Pith. sign in

REVIEW 3 major objections 4 minor 35 references

Efficient training for large-scale optical neural network using an evolutionary strategy and attention pruning

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A CMA-ES plus attention-pruning training algorithm for block optical neural networks prunes 60-80% of parameters with under 5% accuracy loss and shows improved noise robustness.

desk verdict The pruning contribution is real and worth peer review, but the robustness advantage over CMA-ES is not supported by the reported metrics. read the letter →

arxiv 2505.12906 v1 pith:DEIPAPVX submitted 2025-05-19 cs.LG physics.optics

Zhiwei Yang , Zeyang Fan , Yihang Lai , Qi Chen , Tian Zhang , Jian Dai , Kun Xu This is my paper · ORCID
classification cs.LGphysics.optics
keywords algorithmbonnslarge-scaleparametersdatasetmatrixmnistnetwork
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

Optical neural networks process information with light, using chips made of many small interferometers. A large network has thousands of adjustable phase values, and training is hard because hardware noise makes gradients unreliable. The authors combine two known ideas. First, they use CMA-ES, an evolutionary search that keeps a population of candidate settings and adapts its search distribution, so it does not need exact gradients. Second, they attach a learnable attention score to each block of the weight matrix, train the scores for part of the run, then delete the blocks with the lowest scores and retrain the remaining network.

On MNIST, the method removes 60 percent of parameters while losing 3.3 percentage points of accuracy. On Fashion-MNIST, it removes 80 percent while losing 4.7 points. These losses are smaller than those from random pruning. Under a simulated noise model for phase shifters, the pruned network degrades less than two comparison training algorithms. The authors also ran a small hardware experiment with a 4x4 interferometer chip on a simplified MNIST task and reached 88.5 percent accuracy at 60 percent pruning, close to the no-noise simulation.

The main caveat is that the results are on small datasets and a simplified experimental setup. The robustness comparison also depends on the assumed noise model rather than measured noisy hardware. The paper is a practical step for making optical neural networks easier to train and compress, not a conceptual breakthrough.

Extended reading notes

Core claim

The paper asserts that 'the CAP algorithm can prune 60% and 80% of the parameters for MNIST and Fashion-MNIST datasets, respectively, while only degrades the performance by 3.289% and 4.693%', and that under dynamic phase-shifter noise with standard deviation 0.5, CAP's performance degradation (22.327% MNIST, 24.019% Fashion-MNIST) is the smallest among BAT (43.963%, 41.074%) and CMA-ES (25.757%, 32.871%). This robustness ordering is the central load-bearing claim.

Load-bearing premise

The robustness comparison relies on the assumption, introduced in Section 3.2, that dynamic phase-shifter error is an additive Gaussian random variable with mean zero and standard deviation sigma_P, and that static error is captured by the extinction ratio of the directional coupler. The headline advantage of CAP over BAT and CMA-ES is demonstrated under this simulated noise model, not under measured noise on the fabricated chip. If real chip noise is correlated, non-Gaussian, or drift-dominated, the reported robustness ordering may not transfer to hardware.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes CAP, an on-chip training algorithm for large-scale MZI-based block optical neural networks (BONNs) that combines CMA-ES with attention-based block pruning. Learning attention coefficients for each matrix block, the method prunes low-attention blocks and then retrains the remaining phase shifters with CMA-ES. The authors report that CAP can prune 60% and 80% of parameters on MNIST and Fashion-MNIST with test-accuracy degradation of 3.289% and 4.693%, respectively. Under simulated dynamic Gaussian phase-shifter noise with standard deviation 0.5, they report that CAP exhibits smaller accuracy degradation (22.327% MNIST, 24.019% Fashion-MNIST) than their previously reported BAT algorithm (43.963%, 41.074%) and CMA-ES (25.757%, 32.871%). They also experimentally train a 4x4 MZI PIC on a simplified MNIST task, achieving 88.5% accuracy with 60% pruning, close to the noiseless simulation value of 92.1%, and they demonstrate that omitting external phase shifters reduces parameter count without major accuracy loss.

