Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Geometric Generative Modeling with Noise-Conditioned Graph Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A graph generative model's architecture should be conditioned on noise level: its Dynamic Message Passing, which widens its reach and coarsens its view as noise rises, beats fixed-architecture GNNs on 3D shapes, cell data, and images.

desk verdict The DMP architecture is a genuinely useful idea with credible strong results, but the motivating theory has a real algebra error that breaks Theorem 3.2; fix the appendix or reframe the constraint. read the letter →

arxiv 2507.09391 v1 pith:KHOMGWJR submitted 2025-07-12 cs.LG

Peter Pao-Huang , Mitchell Black , Xiaojie Qiu This is my paper
classification cs.LG
keywords noise-conditionedgraphnetworksdynamicmessagepassinggeometricgenerationflow-basedgenerativemodelsneuralnoise-adaptivearchitecture3Dpointcloudspatiotemporaltranscriptomics
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

Flow-based generative models of geometric graphs learn to reverse a noising process that gradually turns data into Gaussian noise, and the standard recipe runs the same fixed graph neural network at every noise level. This paper tries to establish that the architecture itself should follow the noise: the noisier the graph, the more distant the neighbors that carry useful signal, and the fewer nodes needed to represent it faithfully. From that principle it builds Dynamic Message Passing (DMP), which schedules message-passing range and graph resolution as monotone functions of the noise level, interpolating from a sparse fine-grained graph at low noise to a dense coarse-grained one at high noise while keeping linear-time message passing. DMP consistently outperforms the noise-independent baselines, with a 16.15% average improvement in Wasserstein distance on ModelNet40, better-or-comparable results on simulated spatiotemporal gene-expression graphs, and an ImageNet FID of 63.983 versus 84.051 for DiT at identical computational cost.

What carries the argument

The object carrying the argument is the noise-conditioned graph network (NCGN): a GNN whose message-passing range and node resolution are scheduled by the noise level $t$, in contrast to the fixed architecture of prior work. Its concrete realization is Dynamic Message Passing (DMP), whose forward pass coarsens the noisy graph to $s_t$ supernodes, builds edges among supernodes with range $r_t$ (e.g. $k$NN), runs the ordinary GNN layers on that coarse graph, and uncoarsens back to the original nodes so the predicted vector field is defined at full resolution. The two knobs come from a predefined scheduler $f$ satisfying a monotonicity constraint — $r' \ge r$ and $s' \le s$ when noise increases — plus boundary conditions. The theoretical motor is Lemma 3.1's exact formula $$ I($x_1^{{(i)}}$, $Y_t^{{(i,r)}}$) = \frac{1}{2}\log\!\left(\frac{2r/\mathrm{SNR}(t) + A}{2r/\mathrm{SNR}(t) + A - $B^{2}$}\right), \quad A = \iint_{D_r(i)} \rho(\$eta^{{(j)}}$,\$eta^{{(k)}}$)\,dj\,dk,\; B = \int_{D_r(i)} \rho(\$eta^{{(i)}}$,\$eta^{{(j)}}$)\,dj, $$ which converts 'how far should a message travel' into a calculus problem whose optimum provably moves outward as the signal-to-noise ratio drops (Theorem 3.2); the resolution conclusion follows because aggregation is itself a form of pooling, so the same radius result justifies coarsening at high noise.

What would settle it

A decisive experiment is a per-noise-level search: train on ModelNet40 or the transcriptomics dataset with the schedule frozen at several fixed noise levels (or sweep range and resolution at each level) and record the Wasserstein distance — if the best range does not grow, or the best resolution does not shrink, monotonically with noise, the design principle behind DMP is refuted. The position-noising justification is checkable even more cheaply: Proposition 3.3's closed form, evaluated at $t=0$, should return the clean squared distance $\gamma^2$, whereas the stated expression returns $2(1+\rho(\gamma))$, so the formula as written fails its own zero-noise boundary condition.

