Pith. sign in

REVIEW 5 major objections 6 minor 33 references

KITINet: Kinetics Theory Inspired Network Architectures with PDE Simulation Approaches

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing residual connections with a stochastic particle-collision simulation during training improves accuracy on image, text, and PDE operator tasks while leaving the inference architecture unchanged.

desk verdict A genuinely new training-time collision rule for residual connections, but the 'architecture' claim oversells it: inference is identical to a baseline residual net, and no dropout/noise control separates the effect from generic stochastic regularization. read the letter →

arxiv 2505.17919 v1 pith:B2YLPKKB submitted 2025-05-23 cs.LG

Mingquan Feng , Yifan Fu , Tongcheng Zhang , Yu Jiang , Yixin Huang , Junchi Yan This is my paper · ORCID
classification cs.LG
keywords KITINetresidualconnectionsBoltzmanntransportequationDirectSimulationMonteCarloparticlecollisionsparametercondensationPDEoperatorlearningphysics-inspiredneuralnetworks
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

KITINet proposes a residual module that, during training only, treats feature channels as particles in a low-dimensional space and simulates their collisions using a rule adapted from the Boltzmann transport equation and the Direct Simulation Monte Carlo method. At inference the module is bypassed, so the deployed network is identical to the original ResNet, BERT, FNO, or OFormer. The paper reports consistent accuracy gains on CIFAR-10/100, IMDb and SNLI, and lower error on several PDE operator-learning benchmarks, with negligible added FLOPs, and links the gains to accelerated parameter condensation during training.

What carries the argument

The central object is the KITINet residual module, a DSMC-inspired collision operator that replaces the residual connection during training. It computes pairwise relative position and velocity, center-of-mass position and velocity, and an exponential mean-free-path factor Ur = exp(-Xr); a collision is accepted when the scaled relative velocity exceeds 1 - coll_coef. Accepted collisions update velocities via the hard-sphere rule and update positions by averaging the collision positions with the initial position, then advancing by dt times the new velocity. At test time the module is bypassed, returning x + dt*v, which makes the collision dynamics a training-only perturbation of standard residual learning.

What would settle it

A matched-noise control experiment: take a ResNet baseline and an FNO baseline and, during training only, inject stochastic pairwise feature mixing with the same acceptance probability and update magnitude as KITINet's collisions. If this control reproduces KITINet's accuracy and condensation curves on CIFAR-10 and the Burgers or heat operator tasks, then the specific collision structure is not the cause of the gains; if it does not, the geometry of the collision rule is doing genuine work.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that residual learning can be reformulated as a kinetic particle system. KITINet reshapes the feature vector and residual into n_divide by N matrices, treating each row as a particle that moves and collides. The update estimates pairwise distances and velocities, accepts collisions between pairs with a probability governed by an exponential mean-free-path factor and a coll_coef threshold, and applies a hard-sphere center-of-mass velocity exchange together with a position-averaging update. Because the module returns x plus dt times the residual at inference, the trained model is architecturally identical to the baseline, so the measured gains are attributed entirely to training dynamics, which the paper connects to network parameter condensation.

Load-bearing premise

The entire reported benefit rests on the assumption that the collision dynamics themselves, rather than the extra randomness they inject, drive the test-set gains, because the module is removed at inference.

Editorial extensions

If this is right

  • Because the module is disabled at inference, any measured benefit comes from the training trajectory rather than added capacity, so adopting KITINet adds no inference-time cost.
  • The paper reports gains across ResNet variants on CIFAR, BERT on IMDb and SNLI, and FNO and OFormer on PDE benchmarks, indicating the mechanism is architecture-agnostic.
  • The explicit position-update term is load-bearing: ablating it consistently increases error on the Burgers, Navier-Stokes, and heat equations.
  • The two hyperparameters n_divide and coll_coef have a large effect on performance, so the reported gains depend on per-task tuning.
  • The module accelerates parameter condensation in synthetic fully connected and skip-connected networks, linking the collision dynamics to a known signature of generalization.