Significance. If the central robustness ordering holds, CAP would be a meaningful practical contribution to on-chip training of optical neural networks, combining evolutionary optimization with attention-based structural pruning in a setting where gradient signals are noisy and chip area is constrained. The paper includes a genuine hardware experiment, which supports the feasibility of online evolutionary training on a fabricated PIC. The population-based approach is a reasonable alternative to gradient-based methods for noisy hardware. However, the statistical support for the headline robustness claim is weak, and the noise model is assumed rather than measured, so the significance of the claimed advantage over CMA-ES on MNIST in particular is not established. The pruning results are directionally interesting but are also selected on test-set accuracy without a validation split or repeated trials.

major comments (3)
  1. [Section 3.2, Fig. 4(d), Abstract] The headline robustness advantage of CAP over CMA-ES on MNIST is not supported by the reported data. The performance-degradation metric in the abstract is computed relative to different noiseless baselines: CAP's pruned baseline is 89.896% (Fig. 4(a), S=40%) while CMA-ES's unpruned baseline is 93.185% (Fig. 2(a)). At sigma_P=0.5, CAP's absolute avg10acc is 67.569% and CMA-ES's absolute avg10acc is 93.185% - 25.757% = 67.428%, a gap of only 0.14 percentage points. No error bars or repeated-seed statistics are reported for CAP or CMA-ES; the only multi-trial band shown is for BAT (Fig. 4(d)). A 0.14-point separation is well within run-to-run variation for evolutionary training, so the claim that CAP is more robust than CMA-ES on MNIST is not statistically established. On Fashion-MNIST the absolute gap is larger (56.519% vs 52.36%), but even there no error bars are given. The paper should either report absolute noisy accuracies with uncertainty or restrict the robustness claim to Fashion-MNIST with appropriate error characterization.
  2. [Section 3.2, noise model; Section 3.3, experiment] The entire robustness comparison relies on the assumption introduced in Section 3.2 that dynamic phase-shifter error is an additive Gaussian random variable with mean zero and standard deviation sigma_P, and that static error is fully captured by the extinction ratio of the directional coupler. The hardware experiment in Section 3.3 does not measure this noise distribution on the fabricated chip, nor does it inject calibrated dynamic noise to validate the simulation model. The authors attribute the small simulation-experiment gap for the BONN without external phase shifters to external phase-shifter noise, but no noise characterization is provided. Consequently, the transferability of the reported robustness ordering to real hardware is not demonstrated, and the claim that CAP is 'strongest robustness' under real device conditions remains an extrapolation from an unvalidated model.
  3. [Sections 3.1 and 3.2, hyperparameter selection] The hyperparameters S (attention-training ratio), R (pruning ratio), and population size h are all selected using avg10acc, which is the test-set accuracy, and no validation split is used. For example, S=40% is chosen because it gives the highest avg10acc on the MNIST test set (Fig. 4(a)), and the reported pruning degradations (3.289% and 4.693%) are the results at the selected R and S. This creates an optimistic bias in the headline numbers, since the same test set is used for both model selection and evaluation. In addition, no repeated trials are reported for CAP or CMA-ES, so the reported gaps (e.g., the 4.159 percentage-point advantage on Fashion-MNIST) cannot be distinguished from noise. The authors should use a validation set for hyperparameter selection, report means and standard deviations over multiple independent runs, and state the number of seeds.
minor comments (4)
  1. [Abstract and throughout] There are numerous grammatical errors and typos, e.g., 'have increasingly drawn attentions', 'the CAP algorithm show excellent potential', and inconsistent spacing in equations. A thorough language edit is needed.
  2. [References [7] and [8]] In the Introduction, the statement that MNIST requires around one thousand ports and CIFAR-10 requires thousands or tens of thousands of ports is cited to references [7] and [8], which are papers about coin betting optimization and mixing Adam/SGD, respectively. These citations appear incorrect and should be replaced with appropriate sources.
  3. [Equation (7)] The cross-entropy fitness formula appears garbled in the typeset equation; the softmax normalization and logarithm are not clearly expressed. Please rewrite the equation and define all symbols explicitly.
  4. [Figures 4 and 5] The figures show solid lines for CAP and CMA-ES without error bars, while the sand-colored bands are only for BAT. Adding shading or error bars for all methods would make the comparison more transparent. Also, the x-axis labels for Fig. 4(d) and Fig. 5(d) should clearly state whether the curves correspond to training with noise or testing an already-trained network under noise, since the text discusses both settings.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the CAP claims are empirical results benchmarked against external data and baselines, with no equation-level reduction to inputs.

