OscNet v1.5: Energy Efficient Hopfield Network on CMOS Oscillators for Image Classification

Iris Wang; Thomas H. Lee; Wenxiao Cai; Yu-Neng Wang; Zongru Li

REVIEW 5 major objections 5 minor 26 references

OscNet v1.5: Energy Efficient Hopfield Network on CMOS Oscillators for Image Classification

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

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

Pith's one-line read OscNet v1.5 claims that a forward-only, sparsely connected Hopfield network can classify binarized MNIST digits and map onto CMOS oscillator arrays with negligible accuracy loss.

desk verdict A forward-only sparse Hopfield pipeline with a plausible sparsification rule, but the headline numbers are overstated and the oscillator convergence claim is unsupported. read the letter →

arxiv 2506.12610 v2 pith:ZDMNRB3O submitted 2025-06-14 cs.CV

Wenxiao Cai , Zongru Li , Iris Wang , Yu-Neng Wang , Thomas H. Lee This is my paper

classification cs.CV

keywords OscNetHopfieldnetworkCMOSoscillatorarrayssubharmonicinjectionlockingHebbianlearningsparseconnectivityKuramotodynamicsMNISTimageclassification

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

This paper tries to establish that image classification can be done by a Hopfield-style network whose weights are learned by local Hebbian updates alone, then pruned to a sparse set, and whose inference runs as coupled-oscillator phase dynamics rather than digital backpropagation. The authors report that this forward-only pipeline beats a backpropagation-trained autoencoder baseline on binarized MNIST, and that in oscillator-based inference only 24% of the connections of a fully connected Hopfield network are needed, with just a 0.1% drop in accuracy. The point matters because the method is designed around CMOS ring-oscillator arrays, where the phase of each oscillator encodes a neuron state and coupling is implemented by programmable resistive elements or transmission gates, offering a plausible low-power path for edge inference.

What carries the argument

The carrying objects are the class-specific similarity matrices $W^{(k)}$, the row-wise pruning rule that keeps each pixel's top-$N_2$ most informative connections, and the modified Kuramoto oscillator update $d\theta_i/dt = -\epsilon_i \sin(2\theta_i) - \sum_j w_{ij}\sin(\theta_i-\theta_j)$, where the first term is subharmonic injection locking (SHIL) that stabilizes each phase to $0$ or $\pi$. The final spin state is decoded as $s_i = \operatorname{sign}[\cos(\theta_i)]$, and the class with the lowest Ising-form energy $E = -\sum_{i,j} w_{ij}s_i s_j$ is chosen, tying the continuous oscillator dynamics to the discrete Hopfield attractor picture.

What would settle it

Simulate Eq. (12) with weights trained on two digits, prune to 24% connectivity as in Section 3.2, initialize with every MNIST test image plus noise, and check whether the final phases always settle in the energy minimum of the correct class; a single case where a wrong class has lower final energy, or where the dynamics cycle without converging, would falsify the convergence claim.

Watch

Extended reading notes

Core claim

The central claim is that a Hopfield network trained with the Hebbian rule $w^{(k)}_{ij} = \sum_{\text{images}\, \in\, C_k} p_i p_j$, followed by entropy-based normalization and per-pixel top-$N_2$ pruning, stores each digit class as a low-energy attractor that can be reached by continuous Kuramoto dynamics $d\theta_i/dt = -\epsilon_i \sin(2\theta_i) - \sum_j w_{ij}\sin(\theta_i-\theta_j)$. The authors show that this learned sparse representation outperforms the raw image as a feature for a shallow linear classifier, and that oscillator-based inference using only 24% of the connections reaches 75.3% accuracy versus 75.4% for the fully connected version. Classification selects the class whose Hopfield energy $E = -\sum_{i,j} w_{ij} s_i s_j$ is minimized after the oscillator phases converge, with subharmonic injection locking (SHIL) pinning phases to binary values.

Load-bearing premise

The paper assumes that after pruning, the sparse and no longer symmetric weight matrix still makes the oscillator dynamics converge to the intended class attractor, even though the cited energy-minimization guarantee only covers symmetric couplings and no proof is given for the pruned case.