Reading between the lines

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

  • Editorial: A direct next test would replace KITINet's collision pair selection with isotropic feature dropout or random pair mixing matched for acceptance probability and update magnitude; if the gains survive, the specific collision geometry is not the active ingredient.
  • Editorial: Since the inference graph is identical to the baseline, KITINet is effectively a training schedule for standard architectures, which suggests it could be adopted cheaply as a drop-in regularizer for any residual network.
  • Editorial: The exponential mean-free-path factor makes collisions more likely between nearby features, so KITINet is actively mixing similar representations; measuring whether this structured mixing, rather than generic noise, drives condensation would separate mechanism from regularization.
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.

Referee Report

5 major / 6 minor

Summary. The paper proposes KITINet, a residual module that treats feature maps as particle positions and residuals as velocities, and simulates their evolution during training with a discretized, DSMC-inspired collision rule derived from the Boltzmann transport equation. At inference, Algorithm 1 bypasses the module entirely and returns x + dt*v, i.e. the standard residual update. The paper reports improved accuracy or error over ResNet, FNO, OFormer, and BERT baselines on CIFAR-10/100, Burgers'/NS/Heat PDE operator learning, the Airfoil problem, IMDb, and SNLI, and claims that the module induces network parameter condensation during training.

Significance. If the reported gains were robust and attributable to the proposed mechanism, the result would be significant: it would show that a physics-inspired, training-time stochastic collision operator can improve the generalization of standard architectures with negligible FLOPs overhead, while also linking the training dynamics to parameter condensation. The paper is broad in scope and includes several useful ablations, including the effect of position updates and hyperparameter sensitivity. However, the central claim is not established by the current evidence: because inference is identical to the baseline, all measured improvements must come from training-time stochasticity, yet no control separates the collision rule from generic stochastic regularization. The condensation analysis is descriptive and does not establish causality. The paper would need a matched control arm, multiple seeds with error bars, and a clear statement of the hyperparameter values used for every reported result before the claims can be accepted.

major comments (5)
  1. [Section 4, Algorithm 1] Line 3 of Algorithm 1 returns x + dt*v at inference, which is exactly the standard residual update. The KITINet module is therefore not present in the deployed model, and all reported gains must arise from training-time dynamics alone. Yet the paper repeatedly frames the contribution as an architectural improvement with 'adaptive feature refinement via physics-informed interactions.' This is a load-bearing mismatch: the experiments cannot distinguish the proposed collision rule from any stochastic perturbation applied during training. A matched control, such as applying dropout to the residual path, adding Gaussian noise to v, or randomly selecting pairs without using positions and velocities, is required to support the claim that the kinetics-inspired mechanism, rather than generic stochastic regularization, is responsible for the improvements.
  2. [Section 5, Tables 1-3] No error bars, no multiple seeds, and no significance tests are reported for any of the main results. Many gains are small in absolute terms (for example, ResNet-34 CIFAR-10 improves from 94.48% to 95.04%, and the NS equation MSE improves from 0.12023 to 0.11346), so without variance estimates the improvements could easily be within training noise. The paper should report mean and standard deviation over at least three to five independent runs and, where appropriate, a paired significance test against the baseline.
  3. [Section 5.5, Figure 3] The two method-specific hyperparameters n_divide and coll_coef strongly influence performance, and Figure 3 shows that the best setting varies by task. The specific values used for each result in Tables 1-3 are not reported, and the sensitivity curves resemble tuning a regularization strength. This is especially problematic because the method is only active during training: the reported improvements could be the result of selecting a favorable noise schedule. The authors should report the chosen hyperparameter values for every experiment and include a sensitivity analysis that compares KITINet with a generic stochastic regularizer at matched computational cost.
  4. [Section 4, Equation (10)] Equation (10) is presented as the collision acceptance rule, but it contains no random draw: the condition (Vr)_{i,j} * (Ur)_{i,j} / vmax_r > 1 - coll_coef is deterministic given the current positions and velocities. This differs qualitatively from the DSMC rejection sampling in Equation (4), which explicitly uses a uniform random number. The only stochastic element in the module is the random unit vector in Equation (9). The paper should either revise the description of the mechanism as 'stochastic' or clarify the precise role of randomness in the collision rule, because the claimed connection to DSMC/BTE is part of the core motivation.
  5. [Section 5.6] The parameter condensation analysis is observational and does not establish that condensation is the cause of the reported test-set improvements. The synthetic experiments show that KITINet accelerates condensation, but they do not show that this acceleration is specific to the collision rule, nor that condensation is the mediating factor behind the gains in Tables 1-3. Since other stochastic training mechanisms can also induce clustering or implicit regularization, a controlled comparison is needed before condensation can be presented as the explanatory mechanism.
minor comments (6)
  1. [Section 6] The phrase 'KITINetic theory' in the first sentence of the conclusion appears to be a typo and should read 'kinetic theory.'
  2. [Appendix C] The heading 'PDE Datsets Generation' contains a typo; it should be 'PDE Datasets Generation.'
  3. [Section 5.1] The text 'OUr datasets generation method' contains a typo and should read 'Our datasets generation method.'
  4. [Tables 2 and 3] The paper should state explicitly which n_divide and coll_coef values were used for each model and dataset, since the sensitivity analysis in Section 5.5 shows that these choices materially affect results.
  5. [Appendix B] The 'a-edition KITINet' variant is described only in the appendix and is reported to have failed; a brief mention in the main text would help readers understand why the presented formulation was chosen.
  6. [General] No code or reproducibility statement is provided; releasing the implementation and configuration files would substantially strengthen the paper, especially given the stochastic nature of the training procedure.

