Pith. sign in

REVIEW 5 major objections 5 minor 65 references

Brain-inspired Chaotic Graph Backpropagation for Large-scale Combinatorial Optimization

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding a chaotic local loss term to GNN training, then annealing its strength, lets the network escape local minima and solve large-scale combinatorial optimization problems at least as well as specialized…

desk verdict A useful empirical training trick for GNN solvers, wrapped in an unsupported global-optimality narrative and a SOTA claim its own Table 1 contradicts. read the letter →

arxiv 2412.09860 v1 pith:K4BWBQ2Q submitted 2024-12-13 cs.LG cs.AIcs.NE

Peng Tao , Kazuyuki Aihara , Luonan Chen This is my paper · ORCID
classification cs.LGcs.AIcs.NE
keywords combinatorialoptimizationgraphneuralnetworkschaoticbackpropagationsimulatedannealingmaximumcutcoloringindependentsetMarottochaos
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

The paper introduces chaotic graph backpropagation (CGBP), a training rule for graph neural networks that adds a local "chaotic loss" to the physics-inspired Hamiltonian loss used for combinatorial optimization. The claim is that when the chaotic strength is large, the weight updates become genuinely chaotic (Marotto chaos, high-dimensional topological chaos), and the resulting global ergodicity and pseudo-randomness let the network explore the loss landscape and escape local minima that trap ordinary backpropagation. Annealing the chaotic strength down to zero then lets the dynamics settle into a good solution. On maximum independent set, maximum cut, and graph coloring benchmarks, CGBP-trained GNNs outperform standard GNN training and, on several instances, match or beat classical solvers. Because the extra loss is a plug-in term, the paper argues CGBP can improve any GNN training method rather than only this particular solver.

What carries the argument

The load-bearing object is the chaotic loss term $\mathcal{L}_C = -z \sum_{l,j} [I_0 \ln o^{(l)}_{dj} + (1-I_0)\ln(1-o^{(l)}_{dj})]$, added to the Hamiltonian loss $\mathcal{L}_H$. Its gradient with respect to a weight is proportional to $-z(I_0 - o^{(l)}_{dj}) h^{(l)}_i c_{kj}$, so the term acts as local negative feedback and, for large chaotic strength $z$, makes the discrete weight-update map $W(t+1)=F(W(t))$ a snap-back repeller system exhibiting Marotto chaos. Annealing $z$ by the rule $z\leftarrow \beta z$ with $\beta<1$ implements chaotic simulated annealing: the trajectory is globally exploratory early and reduces to gradient dynamics late.

What would settle it

Use a small max-cut or independent-set instance with a known optimum and compare, over many random seeds, the same GNN trained with the chaotic term $z>0$ against the identical network with $z=0$. If the chaotic runs never reach the known optimum while multi-start backpropagation does, or if runs with positive Lyapunov exponents systematically end at higher loss than $z=0$ runs, the central claim that chaos enables global optimization fails. A best-of-100 comparison on a fixed 3-regular graph with $n=100$ would be decisive, since the paper's own results show seed-dependent overlap between BP and CGBP on such instances.

Watch

Extended reading notes

Core claim

The central discovery is that chaos in weight space, induced deliberately through a cross-entropy-like local loss whose gradient acts as negative feedback on each weight, is not a nuisance but a global search mechanism. For sufficiently large chaotic strength $z$, the update map $W(t+1)=F(W(t))$ is shown to exhibit Marotto chaos; with $z$ annealed by $z\leftarrow \beta z$, training starts in a chaotic regime, passes through bifurcations, and ends in gradient descent. The paper reports that this schedule consistently lowers the Hamiltonian loss and raises solution quality relative to backpropagation on 3-regular graphs, and that on the Gset max-cut instances and the Queen and Citation coloring instances, the CGBP versions reach or exceed the best known results of classical state-of-the-art methods, including optimal cuts on G49 and G50.

Load-bearing premise

The load-bearing premise is that chaos in the weight updates—the property that the training trajectory wanders ergodically over a strange attractor—by itself means the trajectory will find the global minimum of the loss function; the paper gives numerical evidence but no proof that ergodicity in weight space implies reaching the loss minimum.

Editorial extensions

If this is right

  • Any existing GNN-based combinatorial optimization solver can be upgraded by adding the chaotic loss term, without changing the problem encoding or network architecture.
  • The linear time complexity of the GNN solver is preserved, so the improvement applies to graphs with up to millions of nodes, although optimality is not guaranteed at that scale.
  • The benefit is robust across optimizers: CGBP lifts SGD, SGDM, and Adam to similar quality, removing a major source of variance in unsupervised GNN solvers.
  • On max-cut instances G49 and G50, CGBP attains the known optimal cut value, and on graph-coloring benchmarks it reduces conflicting edges far below the backpropagation-trained baselines, for example to 2 on Pubmed.

