Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Is Quantum Optimization Ready? An Effort Towards Neural Network Compression using Adiabatic Quantum Computing

T0 review · 3 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 This paper claims that joint pruning and quantization of CNNs can be cast as a QUBO problem and solved by adiabatic quantum computers, delivering under 1% accuracy loss at up to 97% compression, faster than genetic and reinforcement…

desk verdict A genuinely new QUBO formulation for joint pruning-quantization, but the empirical comparison is not controlled and test-set leakage makes the headline claims premature. read the letter →

arxiv 2505.16332 v2 pith:IRJTDQIE submitted 2025-05-22 quant-ph cs.AIcs.PF

Zhehui Wang , Benjamin Chen Ming Choong , Tian Huang , Daniel Gerlinghoff , Rick Siow Mong Goh , Cheng Liu , Tao Luo This is my paper · ORCID
classification quant-phcs.AIcs.PF
keywords quantumannealingadiabaticcomputingQUBOneuralnetworkcompressionpruningquantizationmodelD-Wave
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 adiabatic quantum computing (AQC) can be applied to a real deep-learning task—joint filter/channel pruning and bit-width quantization of convolutional networks—by reformulating the compression objective strictly as a quadratic unconstrained binary optimization (QUBO) problem. The authors show that on small CNN benchmarks, annealing hardware finds compression solutions that reduce model size by roughly 88-97% while keeping accuracy loss under 1% after fine-tuning. They also report that their quantum annealing approach reaches these solutions faster than genetic algorithm and reinforcement learning baselines, and they argue that the QUBO formulation is the key enabling step. The paper is an exploratory demonstration rather than a scaling claim: it candidly notes that qubit count and connectivity currently limit the method to small models. A sympathetic reader would take away that quantum optimization is ready for small, structured compression problems today, with the path to larger models blocked mainly by hardware, not by the formulation.

What carries the argument

The load-bearing object is the QUBO Hamiltonian $H = L_p(P) + \beta L_q(Q) - \gamma R(P,Q)$ with binary variables $p_{n,i}$ (prune filter or channel $i$ in layer $n$) and $q_{n,k}$ ($k$-th bit of the bit-width reduction for layer $n$), assembled into $H = x^T U x$ with all coefficients precomputed from the pretrained weights. Pruning loss $L_p$ squares the layer-wise $\ell^1$ weight magnitude, quantization loss $L_q$ squares the base-2 logarithm of the RMSE proxy, and the reduction term $R$ counts saved bits while penalizing pruned-quantized combinations; $\beta$ and $\gamma$ trade off loss versus compression. The companion mechanism is the iterative hyperparameter search (Algorithms 1-3) that initializes $\beta = \|A\|_1/\|B\|_1$ and then binary-searches the largest $\gamma$ and smallest $\beta$ that keep post-fine-tuning accuracy above a threshold. The QUBO structure is what lets the annealing hardware treat compression as a ground-state search, and the search algorithm is what converts raw annealing samples into a usable accuracy-compression point.

What would settle it

Take a small trained CNN, enumerate all compression solutions near the optimum, and compare the QUBO energy H of pairs (A, B) with the accuracy each achieves after identical fine-tuning; if a sizable fraction of pairs with H(B) < H(A) nevertheless give accuracy(B) < accuracy(A), the surrogate objective is not aligned with the stated goal. A simpler check is to run the same QUBO problem on a classical exact solver for a small model and confirm that the lowest-energy configuration is also the one with the best fine-tuned accuracy.

Watch

Extended reading notes

Core claim

The central claim is that joint pruning-quantization of CNN weights can be expressed as a strictly quadratic, unconstrained binary objective, $H = x^T U x$, whose ground state encodes the optimal pruning mask and per-layer quantization bit-widths. The paper constructs $H$ from three energy components: a squared $\ell^1$-weight-magnitude pruning loss, a squared logarithmic RMSE quantization loss built on the learned step-size rule $\mathrm{RMSE}_n \approx c_n \cdot 2^{\sum_k 2^k q_{n,k}}$, and a compression-rate reward that couples pruning and quantization variables. On LeNet-5 (MNIST), GTSR-CNN (GTSRB), ResNet-9 (CIFAR-10), and VGG-16 (CIFAR-100), the authors report that solutions returned by the D-Wave Advantage and Fujitsu Digital Annealer keep accuracy drop below 1% at 88.4-96.8% model reduction, and that in head-to-head runs the annealing approach matches or exceeds genetic and reinforcement learning baselines while using far less optimization time. The paper's own experiments also reveal that hardware capacity, not the QUBO formulation, is the current bottleneck: the D-Wave device could not embed ResNet-9's 2264-variable or VGG-16's 4263-variable problems.

