pith. sign in

arxiv: 2411.18954 · v4 · submitted 2024-11-28 · 💻 cs.LG · cs.AI

ReMAP: Neural Reparameterization for Scalable MAP Inference in Arbitrary-Order Markov Random Fields

Yaomin Wang , Chaolong Ying , Xiaodong Luo , Tianshu Yu This is my paper

Pith reviewed 2026-05-23 17:22 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords MAP inferenceMarkov Random Fieldsenergy minimizationgraph neural networksreparameterizationrelaxed optimizationscalable inference
0
0 comments X

The pith

ReMAP optimizes a differentiable relaxation of arbitrary-order MRF energies via per-instance graph neural networks to recover high-quality MAP assignments.

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

The paper establishes that treating each MRF as an independent optimization problem, where a graph neural network outputs continuous node-wise label distributions and gradient descent minimizes the relaxed energy, yields scalable inference without any supervised training data. This matters because traditional message-passing approximations lose quality on dense or high-order factors while exact solvers such as Toulbar2 become prohibitive at large scale. The approach directly supports pairwise and higher-order factors, heterogeneous label sets, and GPU execution. Empirical results on synthetic instances, UAI 2022 benchmarks, and real PCI problems show consistent outperformance of approximate methods and, on hard cases, lower energies than exact solvers within practical time limits. The authors prove consistency of the relaxation with the discrete objective and argue that the neural over-parameterization opens optimization trajectories unavailable in the original discrete space.

Core claim

ReMAP is an instance-wise neural reparameterization framework that directly optimizes a differentiable relaxation of the original MRF energy. A Graph Neural Network produces node-wise label distributions, and gradient-based optimization searches for a low-energy discrete solution in an over-parameterized continuous space. The method supports pairwise and arbitrary-order factors, heterogeneous label cardinalities, and efficient GPU execution, without requiring labeled solutions. The relaxed objective is consistent with the discrete MAP problem, and neural over-parameterization exposes low-energy optimization paths unavailable in the original discrete space.

What carries the argument

Instance-wise graph neural network that produces node-wise continuous label distributions for gradient descent on a relaxed MRF energy function.

If this is right

  • ReMAP scales to arbitrary-order factors and heterogeneous label cardinalities on GPU without needing labeled training data.
  • On UAI 2022 inference benchmarks and real PCI problems, the method produces lower energies than standard approximate message-passing baselines.
  • For hard large-scale synthetic and real instances, ReMAP recovers lower-energy assignments than Toulbar2 within practical time budgets.
  • The same per-instance optimization procedure works for both pairwise and high-order MRFs without modification to the core machinery.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same per-instance neural reparameterization idea could be applied to other discrete energy minimization tasks such as weighted constraint satisfaction problems.
  • Hybrid pipelines that warm-start exact solvers with ReMAP solutions might reduce overall runtime on borderline instances.
  • Because each MRF is optimized independently, the method naturally supports online or streaming settings where new factor graphs arrive sequentially.

Load-bearing premise

The continuous relaxation of the MRF energy remains faithful to the original discrete MAP objective after the neural reparameterization.

What would settle it

A large-scale benchmark instance where ReMAP returns a solution whose energy is strictly higher than the energy returned by Toulbar2 when both are run to the same practical time limit.

Figures

Figures reproduced from arXiv: 2411.18954 by Chaolong Ying, Tianshu Yu, Xiaodong Luo, Yaomin Wang.

Figure 1
Figure 1. Figure 1: An overview of NEUROLIFTING. The energy function of this problem is E(X) = θC1 (x1, x2, x3) + θC2 (x3, x4, x5) + θC3 (x2, x3, x5, x6). H (K) T is the output of the model after the T-th iteration. of this function (depicted in the leftmost shaded diagram) undergoes a transformation to a graph￾based perspective, which subsequently integrates into the network architecture. To address the absence of inherent n… view at source ↗
Figure 2
Figure 2. Figure 2: This illustrates the padding procedure for unary loss terms [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The loss curves of the Segmentation_14, P_potts_6 and P_potts_8 from pairwise potts [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The landscape of instance Segmentation_19. From top to the bottom, each [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The average loss curves over UAI inference competition 2022 pairwise cases, PCI instances [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The average loss curves over UAI inference competition 2022 pairwise cases using different [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The training loss of instance Segmentation_19 after convergence of using network layer [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
read the original abstract