Reading between the lines

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

  • If chaos is the operative mechanism, the same chaotic-loss recipe should transfer to non-graph neural architectures and even to classical local-search heuristics; the paper tests only GCN and GraphSAGE, so a direct test on MLPs, Transformers, or tabu search would clarify the mechanism.
  • The paper links global ergodicity to global optimization, but ergodicity on a strange attractor in weight space does not by itself guarantee convergence to the global minimum of the loss; a proof would need to show the chaotic invariant measure concentrates on low-loss regions.
  • Because the chaotic loss is evaluated at intermediate neuron outputs, its benefit may come from local credit assignment rather than chaos per se; comparing against a non-chaotic local loss with the same annealing schedule would separate the two.
  • A robustness check beyond the reported hyperparameter grid would be to vary $z$ and $\beta$ jointly over a wider range and verify that the best-of-many-seeds solution quality remains stable; the paper reports medians over 100 runs for fixed pairs only.
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

5 major / 5 minor

Summary. This paper proposes CGBP (chaotic graph backpropagation), a training algorithm that adds a 'chaotic loss' term to the GNN objective so that the weight-update dynamics become chaotic. The authors claim that the global ergodicity and pseudo-randomness of these chaotic dynamics let the GNN escape local minima and find globally optimal solutions, and they apply CGBP to PI-GNN-based solvers for maximum independent set, maximum cut, and graph coloring. They report improvements over standard backpropagation on synthetic 3-regular graphs, compare against established solvers on Gset, Queen, and Citation benchmarks, and claim linear time complexity and plug-in usability for any GNN-based method.

Significance. If the central claims held, the paper would make a useful contribution by showing that a chaotic training rule can improve unsupervised GNN solvers for large combinatorial optimization problems while preserving linear scaling. The paper has several positive features: public code is available, the empirical study uses public benchmarks, and the scaling experiment on 3-regular graphs up to 10^5 nodes is informative. However, the main theoretical claim connecting chaos to global optimality is not established, the empirical protocol is asymmetric, and the headline 'outperforms SOTA' statement is contradicted by the authors' own Table 1 on three of five Gset instances. As presented, the contribution is a heuristic training rule with partial empirical evidence rather than the theoretically grounded global optimization algorithm the paper claims.

