Deep Equilibrium Algorithmic Reasoning

Neural Algorithmic Reasoning (NAR) research has demonstrated that graph neural networks (GNNs) could learn to execute classical algorithms. However, most previous approaches have always used a recurrent architecture, where each iteration of the GNN matches an iteration of the algorithm. In this paper we study neurally solving algorithms from a different perspective: since the algorithm's solution is often an equilibrium, it is possible to find the solution directly by solving an equilibrium equation. Our approach requires no information on the ground-truth number of steps of the algorithm, both during train and test time. Furthermore, the proposed method improves the performance of GNNs on executing algorithms and is a step towards speeding up existing NAR models. Our empirical evidence, leveraging algorithms from the CLRS-30 benchmark, validates that one can train a network to solve algorithmic problems by directly finding the equilibrium. We discuss the practical implementation of such models and propose regularisations to improve the performance of these equilibrium reasoners.

Paper

References (51)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer E2D46/10 · confidence 1/52024-06-14

Summary

This work builds deep equilibrium graph neural networks for algorithmic reasoning. The paper tests their models on a variety of algorithm problems and finds mixed results with some positive and encouraging observations. One focus of the work is on speeding up NARs, and the paper also proposes regularizers to boost performance. To be honest, I am not an algorithms person, and a lot of ideas in the paper are foreign to me, so I am not confident about my review.

Strengths

The paper is well-written, polished, and clear. In general, the idea of learning algorithms with neural networks has the potential for practical value, although it seems to me that this paper focuses on emulating existing known algorithms and not learning new ones. Also, the experiments show encouraging results.

Weaknesses

The results seem mixed and not entirely positive. In my own experiments with DEQs, I’ve found that often they overfit to the solver used during training. Other solvers can find fixed points, but those fixed points don’t always map on to solutions. Whether or not the model actually has unique fixed points corresponding to solutions may be worth studying. The paper mentions another paper on DEQs for algorithmic reasoning (“Deep Equilibrium Models For Algorithmic Reasoning”), but does not discuss it in detail. This other paper has a nearly identical title and probably demands a more careful discussion and contextualization. As is, it would be easy for a reader to think that this paper is the first one to think about DEQs for algorithmic reasoning. It might be worth discussing the work on non-GNN recurrent networks for learning algorithms. For example, I have found that a model from "End-to-end Algorithm Synthesis with Recurrent Networks: Logical Extrapolation Without Overthinking" actually behaves similar to DEQs in practice.

Questions

N/A

Rating

6

Confidence

1

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors adequately discuss limitations throughout the paper.

Authorsrebuttal2024-08-12

Thank you for taking the time to read our rebuttal. We believe we have addressed all your concerns, therefore could we kindly ask if there is any particular reason why you have not increased your score or confidence? We found your review insightful and would be happy to listen to further comments. We are available for any additional concerns or questions.

Reviewer tzFP6/10 · confidence 1/52024-06-14

Summary

This paper proposes to solve the neural algorithmic reasoning by attacking the equilibrium solutions directly, without leveraging the recurrent structures which imitate the iterations in algorithms. They proposed deep equilibrium algorithmic reasoner (DEAR), and compare it with baselines including NAR (w/ and w/o Triplet-MPNN).Their training dynamics is shown to be more stable than baselines, and inference time is smaller than than baselines, although the accuracy is sometimes worse than baselines.

Strengths

* The motivation is clear. * Experiments are thorough.

Weaknesses

* It is not immediately clear what is the takeaway from the experiments. My understanding is that efficiency is the key selling point of the new algorithm. However, there is only one table for efficiency in Appendix G. * Section 4 is notation heavy and hard to understand. * I fully appreciate authors' honesty to report the results (Table 1) which are not to their advantage, but the inferior accuracy seems to limit the usefulness of the method.

Questions

* I'm not an expert so I have trouble understanding Section 4. Is it possible to make it clearer how does it connect to the proposed algorithm? * Eq. (4) seems still a recurrent structure? I was thinking the point of the new method is to get rid of recurrent structures? But looks like from Figure 1, the point is to reduce the number of recursive steps. Is this correct?

Rating

6

Confidence

1

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors adequately addressed the limitations.

Authorsrebuttal2024-08-12

We would like to sincerely thank you again for the insightful feedback from your rebuttal, and for additionally raising your score. We truly believe you helped strengthen our paper. Please let us know if you have any further suggestions to help us improve our work.

Reviewer fLfp5/10 · confidence 4/52024-07-01

Summary

This paper explores a novel approach to NAR using GNNs. Traditional NAR models typically use a recurrent architecture where each iteration of the GNN corresponds to an iteration of the algorithm being learned. Instead, this paper proposes that since many algorithms reach an equilibrium state where further iterations do not alter the outcome, it is possible to directly solve for this equilibrium. By training neural networks to find the equilibrium point, the authors aim to improve the alignment between GNNs and classical algorithms, potentially enhancing both the accuracy and speed of NAR models. Empirical evidence from the CLRS-30 benchmark supports the viability of this equilibrium-based approach.

