Eqs. (8)–(11): The identity TMHA = H · T(Aeff, Oeff) holds by construction once Oeff is defined as the Aeff-weighted average of head maps that makes the (i,j) blocks match. This is a valuable reparameterization and geometric reading, not an independent structural constraint discovered about MHA. The abstract and §1 currently invite a stronger reading (“identify attention as a connection walk” / “exact reduction”). Please state explicitly that the theorem is an exact rewrite under these definitions, and that the scientific contribution is the geometric interpretation, the Laplacian-regime conditions, and the induced diagnostics—not a new restriction on what MHA can compute.
  2. Abstract, §7, and §9: The theory does not predict depthwise stabilization of Aeff or emergence of scaled isometries; those are empirical observations measured with diagnostics motivated by the formalism. Phrases such as “geometric structure consistent with our theory,” “signatures predicted by the theory” (§9), and similar wording overstate the link. Soften to “interpretable within / motivated by the connection-walk view,” and keep the clear descriptive stance already present in Limitations §8 (no causal claim that lower δorth improves performance).
  3. Title and Abstract vs §5.3: The title and opening lead with “Connection Laplacian,” but §5.3 correctly shows that Lmix is generally a directed, non-self-adjoint connection-walk generator; classical PSD connection-Laplacian structure requires reversibility and metric-compatible inverse-consistent transports that causal decoder attention typically lacks. Please align the title/abstract emphasis with the body (e.g., “connection walk / toward connection Laplacians”) so readers are not led to expect a generic Laplacian theorem for trained Transformers.

Circularity Check

2 steps flagged · score 4.0 of 10

Theorem 3 is an exact blockwise reparameterization: Oeff is defined as the attention-weighted mixture that forces TMHA = H·T(Aeff, Oeff); the 'proof' verifies that definition. Empirics are independent diagnostics, not forced predictions.

  1. self definitional [Theorem 3, Eqs. (8)–(11)]
    "Define the effective walk weights (mean attention) by Aeff(i, j) def= 1/H ∑_h A^(h)_ij ≥ 0. When Aeff(i, j)>0, define the effective edge transport by the Aeff-weighted average Oeff_ij def= (1/H ∑_h A^(h)_ij Mh) / Aeff(i, j). Then TMHA = H·T(Aeff, Oeff). Equivalently, Yi = H ∑_j Aeff(i, j) Xj Oeff_ij."

    Oeff is defined exactly as the attention-gated mixture of head transports that makes every (i,j) block of TMHA equal H Aeff(i,j) Oeff_ij^⊤. The proof only verifies this identity case-wise (active vs inactive edges). The claim 'MHA is exactly a single edge-dependent connection walk' therefore holds by construction of Oeff, not as an independent structural discovery.

  2. renaming known result [Theorem 2, §4.1]
    "Assume A is row-stochastic and the edge transport is constant, Oij ≡ M for all active edges (j→i). Then the single-head attention (SHA) Y = AXM is exactly the connection propagation step Y = T(A, O)X. Equivalently, in stacked form, T_SHA = T(A, O) = A ⊗ M^⊤."

    SHA is already Y = AXM. Declaring Oij ≡ M and writing the same formula as connection propagation T(A,O) renames the standard operator in connection-Laplacian vocabulary without adding algebraic content beyond the naming.

full rationale

The paper's central algebraic claim is an identification, not a free prediction. Theorem 2 rewrites single-head attention Y = AXM as connection propagation with constant transport Oij ≡ M—an elementary renaming into connection-walk language. Theorem 3 defines Aeff as the mean of head attentions and Oeff_ij as the Aeff-weighted average of head transports Mh, then proves TMHA = H·T(Aeff, Oeff). The proof is blockwise verification that this definition makes the equality hold; there is no independent content beyond the reparameterization. That is mild self-definitional circularity for the 'exact reduction' claim, but it is the normal structure of an operator-identification paper rather than a fitted-input-called-prediction or a self-citation chain. Conditions for reduction to a classical connection Laplacian (stochasticity, reversibility, metric-compatible transports) are stated as optional sufficient conditions drawn from the external literature, not claimed as theorems forced by the rewrite. Empirical δorth and geometric-drift measurements are descriptive probes of trained models (GPT-2 through Llama-3-8B) under those constructed operators; they do not fit parameters and then re-predict the same quantities. A minor self-citation (Lin et al. 2013, overlapping authors) appears only for the standard connection Dirichlet energy formula and is not load-bearing for Theorem 3. Overall: partial definitional circularity on the exact-identification claim (score 4), with independent empirical content and no load-bearing self-citation chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The central algebraic claim needs only standard linear algebra and the usual definition of multi-head attention; no fitted constants enter Theorems 2–3. Load-bearing modeling choices are the token-graph vector-field view, the definition of effective walk/transport that makes MHA a single connection step, and the empirical probe design (dataset, length, masks, middle-layer focus). Invented entities are definitional operators, not new physical objects.