Load-bearing premise

The entire scheme presumes that the two surrogate costs it sends to the annealer, weight magnitude for pruning and the RMSE step-size proxy for quantization, actually track the real test error after fine-tuning, because the annealer only sees the surrogates and never sees the network's accuracy.

Editorial extensions

If this is right

  • If the surrogate losses are faithful, quantum annealing gives an automated, near-millisecond compression optimizer that beats GA and RL on small CNNs, making it a practical design knob for edge deployments.
  • The QUBO formulation extends to channel and per-weight granularity as written, so scaling to finer pruning on larger models becomes a hardware question, not a reformulation question.
  • The demonstrated under-1% accuracy drop at 88-97% reduction on MNIST, GTSRB, CIFAR-10, and CIFAR-100 benchmarks sets a concrete bar that classical automated compression methods must beat on the same problem sizes.
  • The hyperparameter search algorithm's success suggests that $\beta = \|A\|_1/\|B\|_1$ is a generally useful starting balance between pruning and quantization loss for magnitude-based compression.

Reading between the lines

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

  • A natural next step would be to replace the weight-magnitude proxy with a second-order (Hessian-based) sensitivity that can still be expanded into QUBO coefficients, potentially aligning the annealer's objective with fine-tuned accuracy more closely.
  • The millisecond optimization-time comparison is for the QPU call only; end-to-end time including cloud queuing and one-epoch fine-tuning per search step is dominated by classical overhead, so the practical speedup claim is narrower than it appears.
  • The same QUBO machinery could be applied to other structured compression decisions, like layer-wise bit allocation in transformers or mixed-precision assignment, wherever the cost can be written as a quadratic binary form.
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

3 major / 5 minor

Summary. The paper reformulates joint filter/channel pruning and quantization of CNNs as a QUBO problem, with the Hamiltonian H = L_p + beta L_q - gamma R in Eq. (7) and its final QUBO form in Eq. (14). Pruning loss is based on squared weight-magnitude surrogates (Eqs. (8)-(9)), quantization loss on a log-squared RMSE approximation (Eqs. (10)-(11)), and the reduction term on bit savings (Eqs. (12)-(13)). The authors propose a hyperparameter search over beta and gamma (Algorithm 1), run the resulting QUBOs on the D-Wave Advantage and the Fujitsu Digital Annealer, and report accuracy and reduction ratios for LeNet-5, GTSR-CNN, ResNet-9, and VGG-16 in Table 5. They also compare against GA, RL, and PyTorch pruning/quantization tools in Figure 6, and conclude that AQC is effective for small-scale compression and outperforms classical methods in time and in reaching global optima.

Significance. If the claims were fully established, this would be a useful demonstration of AQC for a real neural-network-design task, and the explicit QUBO coefficients would be a reusable artifact. The paper has genuine strengths: the QUBO mapping is written out with precomputable coefficients, two real annealing platforms are used, and the method handles both filter- and channel-level pruning as well as quantization. However, the significance is currently limited by the small model sizes and, more importantly, by evaluation choices that make the headline performance claims difficult to interpret. The paper does not ship machine-checked proofs or fully reproducible code, and the central empirical comparison is not controlled, so the strength of the contribution depends on additional validation experiments.