Editorial extensions

If this is right

Training requires only forward, local Hebbian updates, so no backpropagation is needed on the learning side.
Sparse connectivity reduces the number of analog couplings, the most area- and power-intensive components in oscillator-array hardware.
The same phase dynamics that classifies also denoises: a noisy test image is pushed toward a clean stored attractor during inference.
SHIL lets the continuous Kuramoto dynamics be read out as discrete Hopfield states, bridging analog oscillator computation and binary classification.
Pairwise elimination extends the binary classifier to all ten MNIST digits while keeping the hardware logic minimal and highly parallelizable.

Reading between the lines

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

The 24%-connectivity claim is measured on binarized MNIST at a fixed resolution; whether the same sparsity ratio survives on higher-resolution or multi-channel inputs is open, since per-pixel top-$N_2$ pruning does not obviously transfer across image sizes.
After pruning, the weight matrix is generally asymmetric, and the energy-minimization guarantee cited for symmetric Kuramoto couplings does not automatically extend; checking whether any Lyapunov function exists for the pruned dynamics would turn the empirical 0.1% drop into a structural property.
The tournament-style 10-class extension makes hard early decisions: one wrong binary comparison eliminates the true class, so end-to-end accuracy on all ten digits may degrade faster than the binary-case numbers suggest.
A direct hardware measurement of phase noise under SHIL, comparing decoded digits to the ODE-solver simulation, would show whether the quoted accuracy holds outside a numerical solver.

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.

Desk Editor's Note

read the letter

The genuinely new piece here is the Acc-TopABS sparsification rule: per-pixel top-k selection on Hebbian weights, which buys accuracy at reduced connectivity. The ablation tables are useful, and the forward-only training idea is a plausible match for oscillator hardware. The authors also released code, which is real evidence. But the paper oversells its results. The abstract's '8% improvement' is against a 73.8% autoencoder baseline, not a strong deep network. The 'competitive accuracy' is on binarized MNIST, and the actual oscillator-based inference (Fig. 8) gets only 75.3% accuracy at 24% connectivity—0.1% below a fully connected network that itself is only 75.4%. That is not competitive by any standard. There is also a sign error in Eq. 11: as written, E_k = Σ w p'_i p'_j is the negative of the Ising energy used in Eq. 14, yet the text says agreement lowers the energy. The normalization step in Sec. 3.3 is never defined, and the sparsity level appears to be selected on the test set. The most serious problem is the convergence claim. Section 3.2 explicitly says the sparse connectivity pattern need not be symmetric, but the energy-minimization guarantee from [8] (Eq. 8) applies to symmetric couplings. Equation 12 with asymmetric w_ij is not a gradient system for Eq. 14, so the statement that the system 'converges naturally to attractor states' is unsupported. The 75.3% figure may be an artifact of ODE trajectories initialized near the test image rather than true attractors that hardware would reach. The stress-test concern lands. None of these flaws are unfixable. The sparsification idea is worth testing further, and the forward-only pipeline is interesting for low-power hardware researchers. But the paper needs major revision: correct the sign, define the normalization, add error bars, report which data split was used for sparsity selection, and either prove convergence for asymmetric weights or re-frame the claim without the energy-minimization language. This deserves peer review because the idea is plausible and the flaws are fixable, but I would not cite it in its current form. For a reading group, it is a good example of why engineering papers should separate CPU simulation results from hardware inference claims.

Referee Report

5 major / 5 minor

Summary. The paper proposes OscNet v1.5, a sparse Hopfield-network classifier for binarized MNIST intended for CMOS oscillator hardware. Training uses one Hopfield network per class with Hebbian forward-only updates, then prunes each pixel to its top-N2 connections. At inference, a test image is mapped to a phase vector and evolved by a modified Kuramoto equation (Eq. 12) with subharmonic injection locking; classification selects the class with minimum energy. Experiments on binary MNIST report roughly 96% accuracy when the learned features are passed to a linear classifier, and 75.3% accuracy at 24% connectivity in the oscillator-style inference, described as only 0.1% below the fully connected version. The paper also sketches a CMOS transmission-gate implementation of signed couplings.