Scalable high-quality MAP inference in arbitrary-order Markov Random Fields (MRFs) remains challenging. Approximate message-passing methods are often efficient but can degrade on dense or high-order instances, while exact solvers such as Toulbar2 become increasingly expensive at scale. We present ReMAP, an instance-wise neural reparameterization framework that directly optimizes a differentiable relaxation of the original MRF energy. Instead of relying on supervised labels or amortized training, ReMAP treats each MRF as an independent optimization problem: a Graph Neural Network produces node-wise label distributions, and gradient-based optimization searches for a low-energy discrete solution in an over-parameterized continuous space. The method supports pairwise and arbitrary-order factors, heterogeneous label cardinalities, and efficient GPU execution, without requiring labeled solutions. We show that the relaxed objective is consistent with the discrete MAP problem and analyze how neural over-parameterization can expose low-energy optimization paths unavailable in the original discrete space. Empirically, on synthetic pairwise and high-order MRFs, UAI 2022 inference benchmarks, and real-world Physical Cell Identity (PCI) problems, ReMAP consistently outperforms approximate baselines and often finds lower-energy solutions than Toulbar2 on hard large-scale instances within practical time budgets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces ReMAP, an instance-wise neural reparameterization method for MAP inference in arbitrary-order MRFs. A GNN produces node-wise label distributions that are used to optimize a differentiable relaxation of the MRF energy via gradient descent; the approach requires no supervised training, supports heterogeneous cardinalities and high-order factors, and is executed on GPU. The manuscript claims that the relaxed objective is consistent with the original discrete MAP problem and that neural over-parameterization exposes lower-energy paths; empirical results on synthetic MRFs, UAI 2022 benchmarks, and real-world PCI problems show consistent outperformance of approximate baselines and, on some hard large instances, lower energies than Toulbar2 within practical time limits.

Significance. If the consistency claim holds and the empirical gains are reproducible, ReMAP would supply a scalable, label-free alternative for high-order MRF inference problems that currently force a trade-off between speed and solution quality. The instance-specific formulation and GPU compatibility are practical strengths for applications where amortized or supervised methods are unavailable.

major comments (2)
  1. [Abstract / consistency claim] Abstract and the consistency statement: the claim that the relaxed objective is consistent with the discrete MAP problem is load-bearing for all performance assertions, yet the construction (GNN node distributions reparameterizing arbitrary-order factors) leaves open whether the continuous minimum is attained only at vertices corresponding to valid discrete assignments. A formal argument or explicit verification that the lifting preserves the exact min (rather than allowing underestimation via convex combinations) is required.
  2. [Experiments / Toulbar2 comparison] Empirical comparisons with Toulbar2: the strongest claim (lower-energy solutions on hard large-scale instances) rests on the assumption that reported energies correspond to feasible discrete assignments. The manuscript must specify how candidate solutions are discretized, whether the final energies are re-evaluated exactly on the original MRF, and the precise time budgets and instance sizes used in the head-to-head tables.
minor comments (2)
  1. [Method] Notation for the reparameterization mapping from GNN outputs to factor potentials should be introduced with an explicit equation early in the method section.
  2. [Implementation] The description of GPU execution and memory scaling for high-order factors would benefit from a complexity table or pseudocode.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the paper accordingly to improve clarity and completeness.

read point-by-point responses

  1. Referee: [Abstract / consistency claim] Abstract and the consistency statement: the claim that the relaxed objective is consistent with the discrete MAP problem is load-bearing for all performance assertions, yet the construction (GNN node distributions reparameterizing arbitrary-order factors) leaves open whether the continuous minimum is attained only at vertices corresponding to valid discrete assignments. A formal argument or explicit verification that the lifting preserves the exact min (rather than allowing underestimation via convex combinations) is required.

    Authors: We appreciate the referee drawing attention to the consistency claim. Section 3.2 of the manuscript presents a formal argument establishing that the relaxed objective is consistent with the discrete MAP problem, with minima attained at valid discrete assignments due to the reparameterization construction for arbitrary-order factors. To further address potential concerns regarding underestimation through convex combinations, we will add an explicit lemma with a short proof sketch in the revised version confirming that no lower energy is possible outside the vertices. revision: yes

  2. Referee: [Experiments / Toulbar2 comparison] Empirical comparisons with Toulbar2: the strongest claim (lower-energy solutions on hard large-scale instances) rests on the assumption that reported energies correspond to feasible discrete assignments. The manuscript must specify how candidate solutions are discretized, whether the final energies are re-evaluated exactly on the original MRF, and the precise time budgets and instance sizes used in the head-to-head tables.