major comments (3)
  1. [§5, Eqs. (7)–(14), Table 5] The load-bearing claim that AQC can achieve effective compression rests entirely on the surrogates in H: the weight-magnitude pruning loss (Eqs. (8)-(9)) and the RMSE quantization loss (Eqs. (10)-(11)). The paper never checks whether decreasing H along the optimizer's search path decreases post-fine-tuning test error. Without this check, a solution that minimizes H may still be a poor compression choice, so the accuracy drops in Table 5 do not demonstrate that the QUBO objective tracks the real compression objective. I request a direct validation experiment: sample a set of pruning/quantization configurations, compute H for each, fine-tune them briefly, and report the rank correlation (e.g., Spearman) between H and validation/test accuracy, together with the resulting Pareto front.
  2. [§6, Algorithm 1; §7.2] The hyperparameter search uses test accuracy as both the validity criterion and the selection signal: acc(·) fine-tunes for one epoch and evaluates on the test dataset (Algorithm 1, lines 3, 8, and 18), and the same test set is used to report final accuracy in Table 5 and Figure 5. This is test-set leakage and makes the reported accuracy drops optimistically biased. The authors should use a separate validation split for thresholding and for selecting beta and gamma, and evaluate the final model on the test set only once at the end of the pipeline.
  3. [§7.4, Figure 6; §8.3] The comparison with GA, RL, and PT is not controlled. QA optimizes the surrogate H, while GA and RL are stated to optimize measured accuracy (Section 7.4); the QA hyperparameters are selected using the test set, while the GA/RL baselines are tuned until convergence with no stated budget; and the reported optimization time for the annealers appears to exclude the queuing and scheduling overhead that Section 8.3 admits dominates. Under this protocol, a speed or optimality advantage of QA would be expected even with no quantum benefit. Please compare all methods on the same objective, use the same validation-based selection protocol, report end-to-end wall-clock time including embedding and queue time, and remove or substantiate the 'global optima' claim in the abstract, since no global optimum is known or certified for these problems.
minor comments (5)
  1. [§3, Eq. (1)] Equation (1) calls U a unitary matrix; for a QUBO problem U is a real symmetric (or upper-triangular) matrix and is not required to be unitary.
  2. [§7.3, Table 5] Table 5 reports single optimization runs with no variance; at least a few repeats with different random seeds or read numbers should be reported to assess stability.
  3. [§7.4, Figure 6] The labels 'L', 'R', 'P', and 'T' in Figure 6 are not explained in the caption; they should be spelled out and matched to the legend entries.
  4. [General] The manuscript contains repeated abstract blocks and garbled table captions (notably Tables 1-3), which should be cleaned before resubmission.
  5. [§7.1, 7.3] The pre-training scheme for the 8-bit models is not fully described, and the one-epoch fine-tuning used inside acc(·) is not specified in enough detail to be reproduced; please provide the training recipe and hyperparameters.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation was found: the QUBO objective is assembled from independent magnitude/RMSE heuristics and exact algebra, and the self-citations are not load-bearing.

full rationale

The paper's central derivation—Eq. (7) with pruning loss as squared average l1 magnitude (Eqs. 8–9), quantization loss as squared log RMSE with B_{n,k,l}=2^{k+l} (Eqs. 10–11), and compression rate as exact bit-count bookkeeping (Eqs. 12–13)—is mathematically self-contained. No parameter in the objective is fitted to the reported accuracy or reduction results: beta and gamma are user-set hyperparameters, and the coefficients A, B, D, E, F are precomputed from weights, layer sizes, and binary encodings. The claim that AQC finds effective compression solutions is therefore not equivalent by construction to the inputs. The main experimental concerns are soundness issues rather than circularity: the pruning/quantization surrogates are not validated against post-fine-tuning accuracy along the search path, and Algorithm 1 selects beta and gamma using test-set accuracy ('Evaluate accuracy a using aqc(.) and acc(.)'), a data-leakage concern that does not make the derivation self-referential. References [24] and [25] are self-citations used for the GA and RL baselines in Section 7.4, but those baselines are re-executed here, and the QUBO derivation does not depend on conclusions imported from those papers. Section 8 candidly lists qubit-count, connectivity, and accessibility limitations, which are external engineering constraints, not evidence of circular reasoning. No specific reduction of a prediction to a fitted input or to a self-citation chain was found.

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

The ledger is light on invented content: no new particles, mediators, forces, or dimensions are introduced. The main unstated inputs are accuracy proxies and hand-selected hyperparameters, which carry the empirical burden of the compression claims.