free parameters (3)
  • context length / probe truncation (default 64)
    Main diagnostics use length-64 WikiText samples; longer lengths are ablations. Choice affects early-layer drift and mildly δorth.
  • middle-layer depth window (roughly 20–80%)
    Stability and isometry claims focus on middle layers by design; boundary layers are shaded out as embedding/projection dominated.
  • temperature τ and head count H in attention definitions
    Standard model hyperparameters; H appears as the scale factor in Theorem 3 and is absorbed into diagnostics by normalization.
assumptions (5)
  • domain assumption Standard multi-head attention formula Y = sum_h A^(h) X M_h with row-wise softmax attention and composite transports M_h = W_V^(h) W_O^(h).
    Taken as the object of study (§2); all exact identities rewrite this formula.
  • domain assumption Token positions form a directed graph; features form a discrete vector field; edge maps O_ij act by right multiplication on row vectors.
    Modeling setup in §3 that makes connection-walk language applicable.
  • ad hoc to paper Aeff and Oeff defined by head averaging (Eqs. 8–9) are the correct effective walk and transport for geometric analysis of MHA.
    This definition is chosen so Theorem 3 holds exactly; alternative aggregations over heads are not explored.
  • standard math Classical random-walk connection Laplacian requires row-stochastic reversible walks and metric-compatible (e.g. orthogonal, inverse-consistent) transports for self-adjoint PSD Dirichlet form.
    Invoked in §5 from Singer–Wu / Bandeira et al. style theory; used to state when attention is a true connection Laplacian vs a directed generator.
  • ad hoc to paper Scaled-isometry diagnostic: O^T O ≈ μ I on active edges is a meaningful proxy for geometry-preserving transport.
    δorth (Eq. 15) is the paper's operational metric; not required for the exact walk identity.
invented entities (2)
  • Effective edge transport Oeff_ij (attention-gated mixture of head transports) independent evidence
    purpose: Collapse multi-head sum into a single connection-walk operator per edge.
    Defined in Eq. 9 specifically so TMHA = H T(Aeff, Oeff); independent evidence is only that trained models make Oeff near scaled-isometric under δorth.
  • Connection-walk generator Lmix = I − T̄_MHA as directed ADR-style operator on token fields
    purpose: Link attention to connection Laplacian / diffusion–advection–reaction interpretations and continuous limits.
    Interpretive generator in §5–6; not a new physical field, and PSD Laplacian properties require extra assumptions the paper does not claim hold generically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Self-Attention to Connection Laplacian: A Unified Operator View of Transformers." pith.science (2026). https://pith.science/paper/NDM2O3JL

@misc{pith2026260710677,
  author       = {Pith},
  title        = {Pith review of: From Self-Attention to Connection Laplacian: A Unified Operator View of Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NDM2O3JL}},
  note         = {Machine review of arXiv:2607.10677}
}
read the original abstract

Self-attention is a ubiquitous primitive in modern sequence models, yet its operator-level geometry is only partially understood. We view a token sequence as a vector field over the token-position graph and identify attention as a connection walk: messages are aggregated by a nonnegative walk matrix while being transported along each edge by a learned linear map. Within this framework, we prove that single-head attention (SHA) is exactly a connection propagation step with constant transport, and that multi-head attention (MHA) is exactly a single edge-dependent connection walk whose effective transport is an attention-gated mixture of headwise transports. We further clarify the conditions under which the corresponding generator reduces to a random-walk connection Laplacian, highlighting the roles of stochasticity, reversibility, and metric-compatible transports. Empirically, we find that trained Transformers across scales (from 124M to 8B) and structures (encoder/decoder) exhibit geometric structure consistent with our theory: effective attention graphs converge to stable geometric operators in deeper layers, learned transports self-organize into approximate scaled isometries, and both phenomena strengthen consistently with scale. Overall, the paper provides a precise connection-walk formalism that links self-attention to classical geometric operators, along with a set of operator-level tools for analyzing transformer models from a geometric perspective.

Figures

Figures reproduced from arXiv: 2607.10677 by the authors.