Circularity Check

1 steps flagged · score 5.0 of 10

KITINet's inference path is the baseline residual by construction, so the claimed architectural gains reduce to training-time stochasticity; no control isolates the physics mechanism.

  1. self definitional [Algorithm 1, step 3; Section 5.4 (Ablation Study)]
    "3: If model is not in the training phase, Return x+dt∗v ... However, to ensure that KITINet can be reduced to a ResNet-like architecture, we set dt=1, making the change in position during a time step non-negligible."

    Algorithm 1 makes the inference-time forward pass exactly the standard residual connection x+dt·v, and Section 5.4 states that dt=1 was chosen specifically so KITINet reduces to a ResNet-like architecture. Thus the deployed model is, by construction, the baseline residual block; the physics module is absent at test time. All reported test improvements must therefore come from the training-time stochastic collision schedule, not from any architectural difference at inference. The paper nevertheless frames these gains as architectural improvements produced by 'physics-informed interactions' and explains them via parameter condensation, without an ablation that replaces the collision rule by an equally cheap stochastic perturbation.

full rationale

This is not a case of fabricated benchmark numbers or a self-citation chain: the experiments are run against external baselines and the condensation literature cited (Xu et al.) is external to the present author set. The central circularity is narrower and definitional. KITINet's inference behaviour is deliberately forced to equal the baseline residual connection (Algorithm 1 returns x+dt*v and dt=1 is set 'to ensure that KITINet can be reduced to a ResNet-like architecture'), so any test-time advantage of 'KITINet' over the baseline cannot be an architectural advantage. The reported gains are training-time effects of a stochastic collision operator that is never isolated from generic stochastic regularization through a matched control. The condensation observation in Section 5.6 is a post-hoc correlate and is used to explain the same runs from which it is measured, which does not establish causality. These issues make the strongest claim partly circular by construction, but do not make the whole empirical study vacuous, so a moderate score is appropriate.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on an analogy between network features and gas particles. The two tunable hyperparameters n_divide and coll_coef are fitted per task and are not reported for the main results. The physical derivation is heuristic, and the inference-time reduction to a standard residual connection is an explicit design choice.