Significance. If established, the forward-only sparse learning rule plus oscillator-based inference could be a genuinely low-power edge-inference scheme, and the public code repository is a useful starting point for reproduction. The paper also includes useful ablations over sparsity level and normalization and a concrete coupling schematic for CMOS oscillators. However, the current evidence does not establish the central claims: the convergence argument for the asymmetric sparse dynamics is missing, the energy formula in Eq. 11 has a sign inconsistency, the normalization procedure is not specified, the experiments lack error bars and a validation protocol, and the headline '8% improvement' is not supported by the numbers in Table 3.

major comments (5)

[§3.4, Eq. (12)] The dynamics in Eq. (12) are described in §3.4 as converging 'naturally to attractor states,' with Eq. (14) serving as the minimized energy. However, the coupling matrix in Eq. (12) is the sparse matrix from §3.2, which the paper explicitly states 'need not be symmetric' because the per-pixel top-N2 selection breaks w_ij = w_ji. The Lyapunov-energy argument quoted from [8] (Eq. 8) applies to symmetric couplings K_ij. For asymmetric couplings, the ODE is not the gradient descent of a single global energy and can exhibit limit cycles rather than fixed-point convergence. Thus the global-minimum interpretation is unsupported, and this is load-bearing for the 24%-connectivity/0.1%-drop result in Fig. 8, which is presented as evidence about oscillator hardware. Please either restrict the sparse pattern to symmetric couplings, provide a convergence proof for the asymmetric case, or reframe the inference as an empirical behavior and remove the global-energy claim. Note also that Eq. (15) is symmetric, but if the sparsification masks from §3.2 are applied, the asymmetry issue reappears.
[§3.1, Eq. (11)] Eq. (11) defines E_k = Σ w_ij p'_i p'_j and the surrounding text says that agreement between p'_i p'_j and w_ij 'lowers the energy,' after which the class with the lowest energy is selected. With the plus sign, agreement increases E_k, so a minimum-energy selection would favor anti-correlated classes. This contradicts the Hopfield energy in Eq. (4), the Ising energy in Eq. (14), and the subsequent discussion. The formula should be E_k = -Σ w_ij p'_i p'_j, or the selection rule should be changed to maximum energy. Since Tables 2 and 4 depend on this rule, the experiments need to be repeated or re-reported with the corrected convention.
[§3.3, Table 4] The normalization step is central to the reported performance: without normalization the accuracy is 15.49%, and with normalization over 'Dim 1,2' it is 86.32% (Table 4). However, no normalization formula is given. The only description is 'After calculating the entropy of each connection,' and no entropy definition appears anywhere in the paper. Table 4's 'Dim 1' and 'Dim 2' are also not defined. The normalization is therefore not reproducible, and because it changes accuracy by about 71 percentage points, it is load-bearing. Please specify the exact operations, including how the normalization constants are computed and whether they are derived from training, validation, or test data.
[Abstract, Table 3] The abstract's claim of 'an 8% improvement in accuracy compared to conventional deep learning models on MNIST' is not supported by the numbers in Table 3. The best reported BackProp autoencoder accuracy is 91.35% (#Hidden=128) against OscNet's 96.08%; the differences at the other hidden sizes are 21.60, 14.81, 10.60, and 4.73 percentage points. None of these equals 8%, and no conventional deep-learning MNIST baseline (e.g., a CNN) is included. Similarly, 'competitive accuracy on MNIST' is misleading because the experiments use binarized MNIST with accuracies around 96%, whereas standard MNIST classifiers exceed 99%. Please state the exact baseline and metric used for the 8% claim and temper the 'competitive' wording.
[§4, Tables 2, 4 and Fig. 8] Tables 2, 4, and Fig. 8 are reported without error bars, number of runs, or a validation split. The sparsity level n and the normalization choice appear to be selected from the test-set accuracies in these tables, so the '0.1% drop at 24% connectivity' and the '1.5% drop at 13% connectivity' are point estimates that partly reflect post-hoc model selection. In addition, Table 2 reports a fully connected accuracy of 78.65%, while Fig. 8's fully connected baseline is about 75.4%; the relationship between the two protocols is not explained. Please report mean ± standard deviation over repeated runs with a fixed train/validation/test split, and clarify which protocol each number refers to.