major comments (5)
  1. [Abstract; Section 4] The central theoretical claim is not established. The abstract and Section 4 assert that the 'global ergodicity and pseudo-randomness' of the chaotic dynamics 'enable CGBP to learn each optimal GNN effectively and globally.' However, the paper does not prove a theorem connecting Marotto chaos of the weight map F in Eq. (10) to global optimality of the Hamiltonian loss loss_H in Eq. (4). Ergodicity on a strange attractor of the weight dynamics does not imply that the trajectory reaches the preimage of the global minimum of H, because the map from weights to node probabilities and then to loss_H is non-injective and the attractor is shaped by the combined loss. Moreover, the annealing schedule in Eq. (11) makes the system non-autonomous, so no invariant measure exists and strict ergodicity does not apply to the training process. The Discussion's own caveat that 'the obtained solution may not [be] optimal' contradicts the global-optimality claim. The authors should either supply a proof for the actual GCN/GraphSAGE updates or substantially weaken the claim to 'chaotic exploration may improve solution quality.'
  2. [Table 1; Abstract] The claimed 'outperform existing SOTA methods' is contradicted by the paper's own Table 1. On G14, G15, and G22, the best CGBP results are 3035, 3016, and 13318, respectively, whereas BLS and KHLWG reach 3064, 3050, and 13359. The text in Section 3.3 correctly downgrades this to 'comparable,' but the abstract and introduction continue to claim outperformance. This overstatement directly affects the paper's central empirical contribution and must be corrected, with conclusions limited to the instances where CGBP actually matches or exceeds SOTA.
  3. [Section 5.3; Tables 1-3] The empirical protocol is asymmetric and does not support the claim that CGBP outperforms existing GNN algorithms. Section 5.3 states that hyperopt is used 'to fully unleash the performance of CGBP' with up to 300 hyperparameter samples, while the baseline PI-GCN and PI-SAGE results appear to be taken from the original publications with their default settings. Under this protocol, the comparison conflates algorithmic improvement with hyperparameter tuning. A fair comparison requires tuning the baselines under the same search budget, or at least reporting baseline results with the same optimizer and search strategy.
  4. [Section 2.4; Eq. (9)] The proof of Marotto chaos is not transferred to the GNN setting. The text asserts in Section 2.4 that when z is sufficiently large, Eq. (9) exhibits Marotto chaos, citing the authors' prior MLP work (ref. 37). But Eq. (9) now contains node selection d and shared weights across all nodes in the GCN/GraphSAGE update; the earlier proof does not automatically apply. The only numerical evidence of positive Lyapunov exponents is on the 3-node toy model in Fig. 2, not on the shared-weight GNN used in the benchmark experiments. The authors should either provide a theorem for the GCN update or state the chaotic behavior as a numerical observation.
  5. [Figs. 3-4; Section 3.2] The experiments do not isolate the effect of chaos. In addition to the chaotic loss, CGBP introduces a stochastic node-selection scheme (CGBP-R in Fig. 2) and an annealing schedule, Eq. (11). The comparison of CGBP against BP therefore changes multiple factors simultaneously, so the observed improvements cannot be attributed specifically to chaotic dynamics. An ablation that replaces the chaotic term with a non-chaotic random perturbation of the same magnitude, or that keeps the same stochastic selection and annealing without the chaotic loss, is needed to support the mechanistic claim in Section 4.
minor comments (5)
  1. [Section 5.3] The text mentions 'the hyperparameters z and β introduced in CSBP'; this should be 'CGBP'.
  2. [Eq. (7), Eq. (8)] The mathematical notation in Eq. (7) and the surrounding derivation appears garbled in the rendered manuscript; please ensure all formulas are typeset correctly.
  3. [Introduction] The phrase 'outperforming not only the existing GNN learning algorithms but also SOTA methods' should be aligned with the more cautious 'comparable' language used in Section 3.3.
  4. [Abstract; Section 4] The claim that CGBP is a universal plug-in for 'any existing method' is not supported by experiments, which only consider PI-GNN; please temper the claim or add evidence with a different base method.
  5. [Fig. 4] Figure 4a reports an approximation ratio 'around 0.95' but does not specify whether this is the median, mean, or best over the 100 runs; please clarify in the caption or text.

Circularity Check

3 steps flagged · score 4.0 of 10

Self-citations carry the chaos-to-global-optimality bridge, and benchmark hyperparameters are tuned on the test instances; the core empirical CGBP-vs-BP comparison remains independent.

  1. self citation load bearing [Section 4 (Discussion), after Eq. (11)]
    "The first reason is that chaotic dynamics has been widely used to solve optimization problems due to its theoretically guaranteed global ergodicity and pseudo-randomness15-18, and existing theory has shown that the chaotic loss introduced by the CGBP method can induce chaotic dynamics when z is sufficiently large 19, which generates Marotto chaos and thus makes the training dynamics rich and global."

    The load-bearing premise that CGBP's chaotic dynamics yield 'global ergodicity and pseudo-randomness' and therefore effective global optimization is not derived in this paper. It is imported from the authors' own earlier work: refs. 15 and 18 (Chen & Aihara) for 'global searching ability' and 'chaotic simulated annealing', and ref. 19 (Chen & Aihara) for the Marotto-chaos claim. These citations are treated as external theorems, but they concern chaotic neural-network models and MLP backpropagation, not the shared-weight GCN/GraphSAGE update in Eq. (9) with the redefined z. Thus the central inference 'CGBP is chaotic => ergodic => globally optimal' rests on the authors' prior results rather than on a proof presented here.

  2. ansatz smuggled in via citation [Section 2.4, Eq. (5), and Section 2.4 text following Eq. (9)]
    "Recently, we proposed a chaotic backpropagation (CBP) 37 algorithm for multilayer perceptron (MLP), which introduces a loss function to simulate chaotic dynamics in the brain. Here, we use a similar strategy as CBP to construct the chaotic graph backpropagation (CGBP) algorithm. That is, an additional chaotic loss function lossC is added to the original loss function lossH."

    The specific form of the chaotic loss, with its cross-entropy-like structure and the claim that it generates useful chaotic dynamics, is adopted by direct analogy from the authors' own CBP paper (ref. 37). The GNN extension changes the gradient structure materially because all nodes share weights and the chaotic loss must sum over selected nodes d, yet no theorem is re-derived for the GCN/GraphSAGE dynamics in Eqs. (8)-(9); the text simply states that Marotto chaos follows 'when z is sufficiently large' and cites prior work. In other words, the chaotic-loss ansatz is inherited from a self-citation rather than independently justified for the setting in which it is now used.