free parameters (4)
  • n_divide = not reported, tuned per task (range 1-210)
    Number of spatial dimensions per particle; controls particle grouping. Strong effect on FNO performance (Figure 3a), tuned per task.
  • coll_coef = not reported, tuned per task (range 0.1-0.9)
    Collision acceptance threshold; best value varies greatly by task (Figures 3b and 7).
  • dt = 1
    Time step; set to 1 so that inference reduces to the standard residual connection x+v (Section 5.4).
  • module placement = last stage of ResNet, final BERT layer, each FNO Fourier layer
    The module is applied selectively; the choice is not justified and affects results.
assumptions (5)
  • ad hoc to paper The feature vector can be partitioned into N particles in n_divide-dimensional space, where rows of X and V are positions and velocities.
    This is the central modeling analogy; no justification beyond the DSMC metaphor (Section 4).
  • ad hoc to paper Collision acceptance in Eq (10) with e^{-|x_i-x_j|} approximates a mean free path and the combined steps approximate BTE/DSMC.
    The paper adapts DSMC equations heuristically; no convergence or approximation error analysis is provided (Section 4, step 3).
  • domain assumption Parameter condensation, measured by cosine similarity, is a valid indicator of generalization.
    Imported from cited prior work [31,32,33]; the paper does not test this assumption in its own setting (Section 3.3).
  • ad hoc to paper A training-time stochastic collision rule can improve test-time generalization even though inference uses the vanilla residual path.
    The paper provides no theoretical argument; this is the load-bearing empirical assumption of all experiments (Algorithm 1).
  • ad hoc to paper The hard-sphere collision model conserves momentum and energy and is appropriate for feature vectors.
    The physical conservation laws are invoked for network features without justification (Sections 3.2 and 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of KITINet: Kinetics Theory Inspired Network Architectures with PDE Simulation Approaches." pith.science (2026). https://pith.science/paper/B2YLPKKB

@misc{pith2026250517919,
  author       = {Pith},
  title        = {Pith review of: KITINet: Kinetics Theory Inspired Network Architectures with PDE Simulation Approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2YLPKKB}},
  note         = {Machine review of arXiv:2505.17919}
}
read the original abstract

Despite the widely recognized success of residual connections in modern neural networks, their design principles remain largely heuristic. This paper introduces KITINet (Kinetics Theory Inspired Network), a novel architecture that reinterprets feature propagation through the lens of non-equilibrium particle dynamics and partial differential equation (PDE) simulation. At its core, we propose a residual module that models feature updates as the stochastic evolution of a particle system, numerically simulated via a discretized solver for the Boltzmann transport equation (BTE). This formulation mimics particle collisions and energy exchange, enabling adaptive feature refinement via physics-informed interactions. Additionally, we reveal that this mechanism induces network parameter condensation during training, where parameters progressively concentrate into a sparse subset of dominant channels. Experiments on scientific computation (PDE operator), image classification (CIFAR-10/100), and text classification (IMDb/SNLI) show consistent improvements over classic network baselines, with negligible increase of FLOPs.

Figures

Figures reproduced from arXiv: 2505.17919 by the authors.

Figure 1
Figure 1. Overview of the proposed architecture: KITINet. It modifies the residual connection by [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FNO’s performance on NS equation, both vanilla and with KITINet applied. Left two: [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The performance of KITINet-FNO with different hyper-parameters [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Results of parameter condensation across network configurations on synthetic data. (a) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Evolution of parameter condensation on Three-layer Fully-connected Network. (Row 1) [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The full architecture of Fourier Neural Operators (FNO) with and without KITINet applied. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: The performance of KITINet-FNO with different hyper-parameter [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Evolution of parameter condensation effect on Six-layer ReLU skip-connected network [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Evolution of parameter condensation effect on Six-layer ReLU skip-connected network [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Evolution of parameter condensation on Six-layer skip-connected Network applying [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 23 canonical work pages

  1. [1]