    Authors: We agree that these experimental details require explicit specification for full reproducibility. Candidate solutions are discretized by taking the argmax over the GNN-produced label distributions at each node. All reported energies are then re-evaluated exactly on the original discrete MRF using these assignments. We will update the experimental section, method description, and all relevant tables to clearly state the discretization procedure, confirm exact re-evaluation on the original energy, and list the precise time budgets together with instance sizes and cardinalities for the Toulbar2 comparisons. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained against external benchmarks

full rationale

The provided abstract and context present ReMAP as an instance-wise neural optimization framework whose central claims rest on a stated consistency between the relaxed objective and discrete MAP (analyzed via over-parameterization) plus empirical comparisons to external solvers (Toulbar2) and benchmarks (UAI 2022, PCI problems). No equations, self-citations, or fitted parameters are quoted that reduce any load-bearing prediction or uniqueness claim to a tautology or prior self-result. The method is treated as an independent optimization procedure without load-bearing self-citation chains or renaming of known results. This matches the default expectation for non-circular papers; the reader's score of 2.0 is consistent with minor or absent circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Ledger entries derived from abstract claims; full paper would provide more precise details on any additional parameters or assumptions.

free parameters (1)
  • GNN model parameters
    Parameters of the graph neural network are optimized during the per-instance gradient-based search.
axioms (1)
  • domain assumption The differentiable relaxation of the MRF energy is consistent with the discrete MAP objective
    Invoked in the abstract as shown by the authors.
invented entities (1)
  • Neural over-parameterization in continuous space no independent evidence
    purpose: To expose low-energy optimization paths not available in discrete space
    Introduced in the abstract to explain the method's advantage.

pith-pipeline@v0.9.0 · 5759 in / 1296 out tokens · 52428 ms · 2026-05-23T17:22:08.394872+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages

  1. [1]

    Markov random field modeling, inference & learning in computer vision & image understanding: A survey

    Chaohui Wang, Nikos Komodakis, and Nikos Paragios. Markov random field modeling, inference & learning in computer vision & image understanding: A survey. Computer Vision and Image Understanding, 117(11):1610–1627, 2013

    work page 2013

  2. [2]

    Image steganography with symmetric embedding using gaussian markov random field model

    Wenkang Su, Jiangqun Ni, Xianglei Hu, and Jessica Fridrich. Image steganography with symmetric embedding using gaussian markov random field model. IEEE Transactions on Circuits and Systems for Video Technology, 31(3):1001–1015, 2021

    work page 2021

  3. [3]

    Markov models applications in natural language processing: A survey

    Talal Almutiri and Farrukh Nadeem. Markov models applications in natural language processing: A survey. International Journal of Information Technology and Computer Science, 14:1–16, 04 2022

    work page 2022

  4. [4]

    Conditional random field autoencoders for unsupervised structured prediction

    Waleed Ammar, Chris Dyer, and Noah A Smith. Conditional random field autoencoders for unsupervised structured prediction. Advances in Neural Information Processing Systems, 27, 2014

    work page 2014

  5. [5]

    Enhanced sequence labeling based on latent variable conditional random fields

    Jerry Chun-Wei Lin, Yinan Shao, Ji Zhang, and Unil Yun. Enhanced sequence labeling based on latent variable conditional random fields. Neurocomputing, 403:431–440, 2020

    work page 2020

  6. [6]

    Graph convolutional networks with markov random field reasoning for social spammer detection

    Yongji Wu, Defu Lian, Yiheng Xu, Le Wu, and Enhong Chen. Graph convolutional networks with markov random field reasoning for social spammer detection. Proceedings of the AAAI Conference on Artificial Intelligence, 34(01):1054–1061, Apr. 2020

    work page 2020

  7. [7]

    Transportation data visualization with a focus on freight: a literature review.Transportation Planning and Technology, 45(4):358– 401, 2022

    Elkafi Hassini Yunfei Ma, Amir Amiri and Saiedeh Razavi. Transportation data visualization with a focus on freight: a literature review.Transportation Planning and Technology, 45(4):358– 401, 2022

    work page 2022

  8. [8]

    Reverend bayes on inference engines: A distributed hierarchical approach

    Judea Pearl. Reverend bayes on inference engines: A distributed hierarchical approach. In Probabilistic and causal inference: the works of Judea Pearl, pages 129–138. 2022

    work page 2022

  9. [9]