1 more flagged steps
  1. fitted input called prediction [Section 5.3, 'Setting of training parameters']
    "To fully unleash the performance of CGBP, we applied a mixed strategy to adjust the hyperparameters. Specifically, we used the hyperopt 65 package to find the optimal combination of hyperparameters, which samples from the hyperparameter range set given by the user (50~300 maximum sampling times according to the difficulty of the problem). The dropout probability ranged from 0 to 0.5, and the learning rate ranged from 0.000001 to 0.1."

    The benchmark comparisons that support the headline 'CGBP can outperform ... SOTA methods' are obtained after hyperopt searches over chaotic strength z, annealing β, dropout, and learning rate on the same Gset, Queen, and Citation instances that are later reported as results. No separate validation split or held-out procedure is described. Consequently, the reported NMC/NGC numbers are best-after-tuning values rather than out-of-sample predictions, so the claim of superiority is partly a fitted-input report rather than a prediction from the method with fixed, independently chosen hyperparameters. This is a moderate fitted-input concern, not a definitional collapse of the algorithm.

full rationale

The paper's empirical core—CGBP versus BP on 3-regular graphs and on public benchmarks—is not circular: the comparisons are actual training runs and the improvement over BP is an independent result, even if hyperparameters are tuned on the test instances. However, the theoretical explanation for why CGBP should achieve global optimization is carried almost entirely by the authors' own earlier theorems on chaotic neural networks and chaotic backpropagation (refs. 15, 18, 19, 37), without re-deriving Marotto chaos for the shared-weight GCN/GraphSAGE dynamics used here. The lossC ansatz is likewise transferred from CBP by citation. These are load-bearing self-citations because the abstract and discussion explicitly attribute CGBP's success to 'global ergodicity and pseudo-randomness' that are 'theoretically guaranteed' by the cited prior work. Since the central empirical claim still has independent content and is not forced by definition, the appropriate score is 4 rather than 6 or higher.

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

The method's reported performance depends on several tuned hyperparameters (z, beta, learning rate, dropout, I0) and on unproved premises connecting chaotic dynamics to global optimization and loss to solution quality. The paper introduces no new physical or conceptual entities beyond the chaotic loss term.

free parameters (5)
  • chaotic strength z (per layer) = 3 and 1 for two GNN layers; 20 for the embedding layer (tuned via hyperopt)
    Controls the amplitude of the chaotic loss and the duration of chaotic exploration; selected per benchmark with hyperopt.
  • annealing constant beta = 0.999 default; 0.99 to 0.9999 tested
    Determines how quickly chaos is annealed into gradient dynamics; treated as a tunable hyperparameter.
  • learning rate = range 1e-6 to 0.1, selected by hyperopt
    Standard training hyperparameter tuned per dataset.
  • dropout probability = range 0 to 0.5, selected by hyperopt
    Regularization hyperparameter tuned per dataset.
  • I0 (chaotic loss constant) = 0.65
    Fixed constant in the chaotic loss, inherited from the authors' previous CBP work; chosen by hand.
assumptions (4)
  • domain assumption The update rule in Eq. (9) exhibits Marotto chaos when z is sufficiently large.
    This is a theorem from the authors' prior work (refs 18, 19, 37) and is not re-derived in this paper; the paper relies on it to justify the 'chaotic' label and global search claim.
  • ad hoc to paper Global ergodicity and pseudo-randomness of the chaotic dynamics imply that CGBP finds the global optimum of the Hamiltonian loss.
    Asserted in the Abstract and Discussion without proof; ergodicity on an attractor does not by itself guarantee convergence to the loss minimum.
  • domain assumption The QUBO/Potts Hamiltonian loss is a faithful proxy for combinatorial solution quality when minimized by GNN outputs.
    The paper admits in Section 4 that 'sometimes the solution with a lower loss has a worse quality', which undermines this assumption.
  • domain assumption Annealing z to zero (chaotic simulated annealing) brings the training dynamics to a good, converged solution.
    The annealing schedule is borrowed from ref 18; no convergence guarantee for the GNN setting is provided in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Brain-inspired Chaotic Graph Backpropagation for Large-scale Combinatorial Optimization." pith.science (2026). https://pith.science/paper/K4BWBQ2Q

@misc{pith2026241209860,
  author       = {Pith},
  title        = {Pith review of: Brain-inspired Chaotic Graph Backpropagation for Large-scale Combinatorial Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4BWBQ2Q}},
  note         = {Machine review of arXiv:2412.09860}
}
read the original abstract