full rationale

The paper's central claims are empirical: CAP's pruning performance (3.289% degradation for MNIST at 60% pruning and 4.693% for Fashion-MNIST at 80% pruning) and robustness (22.327% and 24.019% degradation under dynamic phase-shifter noise with sigma_P = 0.5) are measured from simulations and an experiment, not derived from a definition. The attention coefficient alpha is a learned parameter fitted from the training data; using it to choose blocks for pruning is a standard learned-importance procedure, and the paper compares CAP against random pruning (RPC), so the pruning advantage is not forced by construction. The robustness comparison adds dynamic Gaussian phase-shifter noise and static directional-coupler error to trained networks; these noise terms are inputs of the experiment, not fitted parameters that predetermine the claimed ordering. The use of the authors' earlier BAT algorithm and their earlier work [23] is as a comparison baseline and a design heuristic, respectively, and both are independently verified within this paper: Fig. 2(c) directly compares BAT, CMA-ES, and BANN, and Fig. 6(c) compares simulation with experiment for the external-phase-shifter-free BONN. No uniqueness theorem or ansatz is imported from self-citations to justify the CAP choice. The assumed Gaussian dynamic-noise model limits external validity to that model, but it is not circular because the model is stated as an assumption rather than fitted to make CAP win. Therefore no circular step can be exhibited using the paper's own equations.

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

The algorithm introduces no new physical entity; its main movable parts are hyperparameters and a noise model. The block size, population size, and attention-training ratio are picked by sweeping on the test set, and the noise level used to claim robustness is a chosen scenario, so these carry much of the epistemic load.

free parameters (5)
  • block size g = 12
    Selected as the value with highest avg10acc on MNIST and Fashion-MNIST among tested block sizes (Fig. 2(a), 3(a)); it changes the balance of nonlinearity and parameter count.
  • population size h = 100 (MNIST), 120 (Fashion-MNIST), 20 (experiment)
    Chosen as the largest or largest useful value tested because accuracy increases with population size (Fig. 2(b), 3(b), 6(b)); no saturation criterion or budget constraint reported.
  • attention-training ratio S = 0.4
    Swept and set to 40% because it gave the best avg10acc on both datasets (Fig. 4(a), 5(a)); too small gives inaccurate attention scores, too large leaves insufficient retraining.
  • initial CMA-ES step size sigma0 = 0.1
    Hand-set at initialization (Section 2, step 1) with no sensitivity study; central to search behavior.
  • noise standard deviation sigma_P = 0.5
    Used as a 'poor fabricated chip' scenario (abstract, Fig. 4/5); chosen to show robustness, not derived from measured hardware.
assumptions (5)
  • domain assumption An arbitrary real-valued weight matrix can be decomposed into a product of unitary matrices and a diagonal matrix, and each unitary block can be realized by a Clements MZI mesh (Refs [21,22]).
    Used to justify constructing block unitary matrices with MZIs; the paper relies on Clements decomposition for architecture.
  • domain assumption Replacing full real-valued weight matrices with block-unitary matrices causes acceptable performance loss.
    Stated in Section 2 and supported by comparing BANN vs BONN and Ref [23]; if false, the BONN architecture would not be a valid approximation.
  • domain assumption The PD array acts as a quadratic nonlinearity after each block, providing enough nonlinearity for classification.
    Section 2, Fig. 1; the nonlinearity is hardware-specific and its sufficiency for deeper tasks is not proven.
  • ad hoc to paper Dynamic phase-shifter error is additive Gaussian noise with zero mean; static error is captured by extinction ratio.
    Introduced in Section 3.2 to model chip imperfections; the robustness claims depend on this model.
  • domain assumption Neuroptica simulator accurately models the MZI mesh and noise behavior.
    Section 2 states PIC is simulated by Neuroptica [24]; simulation-to-experiment agreement is only shown for a small simplified task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient training for large-scale optical neural network using an evolutionary strategy and attention pruning." pith.science (2026). https://pith.science/paper/DEIPAPVX