    Wainwright, T.S

    M.J. Wainwright, T.S. Jaakkola, and A.S. Willsky. Map estimation via agreement on trees: message-passing and linear programming. IEEE Transactions on Information Theory, 51(11):3697–3717, 2005

    work page 2005

  10. [10]

    Application of the mean field methods to mrf optimization in computer vision

    Masaki Saito, Takayuki Okatani, and Koichiro Deguchi. Application of the mean field methods to mrf optimization in computer vision. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1680–1687, 2012

    work page 2012

  11. [11]

    J. Zhang. The mean field theory in em procedures for blind markov random field image restoration. IEEE Transactions on Image Processing, 2(1):27–40, 1993

    work page 1993

  12. [12]

    toulbar2, an exact cost function network solver

    Simon De Givry. toulbar2, an exact cost function network solver. In 24ème édition du congrès annuel de la Société Française de Recherche Opérationnelle et d’Aide à la Décision ROADEF 2023, 2023

    work page 2023

  13. [13]

    Multi-language evaluation of exact solvers in graphical model discrete optimization

    Barry Hurley, Barry O’sullivan, David Allouche, George Katsirelos, Thomas Schiex, Matthias Zytnicki, and Simon de Givry. Multi-language evaluation of exact solvers in graphical model discrete optimization. Constraints, 21:413–434, 2016

    work page 2016

  14. [14]

    The lifted newton method and its application in optimization

    Jan Albersmeyer and Moritz Diehl. The lifted newton method and its application in optimization. SIAM Journal on Optimization, 20(3):1655–1684, 2010

    work page 2010

  15. [15]

    Lift-and-project for mixed 0–1 programming: recent progress

    Egon Balas and Michael Perregaard. Lift-and-project for mixed 0–1 programming: recent progress. Discrete Applied Mathematics, 123(1):129–154, 2002

    work page 2002

  16. [16]

    Lifting the convex conjugate in lagrangian relaxations: A tractable approach for continuous markov random fields

    Hartmut Bauermeister, Emanuel Laude, Thomas Möllenhoff, Michael Moeller, and Daniel Cremers. Lifting the convex conjugate in lagrangian relaxations: A tractable approach for continuous markov random fields. SIAM Journal on Imaging Sciences, 15(3):1253–1281, 2022

    work page 2022

  17. [17]

    Combinatorial optimization and reasoning with graph neural networks

    Quentin Cappart, Didier Chételat, Elias Khalil, Andrea Lodi, Christopher Morris, and Petar Veliˇckovi´c. Combinatorial optimization and reasoning with graph neural networks. Journal of Machine Learning Research, 24(130):1–61, 2023. 11

    work page 2023

  18. [18]

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

    work page 2022

  19. [19]

    Cutting planes in integer and mixed integer programming

    Hugues Marchand, Alexander Martin, Robert Weismantel, and Laurence Wolsey. Cutting planes in integer and mixed integer programming. Discrete Applied Mathematics, 123(1):397–446, 2002

    work page 2002

  20. [20]

    A lift-and-project cutting plane algorithm for mixed 0-1 programs

    Egon Balas, Sebastián Ceria, and Gérard Cornuéjols. A lift-and-project cutting plane algorithm for mixed 0-1 programs. Math. Program., 58:295–324, 05 1993

    work page 1993

  21. [21]

    Dey and Jean-Philippe Richard

    Santanu S. Dey and Jean-Philippe Richard. Linear-programming-based lifting and its application to primal cutting-plane algorithms. INFORMS Journal on Computing, 21(1):137–150, 2008

    work page 2008

  22. [22]

    Lifting for mixed integer programs with variable upper bounds

    Sergey Shebalov, Young Woong Park, and Diego Klabjan. Lifting for mixed integer programs with variable upper bounds. Discrete Applied Mathematics, 186:226–250, 2015

    work page 2015

  23. [23]

    A primal-dual lifting scheme for two-stage robust optimization

    Angelos Georghiou, Angelos Tsoukalas, and Wolfram Wiesemann. A primal-dual lifting scheme for two-stage robust optimization. Oper. Res., 68:572–590, 2020

    work page 2020

  24. [24]

    Adaptive distributionally robust optimiza- tion

    Dimitris Bertsimas, Melvyn Sim, and Meilin Zhang. Adaptive distributionally robust optimiza- tion. Manag. Sci., 65:604–618, 2019

    work page 2019

  25. [25]