free parameters (3)
  • beta (quantization vs pruning balance) = initialized as ||A||_1/||B||_1, refined by binary search
    Balances pruning loss and quantization loss in H; selected per model using test accuracy.
  • gamma (compression reward weight) = randomly initialized, adjusted by FINDUPPERGAMMA and FINDLOWERGAMMA
    Controls the accuracy-compression trade-off; selected using test accuracy thresholds.
  • accuracy threshold a_th = 2% (LeNet-5/GTSR-CNN), 4% (ResNet-9), 10% (VGG-16)
    Hand-set acceptance threshold for valid solutions; directly affects the reported compression rate and which solutions are kept.
assumptions (4)
  • domain assumption Weight magnitude squared is a valid proxy for pruning-induced accuracy loss
    Section 5.1, Eq. (8)-(9); the proxy is not validated against measured accuracy on the search path.
  • domain assumption Quantization RMSE is approximated by s_n/sqrt(12) and step size scales as c_n * 2^(sum 2^k q)
    Section 5.2, Eq. (10)-(11); ignores the fact that step sizes are learned during fine-tuning.
  • domain assumption Unconstrained QUBO with soft penalties yields feasible bit-widths (b_n >= 0) after hyperparameter selection
    Section 5, Eq. (5)-(7); no hard constraint enforces non-negative bit-width, so feasibility depends on the search avoiding invalid regions.
  • domain assumption Pre-trained 8-bit models match FP32 accuracy
    Section 7.1; this is the baseline used for all accuracy-drop calculations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Quantum Optimization Ready? An Effort Towards Neural Network Compression using Adiabatic Quantum Computing." pith.science (2026). https://pith.science/paper/IRJTDQIE

@misc{pith2026250516332,
  author       = {Pith},
  title        = {Pith review of: Is Quantum Optimization Ready? An Effort Towards Neural Network Compression using Adiabatic Quantum Computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRJTDQIE}},
  note         = {Machine review of arXiv:2505.16332}
}
read the original abstract

Quantum optimization is the most mature quantum computing technology to date, providing a promising approach towards efficiently solving complex combinatorial problems. Methods such as adiabatic quantum computing (AQC) have been employed in recent years on important optimization problems across various domains. In deep learning, deep neural networks (DNN) have reached immense sizes to support new predictive capabilities. Optimization of large-scale models is critical for sustainable deployment, but becomes increasingly challenging with ever-growing model sizes and complexity. While quantum optimization is suitable for solving complex problems, its application to DNN optimization is not straightforward, requiring thorough reformulation for compatibility with commercially available quantum devices. In this work, we explore the potential of adopting AQC for fine-grained pruning-quantization of convolutional neural networks. We rework established heuristics to formulate model compression as a quadratic unconstrained binary optimization (QUBO) problem, and assess the solution space offered by commercial quantum annealing devices. Through our exploratory efforts of reformulation, we demonstrate that AQC can achieve effective compression of practical DNN models. Experiments demonstrate that adiabatic quantum computing (AQC) not only outperforms classical algorithms like genetic algorithms and reinforcement learning in terms of time efficiency but also excels at identifying global optima.

Figures

Figures reproduced from arXiv: 2505.16332 by the authors.

Figure 1
Figure 1. (a) An example coupling map of three qubits. The cost function (energy) of the system, H, can be formulated as a QUBO problem. (b) Since each qubit has two possibilities, the system can exhibit eight possible values of H. The AQC paradigm uses quantum mechanics to evolve qubit configurations toward the global minimum energy, seeking the minimum value of H. as genetic algorithms [25] or reinforcement learning algo￾ri… view at source ↗
Figure 1
Figure 1. supopose a 4-2-and 5 filters. and each the bit widht ranges from 1 bit to 8 bit. Pruning-quantization of neural network weights. We depict filter pruning, for which all weights in a filter are replaced with zero. For quantization, weight values are rounded to the nearest low-bit integer value. First Author et al.: Preprint submitted to Elsevier Page 1 of 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. supopose a 4-2-and 5 filters. and each the bit widht ranges from 1 bit to 8 bit. Pruning-quantization of neural network weights. We depict filter pruning, for which all weights in a filter are replaced with zero. For quantization, weight Pruning Parameters (a) values are rounded to the nearest low-bit integer value. First Author et al.: Preprint submitted to Elsevier Page 1 of 1 [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Heat maps of accuracy (left) and model reduction ratio (right) of ResNet-9 as distributions of hyperparameters 𝛽 (x-axis) and 𝛾 (y-axis). Dark areas represent regions where accuracy is completely lost. 5.2. Quantization Loss We use local quantization error as a proxy f…
Figure 5
Figure 5. Figure 5: Heat maps of accuracy (red, top row) and model reduction ratio (blue, bottom row) of (a) LeNet-5 for MNIST, (b) ResNet-9 for CIFAR-10, and (c) VGG-16 for CIFAR-100. The objective is to select a point that is brightly-colored in both accuracy (top) and model reduction r…
Figure 6
Figure 6. Figure 6: Comparison of quantum annealing (QA), genetic algorithms (GA), reinforcement learning (RL), and machine learning package tools (PT) regarding accuracy, model reduction ratio, and optimization time. We implemented the GA algorithms (NSGA-II) using code from Reference [2…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