@misc{pith2026250512906,
  author       = {Pith},
  title        = {Pith review of: Efficient training for large-scale optical neural network using an evolutionary strategy and attention pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEIPAPVX}},
  note         = {Machine review of arXiv:2505.12906}
}
read the original abstract

MZI-based block optical neural networks (BONNs), which can achieve large-scale network models, have increasingly drawn attentions. However, the robustness of the current training algorithm is not high enough. Moreover, large-scale BONNs usually contain numerous trainable parameters, resulting in expensive computation and power consumption. In this article, by pruning matrix blocks and directly optimizing the individuals in population, we propose an on-chip covariance matrix adaptation evolution strategy and attention-based pruning (CAP) algorithm for large-scale BONNs. The calculated results demonstrate that the CAP algorithm can prune 60% and 80% of the parameters for MNIST and Fashion-MNIST datasets, respectively, while only degrades the performance by 3.289% and 4.693%. Considering the influence of dynamic noise in phase shifters, our proposed CAP algorithm (performance degradation of 22.327% for MNIST dataset and 24.019% for Fashion-MNIST dataset utilizing a poor fabricated chip and electrical control with a standard deviation of 0.5) exhibits strongest robustness compared with both our previously reported block adjoint training algorithm (43.963% and 41.074%) and the covariance matrix adaptation evolution strategy (25.757% and 32.871%), respectively. Moreover, when 60% of the parameters are pruned, the CAP algorithm realizes 88.5% accuracy in experiment for the simplified MNIST dataset, which is similar to the simulation result without noise (92.1%). Additionally, we simulationally and experimentally demonstrate that using MZIs with only internal phase shifters to construct BONNs is an efficient way to reduce both the system area and the required trainable parameters. Notably, our proposed CAP algorithm show excellent potential for larger-scale network models and more complex tasks.

Figures

Figures reproduced from arXiv: 2505.12906 by the authors.

Figure 1
Figure 1. (a) The flowchart of the training process for the BONN based on CAP algorithm. (b) The general scheme of the BONN based on MZIs. (c) The architecture of PIC [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 19 canonical work pages

  1. [1]

    C. Teng, X. Tang, Y. Huang, W. Du, W. He, A. Ren, J. W u, Z. Wang, A high -stability all-optical nonlinear activator for optical computing, J . Lightw. Technol. 41(21) (2023) 6817 -6823, http://dx.doi.org/10.1109/JLT.2023.3290021

    arXiv 2023

  2. [2]

    Shen, N .C

    Y. Shen, N .C. Harris, S. Skirlo, M. Prabhu, T. Baehr -Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund, Deep learning with coherent nanophotonic circuits, Nat . Photon. 11(7) (2017) 441-446, http://dx.doi.org/10.1038/NPHOTON.2017.93

    doi:10.1038/nphoton.2017.93 2017

  3. [3]

    Z. Fan, J. Lin, J. Dai, T. Zhang, K. Xu, Photonic Hopfield neural network for the Ising problem, Opt . Express 31(13) (2023) 21340-21350, http://dx.doi.org/10.1364/OE.491554

    work page doi:10.1364/oe.491554 2023

  4. [4]

    M.B. On, Y. -J. Lee, L.E. Srouji, M. Abdelghany, S.B. Yoo, Improving Scalability of Mach -Zehnder Interferometer Based Photonic Computers via Algorithm-Hardware Codesign, J. Lightw. Technol. 42(22) (2024) 8014-8023

    work page 2024

  5. [5]

    Bagherian, S

    H. Bagherian, S. Skirlo, Y. Shen, H. Meng, V. Ceperic, M. Soljacic, On-chip optical convolutional neural networks, a rXiv preprint arXiv:1808.03303 (2018), http://dx.doi.org/10.48550/arXiv.1808.03303

    arXiv doi:10.48550/arxiv.1808.03303 2018

  6. [6]

    https://bm.cnfic.com.cn/sharing/share/articleDetail/206165931/1

    work page arXiv

  7. [7]

    Orabona, T

    F. Orabona, T. Tommasi, Training deep networks without learning rates through coin betting, Advances in Neural Information Processing Systems 30 (2017)

    work page 2017

  8. [8]