Graph neural networks (GNNs) with unsupervised learning can solve large-scale combinatorial optimization problems (COPs) with efficient time complexity, making them versatile for various applications. However, since this method maps the combinatorial optimization problem to the training process of a graph neural network, and the current mainstream backpropagation-based training algorithms are prone to fall into local minima, the optimization performance is still inferior to the current state-of-the-art (SOTA) COP methods. To address this issue, inspired by possibly chaotic dynamics of real brain learning, we introduce a chaotic training algorithm, i.e. chaotic graph backpropagation (CGBP), which introduces a local loss function in GNN that makes the training process not only chaotic but also highly efficient. Different from existing methods, we show that the global ergodicity and pseudo-randomness of such chaotic dynamics enable CGBP to learn each optimal GNN effectively and globally, thus solving the COP efficiently. We have applied CGBP to solve various COPs, such as the maximum independent set, maximum cut, and graph coloring. Results on several large-scale benchmark datasets showcase that CGBP can outperform not only existing GNN algorithms but also SOTA methods. In addition to solving large-scale COPs, CGBP as a universal learning algorithm for GNNs, i.e. as a plug-in unit, can be easily integrated into any existing method for improving the performance.

Figures

Figures reproduced from arXiv: 2412.09860 by the authors.

Figure 1
Figure 1. Schematic diagram of CGBP. a shows the procedure of using GNN to solve the MIS problem. First, solving the MIS problem is mapped to the training process of GNN, and the probability of each node belonging to the MIS is obtained after training. Finally, the probability values are projected to the solution of the MIS problem by setting a proper threshold. b shows a simple two-layer GNN with trainable parameters denoted… view at source ↗
Figure 2
Figure 2. Numerical results on 2-regular graphs. a [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Numerical results on a 3-regular graph with 100 nodes. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Numerical results on large-scale 3-regular graphs for MC problem. a shows the statistical results of the NMC obtained by CGBP in 100 independent training runs as the number of nodes n increases from 102 to 105 , where the approximation ratio of NMC to the theoretical u…
Figure 5
Figure 5. Figure 5: Graph coloring results of CGBP on the Queen dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Graph coloring results of CGBP on the Pubmed graph. [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 64 canonical work pages

  1. [1]

    Machine learning for combinatorial optimization: a methodol ogical tour d’horizon

    Bengio Y , Lodi A, Prouvost A. Machine learning for combinatorial optimization: a methodol ogical tour d’horizon. European Journal of Operational Research 290, 405-421 (2021)

    work page 2021

  2. [2]

    Application of combinatorial optimization strategies in synthetic biology

    Naseri G, Koffas MA. Application of combinatorial optimization strategies in synthetic biology. Nature communications 11, 2446 (2020)

    work page 2020

  3. [3]

    Combinatorial optimization of mRNA structure, stability, and translation for RNA-based therapeutics

    Leppek K , et al. Combinatorial optimization of mRNA structure, stability, and translation for RNA-based therapeutics. Nature Communications 13, 1536 (2022)

    work page 2022

  4. [4]

    Explicit Evolutionary Multitasking for Combinatorial Optimization: A Case Study on Cap acitated Vehicle Routing Problem

    Feng L , et al. Explicit Evolutionary Multitasking for Combinatorial Optimization: A Case Study on Cap acitated Vehicle Routing Problem. IEEE Transactions on Cybernetics 51, 3143-3156 (2021)

    work page 2021

  5. [5]

    Reinforcement learning for combinatorial optimization: A survey

    Mazyavkina N, Sviridov S, Ivanov S, Burnaev E. Reinforcement learning for combinatorial optimization: A survey. Computers & Operations Research 134, 105400 (2021)

    work page 2021

  6. [6]

    Pointer networks

    Vinyals O, Fortunato M, Jaitly N. Pointer networks. In Advances in neural information processing systems, (2015)

    work page 2015

  7. [7]

    Erdos goes ne ural: an unsupervised learning framework for combinatorial optimization on graphs

    Karalias N, Loukas A. Erdos goes ne ural: an unsupervised learning framework for combinatorial optimization on graphs. In Advances in Neural Information Processing Systems, 6659-6672 (2020)

    work page 2020

  8. [8]

    Neural combinatorial optimization with reinforcement learning

    Bello I, Pham H, Le QV , Norouzi M, Bengio S. Neural combinatorial optimization with reinforcement learning. arXiv preprint arXiv:161109940 , (2016)

    work page 2016

Show all 65 references
  1. [9]

    Learning combinatorial optimization algorithms over graphs

    Khalil E, Dai H, Zhang Y , Dilkina B, Song L. Learning combinatorial optimization algorithms over graphs. In Advances in neural information processing systems, (2017)

    2017

  2. [10]