Figure 1
Figure 1. Layerwise diagnostics of effective walks and transports. Shaded re￾gions indicate the initial and final layers, which are often dominated by embedding and output-projection effects; the discussion focuses on the stable middle-layer regime. (a, b) Geometric Drift: Both models exhibit stabilization patterns. GPT-2 Small shows higher variance, whereas Llama-3-8B reaches a lower-drift middle-layer regime (d ℓ→ℓ+1 A ≈ 0)… view at source ↗
Figure 2
Figure 2. Visualization of Learned Operators (GPT-2 Small). (a) Effective walk weights (Aeff): With the attention sink removed, middle layers exhibit a coherent structure, consistent with a stable local-to-global topology. (b) Transport Gram structure (Gij): The Gram matrix is strongly diagonal-dominant. The suppression of off-diagonal elements is consistent with approximate scaled-isometry under the proposed diagnostic, whil… view at source ↗
Figure 3
Figure 3. Sensitivity to context length (GPT-2 Small). We vary the input context length from 16 to 1,024 tokens. Longer contexts increase early-layer drift and mildly affect orthogonality, while deeper layers recover a similar low-drift and low-deviation profile in this tested range. graph is more difficult. However, the overall profile remains similar: deeper layers recover a low-drift and low-deviation regime within the tes… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: GPT-2 Series Evolution. Comparison of geometric stability (Left) and transport orthogonality (Right) across four model sizes. The larger variants in this controlled family show smoother curves and lower orthogonality deviation. 0 2 4 6 8 10 Layer Index 0.00 0.10 0.20 0…
Figure 7
Figure 7. Figure 7: BERT-base-uncased operator diagnostics using the full valid bidirectional non [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 5
Figure 5. Figure 5: Qwen Series Evolution. Both Qwen2.5 and Qwen3 show low middle-layer drift under the geometric diagnostic. 0 5 10 15 20 25 Layer Index 0.00 0.05 0.10 0.15 0.20 D rift ( 1 − c o s simila rity ) Attention Geometric Drift Stable Regime (Middle Layers) 0 5 10 15 20 25 Layer…
Figure 6
Figure 6. Figure 6: Llama Series Evolution. Llama models show low drift and low transport deviation under the proposed diagnostics. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 8
Figure 8. Figure 8: Impact of Sequence Length. Comparing drift and orthogonality across model families for context lengths L = 16 to L = 256. Across the tested architectures, deeper layers often recover lower drift despite increased initial drift for longer sequences. 22 [PITH_FULL_IMAGE…
Figure 9
Figure 9. Figure 9: Connection-energy diagnostics for GPT-2 Small and BERT-base-uncased. The [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Extended context-length diagnostic for GPT-2 Small, probing context lengths [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 4 canonical work pages

  1. [1]

    Gomez and Lukasz Kaiser and Illia Polosukhin , editor =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , editor =. Attention is All you Need , booktitle =. 2017 , timestamp =

  2. [2]

    Rethinking Attention with Performers , booktitle =

    Krzysztof Marcin Choromanski and Valerii Likhosherstov and David Dohan and Xingyou Song and Andreea Gane and Tam. Rethinking Attention with Performers , booktitle =. 2021 , timestamp =

  3. [3]

    2023 , url =

    Yi Tay and Mostafa Dehghani and Dara Bahri and Donald Metzler , title =. 2023 , url =. doi:10.1145/3530811 , timestamp =

  4. [4]

    Wright and Joseph E

    Matthew A. Wright and Joseph E. Gonzalez , title =. CoRR , volume =. 2021 , url =. 2106.01506 , timestamp =

  5. [5]

    CoRR , volume =

    Sankalan Pal Chowdhury and Adamos Solomou and Avinava Dubey and Mrinmaya Sachan , title =. CoRR , volume =. 2021 , url =. 2110.08323 , timestamp =

  6. [6]

    Susskind , editor =

    Shuangfei Zhai and Tatiana Likhomanenko and Etai Littwin and Dan Busbridge and Jason Ramapuram and Yizhe Zhang and Jiatao Gu and Joshua M. Susskind , editor =. Stabilizing Transformer Training by Preventing Attention Entropy Collapse , booktitle =. 2023 , url =

  7. [7]

    On the Role of Attention Masks and LayerNorm in Transformers , booktitle =

    Xinyi Wu and Amir Ajorlou and Yifei Wang and Stefanie Jegelka and Ali Jadbabaie , editor =. On the Role of Attention Masks and LayerNorm in Transformers , booktitle =. 2024 , timestamp =

  8. [8]

    Hopfield Networks is All You Need , booktitle =

    Hubert Ramsauer and Bernhard Sch. Hopfield Networks is All You Need , booktitle =. 2021 , timestamp =

Show all 30 references
  1. [9]

    Kim , editor =

    Seongjun Yun and Minbyul Jeong and Raehyun Kim and Jaewoo Kang and Hyunwoo J. Kim , editor =. Graph Transformer Networks , booktitle =. 2019 , timestamp =

  2. [10]

    Tensorized Self-Attention: Efficiently Modeling Pairwise and Global Dependencies Together , booktitle =

    Tao Shen and Tianyi Zhou and Guodong Long and Jing Jiang and Chengqi Zhang , editor =. Tensorized Self-Attention: Efficiently Modeling Pairwise and Global Dependencies Together , booktitle =. 2019 , url =. doi:10.18653/V1/N19-1127 , timestamp =