    Mixing ADAM and SGD: a Combined Optimization Method

    N. Landro, I. Gallo, R. La Grassa, Mixing Adam and SGD: A combined optimization method, arXiv preprint arXiv:2011.08042 (2020), https://doi.org/10.48550/arXiv.2011.08042

    work page Pith review arXiv doi:10.48550/arxiv.2011.08042 2020

Show all 35 references
  1. [9]

    J. Gu, H. Zhu, C. Feng, Z. Jiang, R. Chen, D. Pan, L2ight: Enabling on-chip learning for optical neural networks via effi cient in-situ subspace optimization, Advances in Neural Information Processing Systems 34 (2021) 8649-8661

    2021

  2. [10]

    C. Feng, J. Gu, H. Zhu, Z. Ying, Z. Zhao, D.Z. Pan, R.T. Chen, A Compact Butterfly-Style Silicon Photonic–Electronic Neural Chip for Hardware-Efficient Deep Learning, Acs Photon. 9(12) (2022) 3906-3916, http://dx.doi.org/10.1021/acsphotonics.2c01188

    2022 doi

  3. [11]

    C. Ding, S. Liao, Y. Wang, Z. Li, N. Liu, Y. Zhuo, C. Wang, X. Qian, Y. Bai, G. Yuan, Circnn: accelerating and compressing deep neural networks using block -circulant weight matrices, Proceedings of the 50th Annual IEEE/ACM International Symposium on Microarchitecture (2017) 395-408

    2017

  4. [12]

    S. Liao, Z. Li, X. Lin, Q. Qiu, Y. Wang, B. Yuan, Energy-efficient, high-performance, highly-compressed deep neural network design using block -circulant matrices, 2017 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) (2017) 458-465

    2017

  5. [13]

    H. Zhou, Y. Zhao, X. Wang, D. Gao, J. Dong, X. Zhang, Self-learning photonic signal processor with an optical neural network chip, arXiv preprint arXiv:1902.07318 (2019) http://dx.doi.org/10.48550/arXiv.1902.07318

    2019 doi arXiv

  6. [14]

    Z. Yang, T. Zhang, J. Dai, K. Xu, Efficient on -chip training of large-scale optical neural network through block adjoint training algorithm, Opt . Express 32(26) (2024) 46633 -46648, https://doi.org/10.1364/OE.537813

    2024 doi

  7. [15]

    Salimans, J

    T. Salimans, J. Ho, X. Chen, S. Sidor, I. Sutskever, Evo lution strategies as a scalable alternative to reinforcement learning, arXiv preprint arXiv:1703.03864 (2017) , https://doi.org/10.48550/arXiv.1703.03864

    2017 doi arXiv

  8. [16]

    Zhang, J

    T. Zhang, J. Wang, Y. Dan, Y. Lanqiu, J. Dai, X. Han, X. Sun, K. Xu, Efficient training and design of photonic neural network through neuroevolution, Opt. Express 27(26) (2019) 37150-37163, http://dx.doi.org/10.1364/OE.27.037150

    2019 doi

  9. [17]

    Hansen, S.D

    N. Hansen, S.D. Mü ller, P. Koumoutsakos, Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (CMA -ES), Evolutionary computation 11(1) (2003) 1 -18, http://dx.doi.org/10.1162/106365603321828970

    2003 doi

  10. [18]

    Chang, K

    H. Chang, K. Futagami, Reinforcement learning with convolutional reservoir computing, Appl . Intell. 50(8) (2020) 2400-2410, http://dx.doi.org/10.1007/s10489-020-01679-3

    2020 doi

  11. [19]

    Miyatake, K

    Y. Miyatake, K. Toprasertpong, S. Takagi, M. Takenaka, Design of compact and low-loss S-bends by CMA-ES, Opt. Express 31(26) (2023) 43850-43863, http://dx.doi.org/10.1364/OE.504866

    2023 doi

  12. [20]

    Y. Su, Z. Chen, F. Xu, Y. Ye, X. Jiang, W. Liu, Y. Xie, A Multi-Objective Particle Swarm Optimization Pruning on Photonic Neural Networks, J. Lightw. Technol. 43(5) (2025) 2213 - 2225, http://dx.doi.org/10.1109/JLT.2024.3486718

    2025

  13. [21]

    Lawson, R.J

    C.L. Lawson, R.J. Hanson, Solving least squares problems, SIAM1995

  14. [22]