    Learning deep graph matching with channel-independent embedding and hungarian attention

    Tianshu Yu, Runzhong Wang, Junchi Yan, and Baoxin Li. Learning deep graph matching with channel-independent embedding and hungarian attention. In International conference on learning representations, 2019

    work page 2019

  26. [26]

    Boosting graph pooling with persistent ho- mology

    Chaolong Ying, Xinjian Zhao, and Tianshu Yu. Boosting graph pooling with persistent ho- mology. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  27. [27]

    Graph learning for combinatorial optimization: A survey of state-of-the-art

    Yun Peng, Byron Choi, and Jianliang Xu. Graph learning for combinatorial optimization: A survey of state-of-the-art. Data Science and Engineering, 6, 06 2021

    work page 2021

  28. [28]

    Unsupervised training for neural tsp solver, 07 2022

    El¯ıza Gaile, Andis Draguns, Emils Ozolins, and Karlis Freivalds. Unsupervised training for neural tsp solver, 07 2022

    work page 2022

  29. [29]

    Unsupervised learning for solving the travelling salesman problem

    Yimeng Min, Yiwei Bai, and Carla P Gomes. Unsupervised learning for solving the travelling salesman problem. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 47264–47278. Curran Associates, Inc., 2023

    work page 2023

  30. [30]

    Maskell, and You Zhou

    Xuan Wu, Di Wang, Lijie Wen, Yubin Xiao, Chunguo Wu, Yuesong Wu, Chaoyu Yu, Douglas L. Maskell, and You Zhou. Neural combinatorial optimization algorithms for solving vehicle routing problems: A comprehensive survey with perspectives, 2024

    work page 2024

  31. [31]

    Weiss and W.T

    Y . Weiss and W.T. Freeman. On the optimality of solutions of the max-product belief- propagation algorithm in arbitrary graphs.IEEE Transactions on Information Theory, 47(2):736– 744, 2001

    work page 2001

  32. [32]

    Felzenszwalb and Daniel P

    Pedro F. Felzenszwalb and Daniel P. Huttenlocher. Efficient belief propagation for early vision. International Journal of Computer Vision, 70:41–54, 2004

    work page 2004

  33. [33]

    A revolution: Belief propagation in graphs with cycles

    Brendan Frey and David Mackay. A revolution: Belief propagation in graphs with cycles. 08 2002

    work page 2002

  34. [34]

    Kolmogorov

    V . Kolmogorov. Convergent Tree-Reweighted Message Passing for Energy Minimization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(10):1568–1583, October 2006

    work page 2006

  35. [35]

    A new look at reweighted message passing

    Vladimir Kolmogorov. A new look at reweighted message passing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(5):919–930, 2015

    work page 2015

  36. [36]

    Inference in probabilistic graphical models by graph neural networks, 2019

    KiJung Yoon, Renjie Liao, Yuwen Xiong, Lisa Zhang, Ethan Fetaya, Raquel Urtasun, Richard Zemel, and Xaq Pitkow. Inference in probabilistic graphical models by graph neural networks, 2019. 12

    work page 2019

  37. [37]

    Neural enhanced belief propagation on factor graphs

    Víctor Garcia Satorras and Max Welling. Neural enhanced belief propagation on factor graphs. In Arindam Banerjee and Kenji Fukumizu, editors, Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, volume 130 of Proceedings of Machine Learning Research, pages 685–693. PMLR, 13–15 Apr 2021

    work page 2021

  38. [38]

    Belief propagation neural networks

    Jonathan Kuck, Shuvam Chakraborty, Hao Tang, Rachel Luo, Jiaming Song, Ashish Sabharwal, and Stefano Ermon. Belief propagation neural networks. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 667–678. Curran Associates, Inc., 2020

    work page 2020

  39. [39]

    Factor graph neural networks

    Zhen Zhang, Fan Wu, and Wee Sun Lee. Factor graph neural networks. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 8577–8587. Curran Associates, Inc., 2020

    work page 2020

  40. [40]

    Variational message passing neural network for maximum-a-posteriori (map) inference

    Zijun Cui, Hanjing Wang, Tian Gao, Kartik Talamadupula, and Qiang Ji. Variational message passing neural network for maximum-a-posteriori (map) inference. In James Cussens and Kun Zhang, editors, Proceedings of the Thirty-Eighth Conference on Uncertainty in Artificial Intelligence, volume 180 ofProceedings of Machine Learning Research, pages 464–474. PMLR...