    NeuroLKH: Combining deep learning model with Lin-Kernighan-Helsgaun heuristic for solving the traveling salesman problem

    Xin L, Song W, Cao Z, Zhang J. NeuroLKH: Combining deep learning model with Lin-Kernighan-Helsgaun heuristic for solving the traveling salesman problem. Advances in Neural Info rmation Processing Systems , 7472-7483 (2021)

    2021

  3. [11]

    An effective implemen tation of the Lin–Kernighan traveling salesman heuristic

    Helsgaun K. An effective implemen tation of the Lin–Kernighan traveling salesman heuristic. European journal of operational research 126, 106-130 (2000)

    2000

  4. [12]

    Combinatorial optimization with physics-inspired graph neural networks

    Schuetz MJA, Brubaker JK, Katzgraber HG. Combinatorial optimization with physics-inspired graph neural networks. Nature Machine Intelligence 4, 367- 377 (2022)

    2022

  5. [13]

    Adam: A met hod for stochastic optimization

    Kingma DP, Ba J. Adam: A met hod for stochastic optimization. arXiv preprint arXiv:14126980, (2014)

    2014

  6. [14]

    Nonlinear dynamics and chaos with student solutions manual: With applications to ph ysics, biology, chemistry, and engineering

    Strogatz SH. Nonlinear dynamics and chaos with student solutions manual: With applications to ph ysics, biology, chemistry, and engineering. CRC press (2018)

    2018

  7. [15]

    Global searching ability of chaotic neural networks

    Chen L, Aihara K. Global searching ability of chaotic neural networks. IEEE Transactions on Circuits and Systems I: Regular Papers 46, 974-993 (1999)

    1999

  8. [16]

    Co mbinatorial optimization by simulating adiabatic bifurcations in non linear Hamiltonian systems

    Goto H, Tatsumura K, Dixon AR. Co mbinatorial optimization by simulating adiabatic bifurcations in non linear Hamiltonian systems. Science Advances 5, eaav2372 (2019)

    2019

  9. [17]

    High-performance combinatorial optimization based on classical mechanics

    Goto H , et al. High-performance combinatorial optimization based on classical mechanics. Science Advances 7, eabe7953 (2021)

    2021

  10. [18]

    Chaotic simula ted annealing by a neural network model with transient chaos

    Chen L, Aihara K. Chaotic simula ted annealing by a neural network model with transient chaos. Neural networks 8, 915-930 (1995)

    1995

  11. [19]

    Chaos and asymptot ical stability in discrete-time neural networks

    Chen L, Aihara K. Chaos and asymptot ical stability in discrete-time neural networks. Physica D 104, 286-325 (1997)

    1997

  12. [20]

    Phas e locking, period-doubling bifurcations, and irregular dynamics in periodically stimul ated cardiac cells

    Guevara MR, Glass L, Shrier A. Phas e locking, period-doubling bifurcations, and irregular dynamics in periodically stimul ated cardiac cells. Science 214, 1350-1353 (1981)

    1981

  13. [21]

    Subthreshold dynamics in periodically stimulated squid giant axons

    Kaplan DT, Clay JR, Manning T, Glass L, Guevara MR, Shrier A. Subthreshold dynamics in periodically stimulated squid giant axons. Physical Review Letters 76, 4074-4077 (1996)

    1996

  14. [22]

    Structures of attractors in periodically forced neural oscillators

    Aihara K, Numajiri T, Matsumoto G, Kotani M. Structures of attractors in periodically forced neural oscillators. Physics Letters A 116, 313-317 (1986)

    1986

  15. [23]

    How brains ma ke chaos in order to make sense of the world

    Skarda CA, Freeman WJ. How brains ma ke chaos in order to make sense of the world. Behavioral and Brain Sciences 10, 161-173 (1987)

    1987

  16. [24]

    Is it healthy to be chaotic? Science 243, 604-607 (1989)

    Pool R. Is it healthy to be chaotic? Science 243, 604-607 (1989)

    1989

  17. [25]

    On chaotic dynamics in transcription factors and the associated effects in differential gene regulation

    Heltberg ML, Krishna S, Jensen MH. On chaotic dynamics in transcription factors and the associated effects in differential gene regulation. Nature Communications 10, 71 (2019)

    2019

  18. [26]

    Oscillations, complex spatiotemporal behavior, and information transport in networks of exci tatory and inhibitory neurons

    Destexhe A. Oscillations, complex spatiotemporal behavior, and information transport in networks of exci tatory and inhibitory neurons. Physical Review E 50, 1594 (1994)

    1994

  19. [27]