Watch

Extended reading notes

Core claim

The central claim is that noise level should be an architectural input, not merely a conditioning scalar: the denoising network should be written $v_\theta(G_t, t) = \mathrm{NCGN}_t(G_t)$, with range and resolution changing in $t$, rather than $v_\theta(G_t, t) = \mathrm{GNN}(G_t)$ on a graph structure fixed across the whole process. The supporting argument analyzes a Gaussianized geometric graph in information-theoretic terms. Lemma 3.1 gives the mutual information between a clean node feature and the aggregation of noised features within radius $r$; Theorem 3.2 shows that under the correlation model $\rho(\eta^{(i)},\eta^{(j)}) = 1-(\eta^{(i)}-\eta^{(j)})^2$ the radius maximizing that mutual information strictly increases as the signal-to-noise ratio falls, and Proposition 3.3 argues the same qualitative direction for noised positions. For resolution, the paper demonstrates with Gromov-Wasserstein distances that heavily noised graphs are better preserved by coarse-graining than at full resolution. The concrete instantiation, DMP, coarsens the graph into supernodes, passes messages among them with range $r_t$, and uncoarsens back to the original nodes, using a scheduler $f(t, r_0, r_1, s_0, s_1)$ constrained to be monotone (wider range and fewer supernodes as noise rises) and to hit boundary conditions at $t=0$ and $t=1$; with endpoints chosen so that $r_t s_t = r_1 N$, message passing stays linear-time throughout generation. The paper reports that this single change beats the noise-independent baselines in every tested domain, including an ImageNet FID improvement from 84.051 to 63.983 at equal compute for a modified DiT.

Load-bearing premise

The load-bearing premise is that the simplified Gaussian analysis — which shows the optimal aggregation radius grows as noise increases, proved for a hand-picked correlation function on a continuum of nodes — transfers to real discrete geometric graphs and learned message passing; the paper's evidence for the transfer is the qualitative attention-weight pattern in Section 3.2.

Editorial extensions

If this is right

  • Noise-independent flow models leave measurable performance unused: adopting a monotone range-and-resolution schedule is a drop-in change that improved Wasserstein distance on ModelNet40 by 16.15% on average across both GCN and GAT backbones.
  • The same exponential schedule transferred across three unrelated domains — 3D shapes, simulated gene-expression graphs, and images — indicating the design rule 'wider and coarser as noise rises' is architecture-agnostic, holding for message-passing GNNs and for transformer-based image diffusion.
  • The principle costs nothing in scaling: with boundary endpoints chosen so that $r_t s_t = r_1 N$, DMP keeps linear-time message passing while outperforming quadratic fully connected baselines.
  • In image generation, patch size and attention range play the roles of resolution and range, so latent diffusion models can adopt noise-conditioned patching as a free improvement at equal compute.
  • Because only monotonicity and boundary conditions are imposed on the scheduler, the theoretical guarantee attaches to the constraint class, not to the particular exponential schedule used in the experiments.

Reading between the lines

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

  • The paper's own layer-count ablation shows DMP degrades like a fully connected network once message-passing layers reach 8 or more; an untested extension is to schedule depth as well as range and resolution — dense communication when noise is high, shallow sparse layers when it is low — which would target the 2-to-4-layer regime where DMP already dominates.
  • The scheduler is hand-picked in this paper; the authors flag learnable range and resolution as future work. If monotonicity is the only constraint that matters, a learned monotone scheduler should beat the fixed exponential form by adapting per dataset.
  • The central theorem is proved in a continuum idealization with one hand-picked correlation, yet the attention analysis suggests the pattern survives in trained GATs; comparing the radius that maximizes Lemma 3.1's mutual information at each signal-to-noise ratio against the empirically optimal radius found by search would localize where the theory stops being predictive.
  • Under the geometric-graph view of images, the ImageNet result predicts that any patch-based latent diffusion model can treat patch size and attention window as noise-conditioned hyperparameters; whether the gain persists at DiT-Large or XL scale, where each patch encodes more learned structure, is a testable next step.
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

4 major / 5 minor

Summary. The paper introduces Noise-Conditioned Graph Networks (NCGNs), in which the connectivity radius and node resolution of a GNN used in flow-based generative modeling are scheduled as functions of the noise level. The authors support this design with an information-theoretic analysis of a continuum geometric graph, a qualitative attention-weight analysis, and three sets of experiments: ModelNet40 point-cloud generation, simulated spatiotemporal transcriptomics, and ImageNet 256×256 generation using a modified DiT. They report substantial improvements, including a 16.15% average Wasserstein reduction on ModelNet40 and FID 63.983 versus 84.051 for DiT at equal compute. The central claim is that as noise increases, message passing should reach farther and use coarser nodes.