    work page 2022

  41. [41]

    GMNN: Graph Markov neural networks

    Meng Qu, Yoshua Bengio, and Jian Tang. GMNN: Graph Markov neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5241–5250. PMLR, 09–15 Jun 2019

    work page 2019

  42. [42]

    Mutual crf-gnn for few-shot learning

    Shixiang Tang, Dapeng Chen, Lei Bai, Kaijian Liu, Yixiao Ge, and Wanli Ouyang. Mutual crf-gnn for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2329–2339, June 2021

    work page 2021

  43. [43]

    Solving mixed integer programs using neural networks, 2021

    Vinod Nair, Sergey Bartunov, Felix Gimeno, Ingrid von Glehn, Pawel Lichocki, Ivan Lobov, Brendan O’Donoghue, Nicolas Sonnerat, Christian Tjandraatmadja, Pengming Wang, Ravichan- dra Addanki, Tharindi Hapuarachchi, Thomas Keck, James Keeling, Pushmeet Kohli, Ira Ktena, Yujia Li, Oriol Vinyals, and Yori Zwols. Solving mixed integer programs using neural net...

    work page 2021

  44. [44]

    Learning to solve the constrained most probable explanation task in probabilistic graphical models

    Shivvrat Arya, Tahrima Rahman, and Vibhav Gogate. Learning to solve the constrained most probable explanation task in probabilistic graphical models. In Sanjoy Dasgupta, Stephan Mandt, and Yingzhen Li, editors,Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, volume 238 of Proceedings of Machine Learning Research,...

    work page 2024

  45. [45]

    Neural network approximators for marginal map in probabilistic circuits

    Shivvrat Arya, Tahrima Rahman, and Vibhav Gogate. Neural network approximators for marginal map in probabilistic circuits. Proceedings of the AAAI Conference on Artificial Intelligence, 38(10):10918–10926, Mar. 2024

    work page 2024

  46. [46]

    Probabilistic graphical models: principles and techniques

    Daphne Koller and Nir Friedman. Probabilistic graphical models: principles and techniques. MIT press, 2009

    work page 2009

  47. [47]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017

    work page 2017

  48. [48]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    work page 2017

  49. [49]

    How powerful are graph neural networks? In International Conference on Learning Representations (ICLR), 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations (ICLR), 2019

    work page 2019

  50. [50]

    Lio, and Yoshua Bengio

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adrian Romero, P. Lio, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations (ICLR), 2018

    work page 2018

  51. [51]

    Projection, lifting and extended formulation in integer and combinatorial optimiza- tion

    Egon Balas. Projection, lifting and extended formulation in integer and combinatorial optimiza- tion. Annals OR, 140:125–161, 11 2005. 13

    work page 2005

  52. [52]

    Combinatorial Optimization: Algorithms and Complexity, volume 32

    Christos Papadimitriou and Kenneth Steiglitz. Combinatorial Optimization: Algorithms and Complexity, volume 32. IEEE, 01 1982

    work page 1982

  53. [53]

    Fixing max-product: Convergent message passing algorithms for map lp-relaxations

    Amir Globerson and Tommi Jaakkola. Fixing max-product: Convergent message passing algorithms for map lp-relaxations. Advances in neural information processing systems , 20, 2007

    work page 2007

  54. [54]

    Representation learning on graphs with jumping knowledge networks

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In International conference on machine learning, pages 5453–5462, 2018

    work page 2018

  55. [55]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference for Learning Representations (ICLR), 2015

    work page 2015

  56. [56]

    Identifying and attacking the saddle point problem in high-dimensional non-convex optimization, 2014

    Yann Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization, 2014

    work page 2014

  57. [57]

    The loss surfaces of multilayer networks, 2015

    Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, and Yann LeCun. The loss surfaces of multilayer networks, 2015

    work page 2015

  58. [58]

    On random graphs i

    P Erdös and A Rényi. On random graphs i. Publicationes Mathematicae Debrecen, 6:290–297, 1959

    work page 1959

  59. [59]

    Visualizing the loss landscape of neural nets

    Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. In Neural Information Processing Systems, 2018. 14 1 1 3 1 1 3 3 3 Figure 2: This illustrates the padding procedure for unary loss terms ϕ(x) and clique loss terms ψ(xi, xj, xk), with |X | = 5 . xmax denotes the variable that has the maximum...

    work page arXiv 2018