    Clements, P.C

    W.R. Clements, P.C. Humphreys, B.J. Metcalf, W.S. Kolthammer, I.A. Walmsley, Optimal design for universal multiport interferometers, Optica 3(12) (2016) 1460 -1465, http://dx.doi.org/10.1364/OPTICA.3.001460

    2016 doi

  15. [23]

    Z. Yang, T. Zhang, J. Dai, K. Xu, Tunable -bias based optical neural network for reinforcement learning in path planning, Opt . Express 32(10) (2024) 18099 -18112, http://dx.doi.org/10.1364/OE.516173

    2024 doi

  16. [24]

    Bartlett, M

    B. Bartlett, M. Minkov, T. Hughes, I. Williamson, Neuroptica: an optical neural network simulator, https://github.com/fancompute/neuroptica (2019)

    2019

  17. [25]

    Shokraneh, M.S

    F. Shokraneh, M.S. Nezami, O. Liboiron -Ladouceur, Theoretical and experimental analysis of a 4× 4 reconfigurable MZI -based linear optical processor, J . Lightw. Technol. 38(6) (2020) 1258-1267, http://dx.doi.org/10.1109/JLT.2020.2966949

    2020

  18. [26]

    Vivien, A

    L. Vivien, A. Polzer, D. Marris -Morini, J. Osmond, J.M. Hartmann, P. Crozat, E. Cassan, C. Kopp, H. Zimmermann, J.M. Fé dé li, Zero-bias 40Gbit/s germanium waveguide photodetector on silicon, Opt. Express 20(2) (2012) 1096 -1101, http://dx.doi.org/10.1364/OE.20.001096

    2012 doi

  19. [27]

    Autere, H

    A. Autere, H. Jussila, Y. Dai, Y. Wang, H. Lipsanen, Z . Sun, Nonlinear optics with 2D layered materials, Adv . Mater. 30(24) (2018) 1705963, http://dx.doi.org/10.1002/adma.201705963

    2018 doi

  20. [28]

    Shorten, T.M

    C. Shorten, T.M. Khoshgoftaar, B. Furht, Text data augmentation for deep learning, J . big Data 8(1) (2021) 101, http://dx.doi.org/10.1186/s40537-021-00492-0

    2021 doi

  21. [29]

    LeCun, The MNIST database of handwritten digits, http://yann

    Y. LeCun, The MNIST database of handwritten digits, http://yann. lecun. com/exdb/mnist/ (1998)

    1998

  22. [30]

    Kingma, J

    D.P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014) , https://doi.org/10.48550/arXiv.1412.6980

    2014 doi arXiv

  23. [31]

    H. Xiao, K. Rasul, R. Vollgraf, Fashion -mnist: a novel image dataset fo r benchmarking machine learning algorithms, arXiv preprint arXiv:1708.07747 (2017) http://dx.doi.org/10.48550/arXiv.1708.07747

    2017 doi arXiv

  24. [32]

    Chetlur, C

    S. Chetlur, C. Woolley, P. Vandermersch, J. Cohen, J. Tran, B . Catanzaro, E. Shelhamer, cudnn: Efficient primitives for deep learning, arXiv preprint arXiv:1410.0759 (2014) http://dx.doi.org/10.48550/arXiv.1410.0759

    2014 doi arXiv

  25. [33]

    Banerjee, M

    S. Banerjee, M. Nikdast, S. Pasric ha, K. Chakrabarty, Pruning coherent integrated photonic neural networks, IEEE J . Sel. Top. Quant. 29(2) (2023) 1 -13, http://dx.doi.org/10.1109/JSTQE.2023.3242992

    2023

  26. [34]

    R. Shao, G. Zhang, X. Gong, Generalized robust training scheme using genetic algorithm for optical neural networks with imprecise components, Photon. Res. 10(8) (2022) 1868 -1876, http://dx.doi.org/10.1364/PRJ.449570

    2022 doi

  27. [35]

    Z. Fan, Y. Dan, J. Lin, T. Zhang, J. Dai, K. Xu, Rapid configuring method for a programmable photonic integrated circuit based on a tandem neural network, Opt. Lett., 50(5) (2025) 1731-1734, https://doi.org/10.1364/OL.551119

    2025 doi

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.