minor comments (5)

[§4 and §6] There are several typos: 'nonlieanr' in Section 4, 'oscill-tor implemention' in Section 6, and 'grobal' in reference [19] should be 'global.'
[Fig. 8] The axes of Fig. 8 are not labeled. Please specify whether the x-axis is the fraction of retained connections and whether the y-axis is the accuracy on the binarized MNIST test set.
[§3.4, Eq. (12)] Eq. (12) does not state the initial conditions for θ_i(t) nor the value(s) of ϵ_i used in the experiments. Please report the ODE solver settings (method, tolerances, integration time) and the SHIL strength.
[§4.1] The pairwise-elimination section does not specify how the 45 pairwise decisions are combined into a final 10-class prediction (e.g., tournament bracket, all-pairs voting, or a fixed elimination order). Please state the decision rule.
[§3.2, Table 2] The notation 'n × 784 connections' is ambiguous because n is used for both the number of connections per pixel and the number of classes. Please disambiguate, and also complete the 'fully connected' row of Table 2, which has no entry in the Acc-TopABS column.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: accuracy claims are measured and no equation-level reduction is present.

full rationale

The paper's training and inference chain is self-contained. Hebbian weights are accumulated from training pixel pairs (Eq. 10), test energies are evaluated on test images (Eq. 11), and classification selects the minimum-energy class. Sparse selection (Sec. 3.2) and normalization (Sec. 3.3) are supervised preprocessing choices, not fitted parameters renamed as predictions; the reported accuracies, including the 0.1% drop at 24% connectivity and the 8% improvement, are measured results rather than derived predictions. The oscillator convergence claim in Sec. 3.4 relies on the Lyapunov/energy-minimization result cited from [8, 9], which include two current co-authors, but this is a standard symmetric-coupling result and is not used to derive the numeric accuracy figures, which are empirical. The cited prior work is externally published and therefore counts as independent support under the review rules. A genuine correctness risk is that the pruned weight matrix is acknowledged to be asymmetric, so the symmetric-coupling Lyapunov guarantee from Eq. 8 does not automatically apply to Eq. 12; however, this is an unsupported assumption, not a circular reduction, because the paper does not define convergence in terms of the reported accuracy and no equation is equivalent to its input by construction. Test-set-based selection of normalization and sparsity levels (Tables 2, 4; Fig. 8) is an experimental methodology concern rather than circularity under the specified criteria.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are proposed. The main free choices are the sparsity level, normalization scheme, and unspecified dynamics parameters; these are tuned on data. The 'entropy' normalization is described but its formula is never given.

free parameters (3)

sparsity level n (connections per pixel) = 84 (best on test set, searched over 10,16,28,56,64,84,128,256)
Table 2 shows accuracy peaks at 84 connections; no validation split is described, so the value is selected on test performance.
normalization choice = normalize over both class (Dim1) and pixel (Dim2) dimensions
Table 4 ablation shows Dim1+Dim2 gives the highest direct classification accuracy; the normalization formula is not given.
SHIL strength epsilon_i and ODE solver settings = unspecified
Section 3.4 mentions epsilon_i but no values or solver tolerances; these would affect oscillator convergence and final accuracy.

assumptions (4)

domain assumption Hebbian weight update w_ij = sum p_i p_j over training images of a class stores class-correlation information.
Used in Eq. 10; standard in Hopfield networks but an assumption about how images are represented.
ad hoc to paper The energy E_k = sum w_ij p'_i p'_j (Eq. 11) ranks classes; lower energy means better match.
The sign of Eq. 11 contradicts the text 'if signs agree it lowers the energy'; the classifier logic depends on an energy sign convention that is not stated consistently.
domain assumption Kuramoto dynamics with SHIL (Eq. 12) converge to binary spin states.
Standard for injection-locked oscillators but the convergence to the global minimum of the Ising energy is not proven for learned sparse weights.
domain assumption Pixel values are binarized and each class has exactly one Hopfield network.
Section 3.1; the single-pattern-per-network assumption in Section 2.3 is contradicted by storing many training images per class.