  1. [1]

    Physics-inspired optimization for quadratic unconstrained problems using adigital annealer

    Aramon, M., Rosenberg, G., Valiante, E., Miyazawa, T., Tamura, H., Katzgraber, H.G., 2019. Physics-inspired optimization for quadratic unconstrained problems using adigital annealer. Frontiers in Physics 7, 48

    work page 2019

  2. [2]

    On the computational complexity of ising spin glass models

    Barahona, F., 1982. On the computational complexity of ising spin glass models. Journal of Physics A: Mathematical and General 15, 3241

    work page 1982

  3. [3]

    Bian, Z., Chudak, F., Macready, W., Roy, A., Sebastiani, R., Varotti, S., 2017. Solving sat and maxsat with a quantum annealer: Foun- dations and a preliminary report, in: Frontiers of Combining Sys- tems: 11th International Symposium, FroCoS 2017, Brasília, Brazil, September 27-29, 2017, Proceedings 11, Springer. pp. 153–171

    work page 2017

  4. [4]

    Quan- tum permutation synchronization, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition,pp.13122– 13133

    Birdal, T., Golyanik, V., Theobalt, C., Guibas, L.J., 2021. Quan- tum permutation synchronization, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition,pp.13122– 13133

    work page 2021

  5. [5]

    Dynamical channel pruning by conditional accuracy change for deep neural networks

    Chen, Z., Xu, T.B., Du, C., Liu, C.L., He, H., 2020. Dynamical channel pruning by conditional accuracy change for deep neural networks. IEEEtransactionsonneuralnetworksandlearningsystems 32, 799–813

    work page 2020

  6. [6]

    Quboformulationsfor training machine learning models

    Date,P.,Arthur,D.,Pusey-Nazzaro,L.,2021. Quboformulationsfor training machine learning models. Scientific reports 11, 10029

    work page 2021

  7. [7]

    Learned step size quantization

    Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S., 2019. Learned step size quantization. arXiv preprint arXiv:1902.08153

    arXiv 2019

  8. [8]

    Knowledgedistillation: Asurvey

    Gou,J.,Yu,B.,Maybank,S.J.,Tao,D.,2021. Knowledgedistillation: Asurvey. InternationalJournalofComputerVision129,1789–1819

    work page 2021

Show all 29 references
  1. [9]

    Benchmarking quantum annealing controls with portfolio optimization

    Grant, E., Humble, T.S., Stump, B., 2021. Benchmarking quantum annealing controls with portfolio optimization. Physical Review Applied 15, 014012

    2021

  2. [10]

    Deepcompression:Compressing deepneuralnetworkswithpruning,trainedquantizationandhuffman coding

    Han,S.,Mao,H.,Dally,W.J.,2015. Deepcompression:Compressing deepneuralnetworkswithpruning,trainedquantizationandhuffman coding. arXiv preprint arXiv:1510.00149

    2015 arXiv

  3. [11]

    Opq: Com- pressing deep neural networks with one-shot pruning-quantization, in:ProceedingsoftheAAAIconferenceonartificialintelligence,pp

    Hu, P., Peng, X., Zhu, H., Aly, M.M.S., Lin, J., 2021. Opq: Com- pressing deep neural networks with one-shot pruning-quantization, in:ProceedingsoftheAAAIconferenceonartificialintelligence,pp. 7780–7788

    2021

  4. [12]

    Benchmarking quantum (-inspired) annealing hardware on practical use cases

    Huang, T., Xu, J., Luo, T., Gu, X., Goh, R., Wong, W.F., 2022. Benchmarking quantum (-inspired) annealing hardware on practical use cases. IEEE Transactions on Computers 72, 1692–1705

    2022

  5. [13]

    The advantage quantum computer

    Inc., D.W.S., 2024. The advantage quantum computer. URL:https: //www.dwavesys.com/solutions-and-products/systems

    2024

  6. [14]

    Traffic signal optimization on a square lattice using the d-wave quan- tum annealer

    Inoue, D., Okada, A., Matsumori, T., Aihara, K., Yoshida, H., 2020. Traffic signal optimization on a square lattice using the d-wave quan- tum annealer. arXiv e-prints , arXiv–2003