Strengths

1. The proposed method is clearly presented. 2. The performance of the proposed method on benchmark datasets is good.

Weaknesses

1. The transition from denotational semantics to the proposed architecture could be further clarified to improve the overall understanding of the method. (Question 1) 2. It is not entirely clear whether the comparison with the baseline is accurate, and further justification or explanation may be needed. (Question 2) 3. The authors might consider revising the presentation of their contributions to more effectively convey the significance and novelty of the paper to the audience. (Question 3) 4. Including real-world experiments in the paper could substantially strengthen the contributions and further demonstrate the practical applicability of the proposed method. (Question 4, 5)

Questions

1. From reviewer's understanding, denotational semantics primarily motivate the equilibrium methods. - The connection between the proposed architecture (PGN with gating) and denotational semantics is not clear. - The motivation for equilibrium methods does not seem to require denotational semantics, as graph algorithms such as Bellman-Ford inherently have equilibrium points. Could the authors provide further clarification on this aspect? 2. It appears that the data for the experiments was generated by the authors rather than from CLRS-30, as they mention "For each algorithm, we generate" in Line 219. However, the baseline scores are identical to those reported in previous papers. To ensure a fair comparison, the authors should consider re-running the baselines on their own dataset. 3. From the reviewer's personal perspective, the introduction and contributions sections may not be well-suited for an academic paper in their current form. The investigated problem could benefit from better motivation, and the contributions might not fully capture the paper's key theoretical and empirical results. 4. The training and test graph sizes used in the experiments seem to be relatively small, although consistent with CLRS-30. The reviewer recommends training on larger graphs (e.g., 50 nodes) and testing on even larger graphs (e.g., 300 nodes) to better demonstrate the performance gains of the proposed methods. 5. The experiments presented in the paper are limited to synthetic graph datasets. The reviewer suggests that demonstrating the performance gains on real tasks (e.g., physical systems [1] or mathematical problems [2]) would enhance the contributions of this paper. 6. The authors mention that DEQ models can be difficult to train, which might be due to the instability of DEQ's gradient. The reviewer recommends considering the use of [3] for training the models, as it provides a more stable approach. [1] Battaglia, Peter, et al. "Interaction networks for learning about objects, relations and physics." NeurIPS 2016. [2] Lample, Guillaume, and François Charton. "Deep learning for symbolic mathematics." ICLR 2020. [3] Fung, Samy Wu, et al. "Jfb: Jacobian-free backpropagation for implicit networks." AAAI 2022.

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

Yes

Reviewer fLfp2024-08-08

I appreciate the authors' efforts to address my concerns, which has led me to increase my evaluation to a score of 5. The size generalization performance of DEQ is particularly noteworthy, aligning with the out-of-distribution generalization capabilities demonstrated in previous DEQ studies. Concerning the first point, the term 'formal motivation' used by the authors in reference to denotational semantics has left me a bit perplexed. While I recognize the attempt to provide a theoretical grounding, it appears to me more as an intuition rather than leading to a rigorous formal derivation that substantiates the utility of DEQ in graph reasoning. Moreover, the connection between denotational semantics and graph reasoning problems seems tenuous, with the only clear link being the capability to solve such problems through programming. As for the second point, **I have reservations about the rigorousness of reusing scores from a previous paper when the test dataset has changed**, but I am not sure about it.