how reviews work

0 comments

Cite this review

Pith. "Pith review of OscNet v1.5: Energy Efficient Hopfield Network on CMOS Oscillators for Image Classification." pith.science (2026). https://pith.science/paper/ZDMNRB3O

@misc{pith2026250612610,
  author       = {Pith},
  title        = {Pith review of: OscNet v1.5: Energy Efficient Hopfield Network on CMOS Oscillators for Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDMNRB3O}},
  note         = {Machine review of arXiv:2506.12610}
}

read the original abstract

Machine learning has achieved remarkable advancements but at the cost of significant computational resources. This has created an urgent need for a novel and energy-efficient computational fabric and corresponding algorithms. CMOS Oscillator Networks (OscNet) is a brain inspired and specially designed hardware for low energy consumption. In this paper, we propose a Hopfield Network based machine learning algorithm that can be implemented on OscNet. The network is trained using forward propagation alone to learn sparsely connected weights, yet achieves an 8% improvement in accuracy compared to conventional deep learning models on MNIST dataset. OscNet v1.5 achieves competitive accuracy on MNIST and is well-suited for implementation using CMOS-compatible ring oscillator arrays with SHIL. In oscillator-based inference, we utilize only 24% of the connections used in a fully connected Hopfield network, with merely a 0.1% drop in accuracy. OscNet v1.5 relies solely on forward propagation and employs sparse connections, making it an energy-efficient machine learning pipeline designed for oscillator computing fabric. The repository for OscNet family is: https://github.com/RussRobin/OscNet .

Figures

Figures reproduced from arXiv: 2506.12610 by the authors.

**Figure 2.** Features learned on three points of the dataset, as shown [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 6.** Left: Input image (noisy digit 0). Right: Reconstructed [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

Figures from the paper (3 more)

**Figure 5.** Figure 5: Energy evolution during convergence. The system [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 7.** Figure 7: Pairwise classification results for five MNIST test digits using Kuramoto dynamics. The first column shows the original image. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Binary MNIST image classification accuracy of oscilla [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

[8]

Smith and Thomas H

Richelle L. Smith and Thomas H. Lee. Polychronous oscillatory cellular neural networks for solving graph coloring problems. IEEE Open Journal of Circuits and Systems, 4:156–164, 2023. 1, 2, 3

work page 2023
[1]

Spatialbot: Precise spatial understanding with vision language models

Wenxiao Cai, Iaroslav Ponomarenko, Jianhao Yuan, Xiaoqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642 ,

arXiv
[2]

Machine learning in finance, volume 1170

Matthew F Dixon, Igor Halperin, and Paul Bilokon. Machine learning in finance, volume 1170. Springer, 2020

work page 2020
[3]

Machine learning in healthcare

Hafsa Habehh and Suril Gohel. Machine learning in healthcare. Current genomics, 22(4):291, 2021. 1

work page 2021
[4]

Power hungry processing: Watts driving the cost of ai deployment? In The 2024 ACM Conference on Fair- ness, Accountability, and Transparency, pages 85–99,

Sasha Luccioni, Yacine Jernite, and Emma Strubell. Power hungry processing: Watts driving the cost of ai deployment? In The 2024 ACM Conference on Fair- ness, Accountability, and Transparency, pages 85–99,

work page 2024
[5]

Evaluating the carbon foot- print of nlp methods: a survey and analysis of existing tools

Nesrine Bannour, Sahar Ghannay, Aur ´elie N ´ev´eol, and Anne-Laure Ligozat. Evaluating the carbon foot- print of nlp methods: a survey and analysis of existing tools. In Proceedings of the second workshop on sim- ple and efficient natural language processing , pages 11–21, 2021

work page 2021
[6]

Reducing the carbon impact of generative ai infer- ence (today and in 2035)

Andrew A Chien, Liuzixuan Lin, Hai Nguyen, Var- sha Rao, Tristan Sharma, and Rajini Wijayawardana. Reducing the carbon impact of generative ai infer- ence (today and in 2035). In Proceedings of the 2nd workshop on sustainable computer systems , pages 1– 7, 2023

work page 2023
[7]