Significance. If correct, this is a practically valuable and broadly applicable architectural principle: converting existing GNN or transformer generative models to noise-conditioned range/resolution schedules requires minimal code changes and preserves computational complexity. The manuscript is clearly written, releases code, and evaluates across three domains; the ImageNet experiment in particular is an attractive proof of concept. However, the theoretical backbone contains algebraic and kernel-validity errors, and the experimental comparisons lack uncertainty information and are partially inconsistent with the stated design. These issues must be repaired before the central claims can be accepted.

major comments (4)
  1. [Appendix A.1, Lemma A.1] The denominator in Lemma A.1 is misderived. Substituting the A and B computed in the same lemma into Lemma 3.1 gives A - B^2 = -4r^6/9, so the denominator is 2r/c - 4r^6/9 = r(2/c - 4r^5/9), not the printed 2/c + 4r^6/9. Lemma A.2 and Lemma A.3 differentiate this incorrect expression. With the corrected κ, the monotonicity claim in Theorem 3.2 fails: for c=1 the corrected κ is maximized at the boundary r=1, while for c=0.1 it has an interior maximum near r≈0.8, so lowering the SNR can decrease, not increase, the optimal radius. The monotonicity constraint imposed on the scheduler in Section 4.1 therefore rests on an invalid theorem rather than on a transfer assumption. In addition, ρ(η(i),η(j)) = 1-(η(i)-η(j))^2 is not a valid correlation kernel on [-r,r]: the Gram matrix at positions 0, 0.5, 1 has determinant -1/8, so the Gaussian mutual-information expression in Lemma 3.1 is not well-defined for the claimed range of (r,c).
  2. [Section 3.1.2 / Appendix A.2, Proposition 3.3] The variance of a difference is computed with the wrong sign. For variance-preserving noise, var(η(i)_t - η(j)_t) = 2(1-t)(1-ρ(γ)), not 2(1-t)+2ρ(γ)(1-t). Consequently the displayed expected squared distance in Proposition 3.3 is incorrect. As written, this proposition does not reliably establish that the distance between nearby nodes grows with the noise level, so it cannot serve as the positional-noising justification for increasing the message-passing radius. The correct formula should be derived and the implications re-evaluated.
  3. [Tables 1-4] All headline experimental results are reported as single point estimates with no number of seeds, standard deviations, or confidence intervals, so the size and reliability of the claimed 16.15% average improvement and the ImageNet FID gap cannot be assessed. Moreover, Table 3 shows that DMP with GAT is worse than RANDOMKNN on gene imputation (2.575 versus 2.527) and gene knockout (2.397 versus 2.390), and only comparable on several other tasks. This contradicts the abstract's statement that DMP “consistently outperforms” noise-independent architectures; the more guarded wording in Section 5.2 is closer to the data.
  4. [Appendix B.2, Table 5] The boundary conditions in Table 5 appear to be reversed relative to the design stated in Section 4.1. Section 4.1 specifies at low noise t=1 full resolution and sparse connectivity, and at high noise t=0 coarse resolution and dense connectivity, but Table 5 lists kNN(t=1)=Fully connected, Clusters(t=1)=[√cN], kNN(t=0)=[3√N], and Clusters(t=0)=N. If the experiments actually used this schedule, they do not instantiate DMP as described; if this is a typo, it must be corrected because the boundary conditions are load-bearing for the complexity argument and for reproducibility.
minor comments (5)
  1. [Figure 2] The axes and the precise family of curves are not labeled; since the figure is based on the erroneous Lemma A.1, it should be regenerated from a corrected expression and preferably with error bars or a quantitative legend.
  2. [Section 3.2] The attention-weight analysis in Figure 3 averages over the dataset but reports no variation across shapes or noise levels; a few representative curves with variance bands would make the claimed qualitative support more convincing.
  3. [Appendix B.2, Table 5] The notation [√cN] and [3√N] is ambiguous: it is not stated whether brackets denote rounding, and whether c is the same c as the signal-to-noise ratio in the theoretical section.
  4. [Section 5.2] The claim that the exponential schedule is used because it “works best” should note that the comparison in Table 2 was run only on ModelNet40 with a diffusion objective; the transfer of this choice to flow-matching on transcriptomics is an assumption rather than an established result.
  5. [Section 5.3] The statement that DiT-DMP has the same computational complexity as baseline DiT is not fully derived; spelling out the tradeoff between the reduced attention range and the reduced sequence length would make the equivalence transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the derivation is self-contained; the appendix algebra error is a correctness concern, not a circular reduction.