    Evidence of chaotic dynamics of brain activity during the sleep cycle

    Babloyantz A, Salazar JM, Nicolis C. Evidence of chaotic dynamics of brain activity during the sleep cycle. Physics Letters A 111, 152-156 (1985)

    1985

  20. [28]

    Chaos and phase locking in normal squid axons

    Matsumoto G, Aihara K, Hanyu Y , Ta kahashi N, Yoshizawa S, Nagumo J-i. Chaos and phase locking in normal squid axons. Physics Letters A 123, 162- 166 (1987)

    1987

  21. [29]

    Criticality between cortical states

    Fontenele AJ , et al. Criticality between cortical states. Physical Review Letters 122, 208101 (2019)

    2019

  22. [30]

    Evidence for quasicritical brain dynamics

    Fosque LJ, Williams-García RV , Beggs JM, Ortiz G. Evidence for quasicritical brain dynamics. Physical Review Letters 126, 098101 (2021)

    2021

  23. [31]

    On redefining a snap-back repeller

    Marotto FR. On redefining a snap-back repeller. Chaos, Solitons & Fractals 25, 25-28 (2005)

    2005

  24. [32]

    Ising formulations of many NP problems

    Lucas A. Ising formulations of many NP problems. Frontiers in physics 2, 5 (2014)

    2014

  25. [33]

    Analysis of the relation between quad ratic unconstrained binary optimization and the spin-glass ground-state problem

    Boettcher S. Analysis of the relation between quad ratic unconstrained binary optimization and the spin-glass ground-state problem. Physical Review Research 1, 033142 (2019)

    2019

  26. [34]

    Semi-supervise d classification with graph convolutional networks

    Kipf TN, Welling M. Semi-supervise d classification with graph convolutional networks. In International Conference on Learning Representations, (2017)

    2017

  27. [35]

    I nductive representation learning on large graphs

    Hamilton W, Ying Z, Leskovec J. I nductive representation learning on large graphs. Advances in neural information processing systems, (2017)

    2017

  28. [36]

    Graph attention networks

    Veli čković P, Cucurull G, Casanova A, Romero A, Lio P, Bengio Y . Graph attention networks. arXiv preprint arXiv:171010903, (2017)

    2017

  29. [37]

    Brain-in spired chaotic backpropagation for MLP

    Tao P, Cheng J, Chen L. Brain-in spired chaotic backpropagation for MLP. Neural Networks 155, 1-13 (2022)

    2022

  30. [38]

    On a response char acteristic of a mathematical neuron model

    Nagumo J, Sato S. On a response char acteristic of a mathematical neuron model. Kybernetik 10, 155-164 (1972)

    1972

  31. [39]

    Chaotic neural networks

    Aihara K, Takabe T, Toyoda M. Chaotic neural networks. Physics Letters A 144, 333-340 (1990)

    1990

  32. [40]

    On the importance of initialization and momentum in deep learning

    Sutskever I, Martens J, Dahl G, Hint on G. On the importance of initialization and momentum in deep learning. Proceedings of the 30th International Conference on Machine Learning 28, 1139-1147 (2013)

    2013

  33. [41]

    Ex tremal cuts of sparse random graphs

    Amir D, Andrea M, Subhabrata S. Ex tremal cuts of sparse random graphs. The Annals of Probability 45, 1190-1217 (2017)

    2017

  34. [42]

    Solvable model of a spin-glass

    Sherrington D, Kirkpatrick S. Solvable model of a spin-glass. Physical review letters 35, 1792 (1975)

    1975

  35. [43]

    Im proved approximation algorithms for maximum cut and satisfiability proble ms using semidefinite programming

    Goemans MX, Williamson DP. Im proved approximation algorithms for maximum cut and satisfiability proble ms using semidefinite programming. Journal of the ACM (JACM) 42, 1115-1145 (1995)

    1995

  36. [44]

    Breakout lo cal search for the max-cutproblem

    Benlic U, Hao J-K. Breakout lo cal search for the max-cutproblem. Engineering Applications of Artificial Intelligence 26, 1162-1173 (2013)

    2013

  37. [45]

    Solving sparse semi definite programs using the dual scaling algorithm with an iterative solver

    Choi C, Ye Y . Solving sparse semi definite programs using the dual scaling algorithm with an iterative solver. Manuscript, Departm ent of Management Sciences, University of Iowa, Iowa City, IA 52242, (2000)

    2000

  38. [46]

    Solving large scale max cut problems via tabu search

    Kochenberger GA, Hao J-K, Lü Z, Wa ng H, Glover F. Solving large scale max cut problems via tabu search. Journal of Heuristics 19, 565-571 (2013)

    2013

  39. [47]