Compute and energy con- sumption trends in deep learning inference

Radosvet Desislavov, Fernando Mart´ınez-Plumed, and Jos´e Hern ´andez-Orallo. Compute and energy con- sumption trends in deep learning inference. arXiv preprint arXiv:2109.05472, 2021. 1

arXiv 2021

Show all 26 references

[9]

Osc- net: Machine learning on cmos oscillator networks

Wenxiao Cai, Zongru Li, and Thomas H Lee. Osc- net: Machine learning on cmos oscillator networks. IEEE Transactions on Circuits and Systems II: Ex- press Briefs, 2025. 1, 2, 3, 5, 6

2025
[10]

Hyperbolic hopfield neural networks for image classification in content-based image re- trieval

K Anitha, R Dhanalakshmi, K Naresh, and D Ruk- mani Devi. Hyperbolic hopfield neural networks for image classification in content-based image re- trieval. International Journal of Wavelets, Multires- olution and Information Processing, 19(01):2050059,
[11]

Training energy-based single-layer hopfield and oscillatory net- works with unsupervised and supervised algorithms for image classification

Madeleine Abernot and Aida Todri-Sanial. Training energy-based single-layer hopfield and oscillatory net- works with unsupervised and supervised algorithms for image classification. Neural Computing and Ap- plications, 35(25):18505–18518, 2023. 1

2023
[12]

Hebb and darwin

Paul Adams. Hebb and darwin. Journal of theoretical Biology, 195(4):419–438, 1998. 1

1998
[13]

Adaptive pattern classification and universal recoding: I

Stephen Grossberg. Adaptive pattern classification and universal recoding: I. parallel development and coding of neural feature detectors. Biological cyber- netics, 23(3):121–134, 1976. 1

1976
[14]

Gradient-based learning applied to document recognition

Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE , 86(11):2278–2324, 1998. 2, 4, 5, 6

1998
[15]

Hopfield networks is all you need

Hubert Ramsauer, Bernhard Sch ¨afl, Johannes Lehner, Philipp Seidl, Michael Widrich, Thomas Adler, Lukas Gruber, Markus Holzleitner, Milena Pavlovi ´c, Geir Kjetil Sandve, et al. Hopfield networks is all you need. arXiv preprint arXiv:2008.02217, 2020. 2

2008 arXiv
[16]

The application of competitive hopfield neural net- work to medical image segmentation

Kuo-Sheng Cheng, Jzau-Sheng Lin, and Chi-Wu Mao. The application of competitive hopfield neural net- work to medical image segmentation. IEEE transac- tions on medical imaging, 15(4):560–567, 1996. 2

1996
[17]

Izhikevich and Frank C

Eugene M. Izhikevich and Frank C. Hoppensteadt. Polychronous wavefront computations. Int. J. Bifurc. Chaos, 19:1733–1739, 2009. 2

2009
[18]

A general theory of injection locking and pulling in electrical oscilla- tors—part i: Time-synchronous modeling and injec- tion waveform design

Brian Hong and Ali Hajimiri. A general theory of injection locking and pulling in electrical oscilla- tors—part i: Time-synchronous modeling and injec- tion waveform design. IEEE Journal of Solid-State Circuits, 54:2109–2121, 2019. 2

2019
[19]

Local and grobal self- entrainments in oscillator lattices

Hidetsugu Sakaguchi, Shigeru Shinomoto, and Yoshiki Kuramoto. Local and grobal self- entrainments in oscillator lattices. Progress of Theoretical Physics, 77:1005–1010, 1987. 2, 4

1987
[20]

F. Y . Wu. The potts model. Reviews of Modern Physics, 54:235–268, 1982. 3

1982
[21]

Shem: A hardware- aware optimization framework for analog computing systems

Yu-Neng Wang and Sara Achour. Shem: A hardware- aware optimization framework for analog computing systems. arXiv preprint arXiv:2411.03557, 2024. 4

2024 arXiv
[22]