  3. [11]

    Communications on pure and applied mathematics , volume =

    Vector diffusion maps and the connection Laplacian , author =. Communications on pure and applied mathematics , volume =. 2012 , publisher =

  4. [12]

    Bandeira and Amit Singer and Daniel A

    Afonso S. Bandeira and Amit Singer and Daniel A. Spielman , title =. 2013 , url =. doi:10.1137/120875338 , timestamp =

  5. [13]

    Gauge Equivariant Convolutional Networks and the Icosahedral

    Taco Cohen and Maurice Weiler and Berkay Kicanaoglu and Max Welling , editor =. Gauge Equivariant Convolutional Networks and the Icosahedral. Proceedings of the 36th International Conference on Machine Learning,. 2019 , timestamp =

  6. [14]

    Bronstein and Joan Bruna and Taco Cohen and Petar Velickovic , title =

    Michael M. Bronstein and Joan Bruna and Taco Cohen and Petar Velickovic , title =. CoRR , volume =. 2021 , url =. 2104.13478 , timestamp =

  7. [15]

    Hamilton and Vincent L

    Devin Kreuzer and Dominique Beaini and William L. Hamilton and Vincent L. Rethinking Graph Transformers with Spectral Attention , booktitle =. 2021 , timestamp =

  8. [16]

    CoRR , volume =

    Ahmed Farooq , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.11043 , eprinttype =. 2506.11043 , timestamp =

  9. [17]

    Slabaugh and Stefanos Zafeiriou , title =

    Francesca Babiloni and Ioannis Marras and Gregory G. Slabaugh and Stefanos Zafeiriou , title =. 2020. 2020 , doi =

  10. [18]

    A Tensorized Transformer for Language Modeling , booktitle =

    Xindian Ma and Peng Zhang and Shuai Zhang and Nan Duan and Yuexian Hou and Ming Zhou and Dawei Song , editor =. A Tensorized Transformer for Language Modeling , booktitle =. 2019 , timestamp =

  11. [19]

    Tensor Product Attention Is All You Need , journal =

    Yifan Zhang and Yifeng Liu and Huizhuo Yuan and Zhen Qin and Yang Yuan and Quanquan Gu and Andrew Chi. Tensor Product Attention Is All You Need , journal =. 2025 , url =. doi:10.48550/ARXIV.2501.06425 , eprinttype =. 2501.06425 , timestamp =

  12. [20]

    CoRR , volume =

    Tianyu Ruan and Shihua Zhang , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.18288 , eprinttype =. 2412.18288 , timestamp =

  13. [21]

    On Layer Normalization in the Transformer Architecture , booktitle =

    Ruibin Xiong and Yunchang Yang and Di He and Kai Zheng and Shuxin Zheng and Chen Xing and Huishuai Zhang and Yanyan Lan and Liwei Wang and Tie. On Layer Normalization in the Transformer Architecture , booktitle =. 2020 , url =

  14. [22]

    Nguyen and Julian Salazar , editor =

    Toan Q. Nguyen and Julian Salazar , editor =. Transformers without Tears: Improving the Normalization of Self-Attention , booktitle =. 2019 , url =

  15. [23]

    Annals of Combinatorics , volume =

    Laplacians and the Cheeger inequality for directed graphs , author =. Annals of Combinatorics , volume =. 2005 , publisher =

  16. [24]

    Nonlinear Phenomena in Complex Systems , volume =

    A Primer on Laplacian Dynamics in Directed Graphs , author =. Nonlinear Phenomena in Complex Systems , volume =. 2020 , doi =

  17. [25]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,

    Jacob Devlin and Ming. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,. 2019 , url =. doi:10.18653/V1/N19-1423 , timestamp =

  18. [26]

    Binbin Lin and Xiaofei He and Chiyuan Zhang and Ming Ji , title =. J. Mach. Learn. Res. , volume =. 2013 , url =. doi:10.5555/2567709.2567755 , timestamp =

  19. [27]

    Anh Tong and Thanh Nguyen. Neural. The Thirteenth International Conference on Learning Representations,. 2025 , timestamp =

  20. [28]

    Osher , title =

    Kelvin Kan and Xingjian Li and Stanley J. Osher , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2501.18793 , eprinttype =. 2501.18793 , timestamp =

  21. [29]

    Neural Sheaf Diffusion:

    Cristian Bodnar and Francesco Di Giovanni and Benjamin Paul Chamberlain and Pietro Li. Neural Sheaf Diffusion:. Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 2...

  22. [30]

    Sheaf Neural Networks with Connection Laplacians , booktitle =

    Federico Barbero and Cristian Bodnar and Haitz S. Sheaf Neural Networks with Connection Laplacians , booktitle =. 2022 , url =

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.