full rationale

The paper's central claims do not reduce to their own inputs by construction. Lemma 3.1 derives a mutual-information formula from stated Gaussian and continuum assumptions on the noising process, and Lemma A.1 evaluates the integrals A and B for a specific, explicitly chosen correlation kernel. Theorem 3.2 and the monotonicity constraint on the DMP scheduler are then consequences of this analytic chain plus the empirical attention and Gromov-Wasserstein analyses in Section 3.2, not of the reported generation results. The DMP scheduler is selected on ModelNet40 and then transferred to other domains; this is model selection on one benchmark, not a fitted parameter being renamed as a prediction, and the reported improvements are not entailed by the scheduler constraint alone. The self-citations to Pao-Huang et al. 2023, Jing et al. 2023, and Qiu et al. 2024 are background or application context and are not load-bearing for any uniqueness claim or architectural choice; no cited result is invoked to forbid alternatives. The paper openly acknowledges that the scheduler is predefined rather than learned, which is an honest limitation rather than a circular step. The principal caveat is a mathematical correctness issue outside the circularity rubric: substituting the paper's own A and B into Lemma 3.1 gives A - B^2 = -4r^6/9, so the denominator in Lemma A.1 should be 2/c - 4r^5/9 rather than 2/c + 4r^6/9, which affects the support for Theorem 3.2. That is an internal algebraic error, not a reduction of the theorem to its assumptions, and the paper does not present this step as an independent empirical prediction. Therefore, no significant circularity is present, and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claims rest on a Gaussianized model with a hand-chosen correlation kernel and a continuum approximation, plus hand-set boundary conditions and a schedule selected on ModelNet40. These choices are stated transparently but limit the proven scope of the theory; the empirical sections show the method works on selected benchmarks but without statistical error bars.

free parameters (3)
  • Adaptive schedule function f (exponential) = exponential interpolation between (r0,s0) and (r1,s1)
    Chosen after comparing linear, exponential, logarithm, and ReLU schedules on ModelNet40 (Table 2); the exponential form is then used for the transcriptomics and ImageNet experiments.
  • Boundary conditions (r0,r1,s0,s1) per experiment = Hand-set per dataset, e.g., ModelNet40: r0=c, r1=N, s0=N, s1=floor(sqrt(c*N))
    They determine the DMP complexity and the range of noise-conditioned behavior. The paper uses these as hyperparameters, not learned.
  • Voxel clustering grid resolution = d√s uniform partitions per position dimension
    Coarsening assigns original nodes to supernodes via a fixed voxel grid; the number of partitions depends on the number of supernodes s_t at each noise level.
assumptions (5)
  • domain assumption Noised positions/features are approximately isotropic Gaussian N(Z_t, diag(σ_t^2))
    Section 3 Preliminaries: 'the node positions or features of a graph can be approximated as an isotropic multivariate Gaussian'. All theoretical analysis uses this.
  • ad hoc to paper Nodes form a continuum indexed by R with η(i)=i
    Appendix A.1 assumption (iii): 'we treat the graph in a continuous manner... with the position η(i)=i'. This makes integrals tractable but is not true for discrete geometric graphs.
  • ad hoc to paper Feature correlation is ρ(η(i),η(j)) = 1 - (η(i)-η(j))^2
    Theorem 3.2 assumes this specific correlation to get closed-form integrals; generality beyond this kernel is only claimed empirically.
  • domain assumption Message passing can be modeled as summation over a ball of radius r
    Lemma 3.1 defines accessible information as Y_t^(i,r) = integral over the ball of noised features; real GNNs use learned attention and nonlinearities, so the optimality result may not transfer.
  • domain assumption Variance-preserving noise schedule for positional noising
    Proposition 3.3 assumes η_t ~ N(η_1, 1-t); other schedules (variance exploding, flow matching interpolants) are not covered.
invented entities (1)
  • Coarse-grained supernodes
    purpose: Message-passing relay nodes that allow DMP to lower resolution at high noise while keeping outputs on original nodes.
    Introduced in Section 4.1 as a computational device; no external falsifiable prediction is made.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometric Generative Modeling with Noise-Conditioned Graph Networks." pith.science (2026). https://pith.science/paper/KHOMGWJR

@misc{pith2026250709391,
  author       = {Pith},
  title        = {Pith review of: Geometric Generative Modeling with Noise-Conditioned Graph Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHOMGWJR}},
  note         = {Machine review of arXiv:2507.09391}
}
abstract