Ising machines as hardware solvers of com- binatorial optimization problems

Naeimeh Mohseni, Peter L McMahon, and Tim Byrnes. Ising machines as hardware solvers of com- binatorial optimization problems. Nature Reviews Physics, 4(6):363–379, 2022. 4

2022
[23]

Classification of hand- written digits using the hopfield network

MA Belyaev and AA Velichko. Classification of hand- written digits using the hopfield network. In IOP conference series: materials science and engineering, volume 862, page 052048. IOP Publishing, 2020. 4

2020
[24]

A 1,968-node coupled ring oscillator circuit for combi- natorial optimization problem solving

William Moy, Ibrahim Ahmed, Po-wei Chiu, John Moy, Sachin S Sapatnekar, and Chris H Kim. A 1,968-node coupled ring oscillator circuit for combi- natorial optimization problem solving. Nature Elec- tronics, 5(5):310–317, 2022. 6

2022
[25]

An ising solver chip based on coupled ring oscillators with a 48-node all-to- all connected array architecture

Hao Lo, William Moy, Hanzhao Yu, Sachin Sapat- nekar, and Chris H Kim. An ising solver chip based on coupled ring oscillators with a 48-node all-to- all connected array architecture. Nature Electronics, 6(10):771–778, 2023. 6, 7

2023
[26]

Hardware acceleration of sparse and irregular tensor computations of ml models: A survey and insights

Shail Dave, Riyadh Baghdadi, Tony Nowatzki, Sasikanth Avancha, Aviral Shrivastava, and Baoxin Li. Hardware acceleration of sparse and irregular tensor computations of ml models: A survey and insights. Proceedings of the IEEE , 109(10):1706–1752, 2021. 7

2021

Pith tools

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

[1] [8]

Smith and Thomas H

Richelle L. Smith and Thomas H. Lee. Polychronous oscillatory cellular neural networks for solving graph coloring problems. IEEE Open Journal of Circuits and Systems, 4:156–164, 2023. 1, 2, 3

work page 2023

[2] [1]

Spatialbot: Precise spatial understanding with vision language models

Wenxiao Cai, Iaroslav Ponomarenko, Jianhao Yuan, Xiaoqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642 ,

arXiv

[3] [2]

Machine learning in finance, volume 1170

Matthew F Dixon, Igor Halperin, and Paul Bilokon. Machine learning in finance, volume 1170. Springer, 2020

work page 2020

[4] [3]

Machine learning in healthcare

Hafsa Habehh and Suril Gohel. Machine learning in healthcare. Current genomics, 22(4):291, 2021. 1

work page 2021

[5] [4]

Power hungry processing: Watts driving the cost of ai deployment? In The 2024 ACM Conference on Fair- ness, Accountability, and Transparency, pages 85–99,

Sasha Luccioni, Yacine Jernite, and Emma Strubell. Power hungry processing: Watts driving the cost of ai deployment? In The 2024 ACM Conference on Fair- ness, Accountability, and Transparency, pages 85–99,

work page 2024

[6] [5]

Evaluating the carbon foot- print of nlp methods: a survey and analysis of existing tools

Nesrine Bannour, Sahar Ghannay, Aur ´elie N ´ev´eol, and Anne-Laure Ligozat. Evaluating the carbon foot- print of nlp methods: a survey and analysis of existing tools. In Proceedings of the second workshop on sim- ple and efficient natural language processing , pages 11–21, 2021

work page 2021

[7] [6]

Reducing the carbon impact of generative ai infer- ence (today and in 2035)

Andrew A Chien, Liuzixuan Lin, Hai Nguyen, Var- sha Rao, Tristan Sharma, and Rajini Wijayawardana. Reducing the carbon impact of generative ai infer- ence (today and in 2035). In Proceedings of the 2nd workshop on sustainable computer systems , pages 1– 7, 2023

work page 2023

[8] [7]

Compute and energy con- sumption trends in deep learning inference

Radosvet Desislavov, Fernando Mart´ınez-Plumed, and Jos´e Hern ´andez-Orallo. Compute and energy con- sumption trends in deep learning inference. arXiv preprint arXiv:2109.05472, 2021. 1

arXiv 2021