    2020

  7. [15]

    Pruning and quantization for deep neural network acceleration: A survey

    Liang, T., Glossner, J., Wang, L., Shi, S., Zhang, X., 2021. Pruning and quantization for deep neural network acceleration: A survey. Neurocomputing 461, 370–403

    2021

  8. [16]

    Maximum cuts and large bipartite sub- graphs

    Poljak, S., Tuza, Z., 1995. Maximum cuts and large bipartite sub- graphs. DIMACS Series 20, 181–244

    1995

  9. [17]

    Pruning tutorial

    PyTorch, 2025a. Pruning tutorial. URL: https://pytorch.org/ tutorials/intermediate/pruning_tutorial.html

  10. [18]

    Quantization

    PyTorch, 2025b. Quantization. URL: https://pytorch.org/docs/ stable/quantization.html

  11. [19]

    Applica- tion of digital annealer for faster combinatorial optimization

    Sao, M., Watanabe, H., Musha, Y., Utsunomiya, A., 2019. Applica- tion of digital annealer for faster combinatorial optimization. Fujitsu Scientific and Technical Journal 55, 45–51

    2019

  12. [20]

    The german trafficsignrecognitionbenchmark:amulti-classclassificationcompe- tition,in:The2011internationaljointconferenceonneuralnetworks, IEEE

    Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C., 2011. The german trafficsignrecognitionbenchmark:amulti-classclassificationcompe- tition,in:The2011internationaljointconferenceonneuralnetworks, IEEE. pp. 1453–1460

    2011

  13. [21]

    An acceleratorarchitectureforcombinatorialoptimizationproblems

    Tsukamoto, S., Takatsu, M., Matsubara, S., Tamura, H., 2017. An acceleratorarchitectureforcombinatorialoptimizationproblems. Fu- jitsu Sci. Tech. J 53, 8–13

    2017

  14. [22]

    Clip-q:Deepnetworkcompressionlearning by in-parallel pruning-quantization, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Tung,F.,Mori,G.,2018. Clip-q:Deepnetworkcompressionlearning by in-parallel pruning-quantization, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 7873– 7882

    2018

  15. [23]

    Graph partitioning using quantum annealing on the d-wave system, in:ProceedingsoftheSecondInternationalWorkshoponPostMoores Era Supercomputing, pp

    Ushijima-Mwesigwa, H., Negre, C.F., Mniszewski, S.M., 2017. Graph partitioning using quantum annealing on the d-wave system, in:ProceedingsoftheSecondInternationalWorkshoponPostMoores Era Supercomputing, pp. 22–29

    2017

  16. [24]

    Edcompress: Energy-aware model compression for dataflows

    Wang, Z., Luo, T., Goh, R.S.M., Zhou, J.T., 2022. Edcompress: Energy-aware model compression for dataflows. IEEE Transactions on Neural Networks and Learning Systems

    2022

  17. [25]

    Evolutionarymulti-objectivemodelcompressionfordeepneuralnet- works

    Wang, Z., Luo, T., Li, M., Zhou, J.T., Goh, R.S.M., Zhen, L., 2021. Evolutionarymulti-objectivemodelcompressionfordeepneuralnet- works. IEEE Computational Intelligence Magazine 16, 10–21

    2021

  18. [26]

    Yang, H., Gui, S., Zhu, Y., Liu, J., 2020. Automatic neural network compression by sparsity-quantization joint learning: A constrained optimization-basedapproach,in:ProceedingsoftheIEEE/CVFCon- ferenceonComputerVisionandPatternRecognition,pp.2178–2188

    2020

  19. [27]

    Adiabatic quantum computing for multi object tracking, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zaech, J.N., Liniger, A., Danelljan, M., Dai, D., Van Gool, L., 2022. Adiabatic quantum computing for multi object tracking, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8811–8822

    2022

  20. [28]

    To prune, or not to prune: exploring the efficacy of pruning for model compression

    Zhu, M., Gupta, S., 2017. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878

    2017 arXiv

  21. [29]

    borealis—a generalized global update algorithm for boolean optimization problems

    Zhu, Z., Fang, C., Katzgraber, H.G., 2020. borealis—a generalized global update algorithm for boolean optimization problems. Opti- mization Letters 14, 2495–2514. Zhehui Wang, Benjamin Chen Ming Choong et al.:Preprint submitted to Elsevier Page 12 of 12

    2020

Pith tools

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