Generative modeling of graphs with spatial structure is essential across many applications from computer graphics to spatial genomics. Recent flow-based generative models have achieved impressive results by gradually adding and then learning to remove noise from these graphs. Existing models, however, use graph neural network architectures that are independent of the noise level, limiting their expressiveness. To address this issue, we introduce \textit{Noise-Conditioned Graph Networks} (NCGNs), a class of graph neural networks that dynamically modify their architecture according to the noise level during generation. Our theoretical and empirical analysis reveals that as noise increases, (1) graphs require information from increasingly distant neighbors and (2) graphs can be effectively represented at lower resolutions. Based on these insights, we develop Dynamic Message Passing (DMP), a specific instantiation of NCGNs that adapts both the range and resolution of message passing to the noise level. DMP consistently outperforms noise-independent architectures on a variety of domains including $3$D point clouds, spatiotemporal transcriptomics, and images. Code is available at https://github.com/peterpaohuang/ncgn.

Figures

Figures reproduced from arXiv: 2507.09391 by the authors.

Figure 1
Figure 1. Comparing the range and resolution between noise-independent graph neural networks and noise-conditioned graph neural networks (GNNs). (Top Row) Noise is introduced in the positions of the star while features are fixed. (Bottom Row) Noise is introduced in the features of the star while positions are fixed. (Left Column) As noise increases, the range of noise-independent GNNs remains the same while noise-conditional … view at source ↗
Figure 2
Figure 2. Optimal Radius for Different Noise Levels. Using the correlation ρ(η (i) , η(j) ) = 1 − (η (i) − η (j) ) 2 , we plot the radius that maximizes mutual information for each signal-to-noise ratio using the formula from Lemma 3.1. For our first theoretical result, we assume that the positions of the nodes are fixed and we only noise the features. Lemma 3.1. Let the signal-to-noise ratio be SNR(t) = σ 2 1 σ 2 t where σ 2… view at source ↗
Figure 3
Figure 3. Attention Distribution over Node Distances. Average attention weight distribution between pairs of nodes as a function of their spatial distance, shown for different noise levels in a graph attention network. bers of clusters, and then (3) compare each noised and coarse-grained version to its original, unnoised shape. To quantify the similarity between shapes, we adopt the Gromov-Wasserstein (GW) distance from objec… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Adaptive Schedules. The schedule determines the level of coarse-graining and connectivity of the graph neural network corresponding to a noise level. (Top) The number of clusters (aka fewer clusters means higher coarse-graining). The higher the noise, the fewer number …
Figure 6
Figure 6. Figure 6: Example of the spatiotemporal data generated by the reaction-diffusion equation. B.3. Spatiotemporal Transcriptomics Dataset Construction. Spatiotemporal transcriptomic measures the gene expression of many single-cells distributed spatially (e.g. a tissue slice) over m…
Figure 7
Figure 7. Figure 7: Turing Network for Bmp-Sox9-Wnt (Raspopovic et al., 2014). Parameter Value Description αsox9 α ∼ U(−0.01, 0.01) Base production rate of Sox9 protein αbmp α ∼ U(−0.01, 0.01) Base production rate of BMP αwnt α ∼ U(−0.01, 0.01) Base production rate of WNT k2 1 Positive re…
Figure 8
Figure 8. Figure 8: Spatiotemporal Transcriptomic Tasks. In addition to unconditional generation, these are several other useful conditional tasks for single cell transcriptomics. For each task, we delineate what the model is given as conditioning information and what the model needs to g…
Figure 9
Figure 9. Figure 9: Select predictions from (Top Row) DiT and (Bottom Row) DiT-DMP. D. Additional Experiments & Ablations D.1. Empirical Analysis: Resolution Using the same dataset from Section 3.2, we also visualize the average Gromov-Wasserstein (GW) distance between the geometric graph…
Figure 10
Figure 10. Figure 10: Gromov-Wasserstein distance between the original unnoised geometric graph and coarse-grained versions of the geometric graph under different levels of noise (α = 1 − t, meaning higher α → higher noise). Two types of coarse-graining (max and mean pooling) are conducted…
Figure 11
Figure 11. Figure 11: Performance of models on the spatiotemporal transcriptomics dataset across an increasing number of message passing layers. We report the Wasserstein distance for the unconditional generation task for DMP, a kNN graph, and a fully connected graph. We train our method a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 18 canonical work pages