    Graph neural networks for maximum constraint satisfaction

    Toenshoff J, Ritzert M, Wolf H, Grohe M. Graph neural networks for maximum constraint satisfaction. Frontiers in artificial intelligence 3, 580607 (2021)

    2021

  40. [48]

    Graph coloring with physics-inspired graph neural networks

    Schuetz MJA, Brubaker JK, Zhu Z, Katzgraber HG. Graph coloring with physics-inspired graph neural networks. Physical Review Research 4, 043131 (2022)

    2022

  41. [49]

    Automating the construction of internet portals with machine learning

    McCallum AK, Nigam K, Rennie J, Se ymore K. Automating the construction of internet portals with machine learning. Information Retrieval 3, 127-163 (2000)

    2000

  42. [50]

    Collective classification in network data

    Sen P, Namata G, Bilgic M, Getoor L, Galligher B, Eliassi-Rad T. Collective classification in network data. AI magazine 29, 93-93 (2008)

    2008

  43. [51]

    Query-driven active surveying for collective classification

    Namata G, London B, Getoor L, Hu ang B, Edu U. Query-driven active surveying for collective classification. In 10th international workshop on mining and learning with graphs, 1 (2012)

    2012

  44. [52]

    Using tabu s earch techniques fo r graph coloring

    Hertz A, Werra Dd. Using tabu s earch techniques fo r graph coloring. Computing 39, 345-351 (1987)

    1987

  45. [53]

    Rethinking Graph Neural Networks for the Graph Coloring Problem

    Li W, Li R, Ma Y , Chan SO, Pan D, Yu B. Rethinking Graph Neural Networks for the Graph Coloring Problem. ArXiv abs/2208.06975, (2022)

    2022 arXiv

  46. [54]

    Back-propagation with chaos

    Fazayeli F, Wang L, Liu W. Back-propagation with chaos. International Conference on Neural Networks and Signal Processing, 5-8 (2008)

    2008

  47. [55]

    Self-backpropagation of synaptic modifications elevates the effici ency of spiking and artificial neural networks

    Zhang T, Cheng X, Jia S, Poo M-m, Zeng Y , Xu B. Self-backpropagation of synaptic modifications elevates the effici ency of spiking and artificial neural networks. Science Advances 7, eabh0146 (2021)

    2021

  48. [56]

    The forward-forward algor ithm: Some preliminary investigations

    Hinton G. The forward-forward algor ithm: Some preliminary investigations. arXiv preprint arXiv:221213345, (2022)

    2022

  49. [57]

    Adabelief optimizer: Adapting stepsizes by the belief in observed gradients

    Zhuang J , et al. Adabelief optimizer: Adapting stepsizes by the belief in observed gradients. Advances in neural inform ation processing systems , 18795-18806 (2020)

    2020

  50. [58]

    Adan: Adaptive Nesterov Momentum Algorithm for Faster Op timizing Deep Models

    Xie X, Zhou P, Li H, Lin Z, Yan S. Adan: Adaptive Nesterov Momentum Algorithm for Faster Op timizing Deep Models. arXiv preprint arXiv:220806677, (2022)

    2022

  51. [59]

    Sparse graph attention networks

    Ye Y , Ji S. Sparse graph attention networks. IEEE Transactions on Knowledge and Data Engineering 35, 905-916 (2021)

    2021

  52. [60]

    Deep graph library: Towards efficient and scalable deep learning on graphs

    Wang MY . Deep graph library: Towards efficient and scalable deep learning on graphs. In ICLR workshop on representation learning on graphs and manifolds, (2019)

    2019

  53. [61]

    Rectified lin ear units improve restricted boltzmann machines

    Nair V , Hinton GE. Rectified lin ear units improve restricted boltzmann machines. International Conference on Machine Learning, 807-814 (2010)

    2010

  54. [62]

    Dropout: a simple way to prevent neur al networks from overfitting

    Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neur al networks from overfitting. The journal of machine learning research 15, 1929-1958 (2014)

    2014

  55. [63]

    Batch normaliza tion: Accelerating deep network training by reducing internal covariate shift

    Ioffe S, Szegedy C. Batch normaliza tion: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, 448-456 (2015)

    2015

  56. [64]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke A , et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural info rmation processing systems 32, 8026-8037 (2019)

    2019

  57. [65]

    Making a science of model search: Hyperparameter optimi zation in hundreds of di mensions for vision architectures

    Bergstra J, Yamins D, Cox D. Making a science of model search: Hyperparameter optimi zation in hundreds of di mensions for vision architectures. In International conference on machine learning, 115-123 (2013)

    2013

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.