    Generalized scheme of the no-time-counter scheme for the dsmc in rarefied gas flow analysis.Computers & fluids, 22(2-3):253–257, 1993

    Takashi Abe. Generalized scheme of the no-time-counter scheme for the dsmc in rarefied gas flow analysis.Computers & fluids, 22(2-3):253–257, 1993

    work page 1993

  2. [2]

    Neural flows: Efficient alternative to neural odes.Advances in neural information processing systems, 34:21325–21337, 2021

    Marin Biloš, Johanna Sommer, Syama Sundar Rangapuram, Tim Januschowski, and Stephan Günnemann. Neural flows: Efficient alternative to neural odes.Advances in neural information processing systems, 34:21325–21337, 2021

    2021

  3. [3]

    Approach to translational equilibrium in a rigid sphere gas.Phys

    GA Bird. Approach to translational equilibrium in a rigid sphere gas.Phys. fluids, 6:1518–1519, 1963

    work page 1963

  4. [4]

    Ludwig Boltzmann. On the relationship between the second fundamental theorem of the mechanical theory of heat and probability calculations regarding the conditions for thermal equilibrium.Entropy, 17(4):1971–2009, 2015

    work page 1971

  5. [5]

    Bowman, Gabor Angeli, Christopher Potts, and Christopher D

    Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Lluís Màrquez, Chris Callison- Burch, and Jian Su, editors,Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal, September 2015. Association f...

    work page 2015

  6. [6]

    Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    2018

  7. [7]

    Efficient and flexible method for reducing moderate-size deep neural networks with condensation.Entropy, 26(7):567, 2024

    Tianyi Chen and Zhi-Qin John Xu. Efficient and flexible method for reducing moderate-size deep neural networks with condensation.Entropy, 26(7):567, 2024

    work page 2024

  8. [8]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

Show all 33 references
  1. [9]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InNorth American Chapter of the Association for Computational Linguistics, 2019

    2019

  2. [10]

    Economon, Francisco Palacios, Sean R

    Thomas D. Economon, Francisco Palacios, Sean R. Copeland, Trent W. Lukaczyk, and Juan J. Alonso. SU2: An Open-Source Suite for Multiphysics Simulation and Design.AIAA Journal, 54(3):828–846, December 2015

    2015

  3. [11]

    Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations.Advances in neural information processing systems, 34:3836–3849, 2021

    Moshe Eliasof, Eldad Haber, and Eran Treister. Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations.Advances in neural information processing systems, 34:3836–3849, 2021

    2021

  4. [12]

    Hamiltonian neural networks.Advances in neural information processing systems, 32, 2019

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks.Advances in neural information processing systems, 32, 2019

    2019

  5. [13]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  6. [14]

    Molecular dynamics simulation for all.Neuron, 99(6):1129–1143, 2018

    Scott A Hollingsworth and Ron O Dror. Molecular dynamics simulation for all.Neuron, 99(6):1129–1143, 2018

    2018

  7. [15]

    Neural controlled differential equations for irregular time series.Advances in neural information processing systems, 33:6696– 6707, 2020

    Patrick Kidger, James Morrill, James Foster, and Terry Lyons. Neural controlled differential equations for irregular time series.Advances in neural information processing systems, 33:6696– 6707, 2020

    2020

  8. [16]

    Transformer for partial differential equations’ operator learning, 2023

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning, 2023

    2023

  9. [17]

    Fourier neural operator for parametric partial differen- tial equations, 2021

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differen- tial equations, 2021. 10

    2021

  10. [18]

    Dendritic integration inspired artificial neural networks capture data correlation.Advances in Neural Information Processing Systems, 37:79325–79349, 2024

    Chongming Liu, Jingyang Ma, Songting Li, and Douglas Dongzhuo Zhou. Dendritic integration inspired artificial neural networks capture data correlation.Advances in Neural Information Processing Systems, 37:79325–79349, 2024

    2024

  11. [19]

    Courier Corporation, 2004

    Leonard B Loeb.The kinetic theory of gases. Courier Corporation, 2004

    2004

  12. [20]

    Pde-net 2.0: Learning pdes from data with a numeric- symbolic hybrid deep network.Journal of Computational Physics, 399:108925, 2019

    Zichao Long, Yiping Lu, and Bin Dong. Pde-net 2.0: Learning pdes from data with a numeric- symbolic hybrid deep network.Journal of Computational Physics, 399:108925, 2019

    2019

  13. [21]

    Pde-net: Learning pdes from data

    Zichao Long, Yiping Lu, Xianzhong Ma, and Bin Dong. Pde-net: Learning pdes from data. In International conference on machine learning, pages 3208–3216. PMLR, 2018

    2018

  14. [22]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y . Ng, and Christopher Potts. Learning word vectors for sentiment analysis. InProceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150,...

    2011

  15. [23]

    Neural ode processes

    A Norcliffe, C Bodnar, B Day, J Moss, P Lio, et al. Neural ode processes. InICLR 2021-9th International Conference on Learning Representations. International Conference on Learning Representations, ICLR, 2021

    2021

  16. [24]

    On second order behaviour in augmented neural odes.Advances in neural information processing systems, 33:5911–5921, 2020

    Alexander Norcliffe, Cristian Bodnar, Ben Day, Nikola Simidjievski, and Pietro Liò. On second order behaviour in augmented neural odes.Advances in neural information processing systems, 33:5911–5921, 2020

    2020

  17. [25]

    Battaglia

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh-based simulation with graph networks, 2021

    2021

  18. [26]

    Latent ordinary differential equations for irregularly-sampled time series.Advances in neural information processing systems, 32, 2019

    Yulia Rubanova, Ricky TQ Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series.Advances in neural information processing systems, 32, 2019

    2019

  19. [27]

    Combinatorial optimization with physics-inspired graph neural networks.Nature Machine Intelligence, 4(4):367–377, 2022

    Martin JA Schuetz, J Kyle Brubaker, and Helmut G Katzgraber. Combinatorial optimization with physics-inspired graph neural networks.Nature Machine Intelligence, 4(4):367–377, 2022

    2022

  20. [28]

    Hamiltonian generative networks.arXiv preprint arXiv:1909.13789, 2019

    Peter Toth, Danilo Jimenez Rezende, Andrew Jaegle, Sébastien Racanière, Aleksandar Botev, and Irina Higgins. Hamiltonian generative networks.arXiv preprint arXiv:1909.13789, 2019

    1909 arXiv

  21. [29]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  22. [30]

    Convection-diffusion equation: a theoret- ically certified framework for neural networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Tangjun Wang, Chenglong Bao, and Zuoqiang Shi. Convection-diffusion equation: a theoret- ically certified framework for neural networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  23. [31]

    An overview of condensation phe- nomenon in deep learning.arXiv preprint arXiv:2504.09484, 2025

    Zhi-Qin John Xu, Yaoyu Zhang, and Zhangchen Zhou. An overview of condensation phe- nomenon in deep learning.arXiv preprint arXiv:2504.09484, 2025

    2025 arXiv

  24. [32]

    Embedding principle of loss landscape of deep neural networks.Advances in Neural Information Processing Systems, 34:14848–14859, 2021

    Yaoyu Zhang, Zhongwang Zhang, Tao Luo, and Zhiqin J Xu. Embedding principle of loss landscape of deep neural networks.Advances in Neural Information Processing Systems, 34:14848–14859, 2021

    2021

  25. [33]

    Towards understanding the condensation of neural networks at initial training.Advances in Neural Information Processing Systems, 35:2184–2196, 2022

    Hanxu Zhou, Zhou Qixuan, Tao Luo, Yaoyu Zhang, and Zhi-Qin Xu. Towards understanding the condensation of neural networks at initial training.Advances in Neural Information Processing Systems, 35:2184–2196, 2022. 11 A Molecule Dynamics and NeuralODE Molecular dynamics (MD) [14]...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.