  1. [1]

    J., Bambrick, J., et al

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, pp.\ 1--3, 2024

    2024

  2. [2]

    S., Boffi, N

    Albergo, M. S., Boffi, N. M., and Vanden-Eijnden, E. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023

    arXiv 2023

  3. [3]

    and Yahav, E

    Alon, U. and Yahav, E. On the bottleneck of graph neural networks and its practical implications. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=i80OPhOCVH2

    work page 2021

  4. [4]

    Flow network based generative models for non-iterative diverse candidate generation

    Bengio, E., Jain, M., Korablyov, M., Precup, D., and Bengio, Y. Flow network based generative models for non-iterative diverse candidate generation. Advances in Neural Information Processing Systems, 34: 0 27381--27394, 2021

    2021

  5. [5]

    Flexivit: One model for all patch sizes

    Beyer, L., Izmailov, P., Kolesnikov, A., Caron, M., Kornblith, S., Zhai, X., Minderer, M., Tschannen, M., Alabdulmohsin, I., and Pavetic, F. Flexivit: One model for all patch sizes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 14496--14506, 2023

    work page 2023

  6. [6]

    G., Gut, G., Del Castillo, J

    Bunne, C., Stark, S. G., Gut, G., Del Castillo, J. S., Levesque, M., Lehmann, K.-V., Pelkmans, L., Krause, A., and R \"a tsch, G. Learning single-cell perturbation responses using neural optimal transport. Nature methods, 20 0 (11): 0 1759--1768, 2023

    work page 2023

  7. [7]

    Diffdock: Diffusion steps, twists, and turns for molecular docking

    Corso, G., St \"a rk, H., Jing, B., Barzilay, R., and Jaakkola, T. Diffdock: Diffusion steps, twists, and turns for molecular docking. arXiv preprint arXiv:2210.01776, 2022

    arXiv 2022

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp.\ 248--255. Ieee, 2009

    2009

Show all 41 references
  1. [9]

    and Lenssen, J

    Fey, M. and Lenssen, J. E. Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428, 2019

    1903 arXiv

  2. [10]

    N., Duvenaud, D., Hern \'a ndez-Lobato, J

    G \'o mez-Bombarelli, R., Wei, J. N., Duvenaud, D., Hern \'a ndez-Lobato, J. M., S \'a nchez-Lengeling, B., Sheberla, D., Aguilera-Iparraguirre, J., Hirzel, T. D., Adams, R. P., and Aspuru-Guzik, A. Automatic chemical design using a data-driven continuous representation of mol...

    2018

  3. [11]

    Neighborhood attention transformer

    Hassani, A., Walton, S., Li, J., Li, S., and Shi, H. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 6185--6194, 2023

    2023

  4. [12]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  5. [13]

    Torsional diffusion for molecular conformer generation

    Jing, B., Corso, G., Chang, J., Barzilay, R., and Jaakkola, T. Torsional diffusion for molecular conformer generation. Advances in Neural Information Processing Systems, 35: 0 24240--24253, 2022

    2022

  6. [14]

    Eigenfold: Generative protein structure prediction with diffusion models

    Jing, B., Erives, E., Pao-Huang, P., Corso, G., Berger, B., and Jaakkola, T. Eigenfold: Generative protein structure prediction with diffusion models. arXiv preprint arXiv:2304.02198, 2023

    2023 arXiv

  7. [15]

    Highly accurate protein structure prediction with alphafold

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Z \' dek, A., Potapenko, A., et al. Highly accurate protein structure prediction with alphafold. nature, 596 0 (7873): 0 583--589, 2021

    2021

  8. [16]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

    2016 arXiv

  9. [17]

    Neural operator: Graph kernel network for partial differential equations

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A. Neural operator: Graph kernel network for partial differential equations. arXiv preprint arXiv:2003.03485, 2020

    2003 arXiv

  10. [18]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  11. [19]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  12. [20]

    Predicting molecular conformation via dynamic graph score matching

    Luo, S., Shi, C., Xu, M., and Tang, J. Predicting molecular conformation via dynamic graph score matching. Advances in Neural Information Processing Systems, 34: 0 19784--19795, 2021

    2021

  13. [21]

    Antigen-specific antibody design and optimization with diffusion-based generative models for protein structures

    Luo, S., Su, Y., Peng, X., Wang, S., Peng, J., and Ma, J. Antigen-specific antibody design and optimization with diffusion-based generative models for protein structures. Advances in Neural Information Processing Systems, 35: 0 9754--9767, 2022

    2022

  14. [22]

    Gromov--wasserstein distances and the metric approach to object matching

    M \'e moli, F. Gromov--wasserstein distances and the metric approach to object matching. Foundations of computational mathematics, 11: 0 417--487, 2011

    2011

  15. [23]

    A., and Battaglia, P

    Nash, C., Ganin, Y., Eslami, S. A., and Battaglia, P. Polygen: An autoregressive generative model of 3d meshes. In International conference on machine learning, pp.\ 7220--7229. PMLR, 2020

    2020

  16. [24]

    Action matching: Learning stochastic dynamics from samples

    Neklyudov, K., Brekelmans, R., Severo, D., and Makhzani, A. Action matching: Learning stochastic dynamics from samples. In International conference on machine learning, pp.\ 25858--25889. PMLR, 2023

    2023

  17. [25]

    and Suzuki, T

    Oono, K. and Suzuki, T. Graph neural networks exponentially lose expressive power for node classification. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=S1ldO2EFPr

    2020

  18. [26]

    Scalable multimer structure prediction using diffusion models

    Pao-Huang, P., Jing, B., and Berger, B. Scalable multimer structure prediction using diffusion models. In NeurIPS 2023 AI for Science Workshop, 2023

    2023

  19. [27]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4195--4205, 2023

    2023

  20. [28]

    stvcr: Reconstructing spatio-temporal dynamics of cell development using optimal transport

    Peng, Q., Zhou, P., and Li, T. stvcr: Reconstructing spatio-temporal dynamics of cell development using optimal transport. bioRxiv, pp.\ 2024--06, 2024

    2024

  21. [29]

    D., Weng, C., Hosseinzadeh, S., Yang, D., Pogson, A

    Qiu, X., Zhang, Y., Martin-Rufino, J. D., Weng, C., Hosseinzadeh, S., Yang, D., Pogson, A. N., Hein, M. Y., Min, K. H. J., Wang, L., et al. Mapping transcriptomic vector fields of single cells. Cell, 185 0 (4): 0 690--711, 2022

    2022

  22. [30]

    Y., Lu, Y., Yao, J., Jing, Z., Min, K

    Qiu, X., Zhu, D. Y., Lu, Y., Yao, J., Jing, Z., Min, K. H., Cheng, M., Pan, H., Zuo, L., King, S., et al. Spatiotemporal modeling of molecular holograms. Cell, 2024

    2024

  23. [31]

    Digit patterning is controlled by a bmp-sox9-wnt turing network modulated by morphogen gradients

    Raspopovic, J., Marcon, L., Russo, L., and Sharpe, J. Digit patterning is controlled by a bmp-sox9-wnt turing network modulated by morphogen gradients. Science, 345 0 (6196): 0 566--570, 2014

    2014

  24. [32]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  25. [33]

    Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics

    Tong, A., Huang, J., Wolf, G., Van Dijk, D., and Krishnaswamy, S. Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics. In International conference on machine learning, pp.\ 9526--9536. PMLR, 2020

    2020

  26. [34]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G., and Bengio, Y. Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482, 2023

    2023 arXiv

  27. [35]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

  28. [36]

    L., Juergens, D., Bennett, N

    Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., et al. De novo design of protein structure and function with rfdiffusion. Nature, 620 0 (7976): 0 1089--1100, 2023

    2023

  29. [37]

    Demystifying oversmoothing in attention-based graph neural networks

    Wu, X., Ajorlou, A., Wu, Z., and Jadbabaie, A. Demystifying oversmoothing in attention-based graph neural networks. Advances in Neural Information Processing Systems, 36, 2024

    2024

  30. [38]

    3d shapenets: A deep representation for volumetric shapes

    Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., and Xiao, J. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1912--1920, 2015

    1912

  31. [39]

    Geodiff: A geometric diffusion model for molecular conformation generation

    Xu, M., Yu, L., Song, Y., Shi, C., Ermon, S., and Tang, J. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923, 2022

    2022 arXiv

  32. [40]

    Pointflow: 3d point cloud generation with continuous normalizing flows

    Yang, G., Huang, X., Hao, Z., Liu, M.-Y., Belongie, S., and Hariharan, B. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4541--4550, 2019

    2019

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.