Authorsrebuttal2024-08-09

We thank the reviewer for promptly replying to our rebuttal and we’re glad we were able to address some of your concerns. *[..] the term 'formal motivation' used by the authors in reference to denotational semantics has left me a bit perplexed. [..] the connection between denotational semantics and graph reasoning problems seems tenuous [..]* We appreciate the comment from the reviewer, and we would like to clarify that we do not aim at providing a connection between "denotational semantics and graph reasoning problems", but rather to provide a connection between finding the fixed point of a function and executing an algorithm to termination. This then provides an analogy between finding the fixed point of a neural algorithmic reasoner and executing an algorithm to termination, which is the base motivation of our proposed method. We will change our language in our paper to remove any ambiguities and make clear that we use the mathematical background as a strong motivating factor for our method. *[..] I have reservations about the ethical implications of reusing scores from a previous paper [..]* We understand the comment from the reviewer and we would like to add a comment, as perhaps the words used should be related to scientific rigour, rather than ethical concerns. Reporting results from other papers, providing proper citation and mention (we agree that it would be extremely unethical without this), is a common practice in literature, especially when the code for some models is not made public. We make sure to use the **same data generation code** used by all considered papers in the literature. For CLRS there isn't a predefined dataset, but rather everyone uses the same data-generation library, ensuring the same data distribution. We note however, that there is some [official downloading code](https://github.com/google-deepmind/clrs/blob/d1c2ad7af8437c7536cf329d9cef8fdf93184d9d/clrs/examples/run.py#L152) for test data generated with CLRS. All algorithms are reported in the results below except for SCC. This algorithm was skipped as pointers in the downloaded dataset were not in the edge-set. PyTorch Geometric does not support this feature and this difference will be mentioned in Appendix B within our paper. We report results with DEAR on all remaining algorithms. They are consistent with our data and even improve on DSP, Bellman-F. and Floyd-W.. | Algorithm | Mean | Std Dev | |--------------------------|---------|----------| | Bellman-F. | 98.59% | 0.34 | | Floyd-W. | 63.92% | 0.04 | | DSP | 92.95% | 1.38 | | MST Prim | 89.33% | 1.09 | | BFS | 99.65% | 0.15 | | DFS | 38.27% | 0.82 | | Search (Binary) | 74.00% | 11.5 | | Minimum | 99.33% | 1.15 | | Sort (Ins) | 85.48% | 6.90 | We thank the reviewer again for all the insightful comments, and remain available for further clarifications if any doubts remain.

Reviewer fLfp2024-08-09

Although the performance of the proposed method is superior, I am not convinced that the motivation outlined by the authors differs significantly from any other theoretically inspired motivations. Therefore, I will maintain my score as a borderline accept.

Reviewer 1Bds6/10 · confidence 3/52024-07-13

Summary

This paper proposes Deep Equilibrium Algorithmic Reasoner (DEAR) which uses a deep equilibrium model (DEQ) to solve algorithmic tasks in CLRS30. The paper first introduces denotational semantics, which can be used to denote programs. It also provides an overview of Domain theory, and uses it to show that algorithms have fixed points. The paper then trains pointer graph network on algorithms from CLRS30 as a DEQ — fixed point solving is done with Anderson solver. The paper also discusses different methods (including failure cases) that were attempted to further improve performance of DEAR — use of Cayley Graph Propagation and alignment loss. DEAR improves performance over NAR (Neural Algorithmic Reasoner, the preivous state-of-the-art) on many algorithms.

Strengths

1. The premise of this paper makes perfect sense. Message passing in GNNs is known to converge to an equilibrium. Therefore, it is obvious to combine NAR with DEQs. 2. DEAR improves OOD generalization of algorithms such as Floyd-Warshall, DFS, SCC, Sorting, and performs comparably to NAR (previous state-of-the-art) on algorithms like Breadth-first search, Minimum. 3. DEAR improves inference speed of many algorithms as shown in Appendix G. Overall, this work is quite novel. While there are some prior works (See Anil et al. 2022) that have applied DEQs on very small GNNs and simple algorithmic tasks, this paper picks up a relatively difficult problem of solving standard algorithms. This hasn't been explored before with DEQ based architecture.

Weaknesses

1. DEAR hurts performance on some algorithms like Binary Search (significant drop in performance), DSP, and MST Prime, and the reasoning is unclear. 2. The authors have done an excellent job at explaining literature that many readers might not be familiar with. However, the readability of paper can be improved further if the authors provide more background on domain theory for those who are not familiar with it. It will also help if some discussion is added in Appendix A. I couldn’t understand it even after attempting to read it multiple times.

Questions

1. Why use a partial function in equation 3? 2. There is some prior work [1] which indicates that more steps to find equilibrium points help with better OOD generalization. From my understanding, DEAR penalizes longer trajectories (due to misalignment with domain theory). Is it possible to get improved performance if we ignore the potential conflict with domain theory? 
[1] Anil, Cem, et al. "Path independent equilibrium models can better exploit test-time computation." Advances in Neural Information Processing Systems 35 (2022): 7796-7809. 3. What loss objective is used to train DEAR (e.g. cross entropy)? I understand that there are auxiliary losses such as alignment loss in Lines 277-298.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations have been discussed.

Authorsrebuttal2024-08-13

Thank you for taking the time to read our rebuttal. We are happy we have addressed all your questions through our detailed response, and will ensure that the updated experiments and additional literature on domain theory will be included in our paper. For these reasons, could we therefore kindly ask if there is any particular reason why you have not increased your score? We found your review insightful and it helped strengthen our paper. Please let us know if you have any further questions or suggestions.

Reviewer tzFP2024-08-12

I want to thank the authors for nicely addressing my concerns. Therefore, I'm raising my score.

Reviewer 1Bds2024-08-12

Thank you for your detailed responses my questions. I hope the authors will include the new experiments as well as additional literature on domain theory as promised. I would like to retain my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC