arxiv: 2604.16228 · v2 · submitted 2026-04-17 · ⚛️ physics.optics · physics.app-ph

Recognition: unknown

TRON: Trainable, architecture-reconfigurable random optical neural networks

Ziao Wang , Fei Xia , Logan G. Wright , Tatsuhiro Onodera , Martin Stein , Jianqi Hu , Peter L. McMahon , Sylvain Gigan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 07:25 UTC · model grok-4.3

classification ⚛️ physics.optics physics.app-ph

keywords optical neural networksin-situ optimizationneural architecture searchmulti-scattering mediumDMDoptical matrix multiplierreconfigurable optical computingmachine learning hardware

0 comments

The pith

TRON shows that in-situ neural architecture search on optical hardware is essential for discovering effective reconfigurable network designs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TRON is an optoelectronic system for deep optical neural networks that treats a multi-scattering medium and a DMD as a trainable high-dimensional optical matrix multiplier. The approach performs in-situ optimization of the optical parameters in the scattering process along with automated neural architecture search directly on the hardware. This combination allows the network to adapt its structure to the demands of a given task while respecting the physical constraints of the optical setup. Experimental results indicate that the in-situ search step is required to obtain architectures that deliver usable performance, which supports further development of large-scale optical processors for machine learning.

Core claim

By exploiting a multi-scattering medium and a DMD as a learnable dense optical matrix multiplier and optimizing both the scattering parameters and the network architecture in situ, TRON implements trainable and architecture-reconfigurable random optical neural networks that combine fixed and tunable optical operations.

What carries the argument

The multi-scattering medium paired with a DMD acting as a learnable high-dimensional dense optical matrix multiplier, optimized in situ together with automated neural architecture search performed directly on the optics.

If this is right

In-situ NAS discovers architectures that adapt to both the computational task and the specific hardware constraints of the optical system.
Processing combines fixed optical operations with tunable ones to exploit massive parallelism and high bandwidth.
The method establishes a viable path toward large-scale optical processors for next-generation machine learning and data-intensive computing.
Reconfigurable computing substrates become feasible without requiring all operations to be realized through external digital control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar in-situ optimization and search methods could be applied to other physical computing platforms that face reconfiguration and noise challenges.
Hybrid opto-electronic pipelines might emerge in which the optical stage handles high-dimensional matrix operations while electronics manage control and readout.
Energy efficiency gains for inference tasks would follow if the optical path can be scaled without proportional increases in noise or alignment overhead.

Load-bearing premise

The multi-scattering medium and DMD can be optimized in situ to implement a broad class of network architectures with sufficient accuracy and scalability even when optical noise, alignment drift, and limited dynamic range are present.

What would settle it

An experiment in which scaling the number of layers or the search space causes accuracy to fall below competitive levels on a standard task because accumulated optical noise cannot be overcome by further in-situ adjustments.

Figures

Figures reproduced from arXiv: 2604.16228 by Fei Xia, Jianqi Hu, Logan G. Wright, Martin Stein, Peter L. McMahon, Sylvain Gigan, Tatsuhiro Onodera, Ziao Wang.

**Figure 2.** Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

read the original abstract

Deep learning has triggered explosive growth in the demand for specialized hardware processors, thus motivating the development of scalable and reconfigurable computing substrates. Optical processors offer a fundamentally different computing paradigm, combining massive parallelism and ultrahigh bandwidth with the potential for substantial energy savings. However, progress has been constrained by the absence of scalable and reconfigurable architectures that can implement a broad class of network architectures. Here, we introduce TRON, a scalable and trainable optoelectronic deep optical neural network that exploits a multi-scattering medium and a DMD as a learnable, high-dimensional dense optical matrix multiplier, processing with fixed and tunable optical operations. We perform in-situ optimization of the optical parameters involved in the scattering process, together with automated neural architecture search (NAS) and optimization directly on optics. The experimental results demonstrate that in-situ NAS is essential to discover architectures that adapt to both the task and hardware constraints, establishing a viable path towards large-scale optical processors for next-generation machine learning and data-intensive computing.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TRON demonstrates an optical matrix multiplier using scattering media and DMD that supports in-situ training plus NAS, but the path to large-scale use hinges on unaddressed noise and drift issues.

read the letter

TRON's main contribution is a hardware platform that turns a multi-scattering medium plus DMD into a trainable dense optical multiplier, with parameters optimized directly on the optics and with neural architecture search run in place to fit both the task and the physical constraints. The experiments indicate that skipping the in-situ NAS produces unusable performance, which is a concrete finding rather than a simulation claim. That combination of learnable scattering operations and on-hardware search is not just a rehash of earlier fixed-architecture optical NNs; it directly tackles the reconfigurability bottleneck mentioned in the abstract. The work is grounded in actual device runs rather than purely theoretical derivations, and the authors show that the setup can handle both fixed and tunable optical steps without external simulation loops. This is useful evidence for anyone trying to move optical computing beyond proof-of-concept stages. The soft spot is the limited treatment of physical error sources. Speckle noise, mechanical drift, and the DMD's restricted modulation depth are known to compound with depth and width; the paper does not appear to include long-term stability traces, noise-floor measurements across runs, or scaling curves that would show how accuracy holds as networks grow. Without those, the assertion that this establishes a viable route to large-scale processors rests on small-scale demonstrations whose extrapolation is not yet supported by the data. The citation pattern looks standard for the field and does not rely on circular self-reference. This paper is for groups working on photonic hardware for machine learning who need concrete examples of hardware-in-the-loop optimization. It is coherent on its own terms and shows clear experimental thinking, so it merits a serious referee even though reviewers will need to press on robustness and scaling. I would send it out for review rather than desk-reject.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces TRON, a trainable and architecture-reconfigurable random optical neural network that employs a multi-scattering medium combined with a DMD to realize a learnable, high-dimensional dense optical matrix multiplier. It performs in-situ optimization of the optical scattering parameters together with automated neural architecture search (NAS) directly on the physical hardware, claiming that experiments demonstrate in-situ NAS is essential for discovering task- and hardware-adapted architectures and thereby establishing a viable path to large-scale optical processors.

Significance. If the experimental results hold with adequate quantification, the work would address a central limitation in optical computing by providing a reconfigurable, trainable substrate that adapts to both algorithmic and physical constraints, potentially enabling energy-efficient, high-bandwidth neural network implementations at scale.

major comments (2)

[Experimental Results] Experimental Results section: the central claim that 'in-situ NAS is essential' is asserted without reported performance numbers, baselines (e.g., fixed-architecture controls), error bars, training curves, or hardware characterization metrics. This absence prevents assessment of whether the observed gains are statistically meaningful or merely due to the specific task/hardware instance.
[Results and Discussion] Results and Discussion sections: no measurements or scaling analysis of optical noise floor, alignment/thermal drift over time, or effective dynamic range of the DMD-plus-scattering system are provided. These quantities directly determine whether the learnable matrix multiplier can maintain usable fidelity as network depth or width increases, which is load-bearing for the scalability claim.

minor comments (1)

[Abstract and Introduction] The abstract and introduction use the phrase 'fixed and tunable optical operations' without clarifying which operations are fixed versus tunable or how the distinction is implemented in the optical path.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each major comment point by point below and propose revisions where appropriate to strengthen the presentation of our results.

read point-by-point responses

Referee: [Experimental Results] Experimental Results section: the central claim that 'in-situ NAS is essential' is asserted without reported performance numbers, baselines (e.g., fixed-architecture controls), error bars, training curves, or hardware characterization metrics. This absence prevents assessment of whether the observed gains are statistically meaningful or merely due to the specific task/hardware instance.

Authors: We agree that the Experimental Results section would benefit from more explicit quantitative support for the claim. In the revised manuscript, we will add direct performance comparisons of in-situ NAS against fixed-architecture baselines, including error bars from multiple experimental runs, representative training curves, and additional hardware characterization metrics. These additions will enable assessment of statistical significance and help distinguish task- and hardware-specific effects. revision: yes
Referee: [Results and Discussion] Results and Discussion sections: no measurements or scaling analysis of optical noise floor, alignment/thermal drift over time, or effective dynamic range of the DMD-plus-scattering system are provided. These quantities directly determine whether the learnable matrix multiplier can maintain usable fidelity as network depth or width increases, which is load-bearing for the scalability claim.

Authors: We acknowledge that quantitative characterization of these system-level properties is important for supporting scalability. While the manuscript discusses experimental stability in the context of the reported results, we will expand the Results and Discussion sections to include explicit measurements of the optical noise floor, alignment and thermal drift over relevant timescales, and the effective dynamic range of the DMD-plus-scattering system, together with a brief scaling analysis. This will provide a clearer basis for evaluating fidelity limits at larger scales. revision: yes

Circularity Check

0 steps flagged

No circularity: experimental demonstration with no derivation chain or fitted predictions

full rationale

The paper introduces TRON as an experimental platform using a multi-scattering medium and DMD for in-situ trainable optical matrix multiplication, with results from automated NAS and optimization on optics. No equations, derivations, or parameter-fitting steps are described that could reduce by construction to inputs, self-citations, or renamed empirical patterns. The central claim rests on physical experimental outcomes rather than any theoretical prediction loop. This matches the default expectation for non-circular experimental work and warrants score 0 with empty steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities. The approach relies on established concepts from optics and machine learning without introducing new unproven physical entities or ad-hoc assumptions visible at this level.

pith-pipeline@v0.9.0 · 5493 in / 1207 out tokens · 47995 ms · 2026-05-10T07:25:40.866044+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

74 extracted references · 8 canonical work pages · 4 internal anchors

[1]

Large-scale 3D medical CT image classification The domain of 3D image processing, especially in clinical settings such as CT imaging, poses distinct challenges arising from the complexity and variability of anatomical structures. Reliable classification of these images is essential for a range of medical tasks, including disease diagnosis, planning therap...
[2]

We selected a dataset from genomics, a domain that requires powerful and precise methods to analyze and categorize high-dimensional, complex data

Computing with RNA sequence for cell type-specific disease classification In the second demonstration, we move beyond image processing and investigate sequential data processing using light. We selected a dataset from genomics, a domain that requires powerful and precise methods to analyze and categorize high-dimensional, complex data. Gene sequence class...
[3]

Purify cells population Exon ExonIntron
[4]

Test tumor samples using RNA-seq Single-cell expression profile FIG. 5. Demonstration of the architecture-optimized TRON for gene classification. (a) Conceptual illustration of the RNAseq dataset, comprising diverse B cells extracted from diseased tissues. (b) Confusion matrix for the architecture optimized TRON tailored to this task. (c) Receiver operati...

2048
[5]

& Hinton, G

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. nature 521, 436–444 (2015)

2015
[6]

Deep learning in neural networks: An overview

Schmidhuber, J. Deep learning in neural networks: An overview. Neural networks 61, 85–117 (2015)

2015
[7]

Deep learning (2016)

Goodfellow, I. Deep learning (2016)

2016
[8]

& Sun, J

He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , 770–778 (2016)

2016
[9]

& Hinton, G

Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems 25, 1097–1105 (2012)

2012
[10]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Dosovitskiy, A. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

work page internal anchor Pith review Pith/arXiv arXiv 2010
[11]

Liu, A. et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[12]

Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems 30 (2017)

2017
[13]

& Toutanova, K

Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , 4171–4186 (2019)

2019
[14]

Brown, T. et al. Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901 (2020)

1901
[15]

Jumper, J. et al. Highly accurate protein structure prediction with alphafold. nature 596, 583–589 (2021)

2021
[16]

Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. nature 542, 115–118 (2017)

2017
[17]

Jouppi, N. P. et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th annual international symposium on computer architecture , 1–12 (2017)

2017
[18]

Raina, R., Madhavan, A. & Ng, A. Y. Large-scale deep unsupervised learning using graphics processors. In Proceedings of the 26th annual international conference on machine learning , 873–880 (2009)

2009
[19]

Silvano, C. et al. A survey on deep learning hardware accelerators for heterogeneous hpc platforms. ACM Computing Surveys 57, 1–39 (2025)

2025
[20]

& Hernandez, D

Amodei, D. & Hernandez, D. Ai and compute. Tech. Rep. (2018)

2018
[21]

& McCallum, A

Strubell, E., Ganesh, A. & McCallum, A. Energy and policy considerations for deep learning in nlp. In Proceedings of the 57th annual meeting of the association for computational linguistics , 3645–3650 (2019)

2019
[22]

Kaplan, J. et al. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

work page internal anchor Pith review Pith/arXiv arXiv 2001
[23]

Patterson, D. et al. Carbon emissions and large neural network training. arXiv:2104.10350 (2021)

work page internal anchor Pith review arXiv 2021
[24]

McMahon, P. L. The physics of optical computing. Nature Reviews Physics 5, 717–734 (2023)

2023
[25]

Wetzstein, G. et al. Inference in artificial intelligence with deep optics and photonics. Nature 588, 39–47 (2020)

2020
[26]

Fu, T. et al. Optical neural networks: progress and challenges. Light: Science & Applications 13, 263 (2024)

2024
[27]

Brunner, D. et al. Roadmap on neuromorphic photonics. arXiv preprint arXiv:2501.07917 (2025)

work page arXiv 2025
[28]

Shastri, B. J. et al. Photonics for artificial intelligence and neuromorphic computing. Nature Photonics 15, 102–114 (2021)

2021
[29]

Bente, I. et al. The potential of multidimensional photonic computing. Nature Reviews Physics 7, 439–450 (2025)

2025
[30]

Liu, J. et al. Research progress in optical neural networks: theory, applications and developments. PhotoniX 2, 1–39 (2021)

2021
[31]

Bernstein, L. et al. Single-shot optical neural network. Science Advances 9, eadg7904 (2023)

2023
[32]

Huang, C. et al. A silicon photonic–electronic neural network for fibre nonlinearity compensation. Nature Electronics 4, 837–844 (2021)

2021
[33]

& Ozcan, A

Mengu, D. & Ozcan, A. All-optical phase recovery: diffractive computing for quantitative phase imaging. Advanced Optical Materials 10, 2200281 (2022)

2022
[34]

& Ozcan, A

Li, Y., Luo, Y., Bai, B. & Ozcan, A. Analysis of diffractive neural networks for seeing through random diffusers. IEEE Journal of Selected Topics in Quantum Electronics 29, 1–17 (2022)

2022
[35]

Wang, T. et al. An optical neural network using less than 1 photon per multiplication. Nature Communications 13, 123 (2022)

2022
[36]

Lin, X. et al. All-optical machine learning using diffractive deep neural networks. Science 361, 1004–1008 (2018)

2018
[37]

& Ozcan, A

Chen, S., Li, Y., Wang, Y., Chen, H. & Ozcan, A. Optical generative models. Nature 644, 903–911 (2025)

2025
[38]

Kalinin, K. P. et al. Analog optical computer for ai inference and combinatorial optimization. Nature 645, 354–361 (2025)

2025
[39]

Xia, F. et al. Nonlinear optical encoding enabled by recurrent linear scattering. Nature Photonics 18, 1067–1075 (2024)

2024
[40]

& Cao, H

Han, S., Xia, F., Gigan, S., Loureiro, B. & Cao, H. Optical kernel machine with programmable nonlinearity. arXiv preprint arXiv:2511.17880 (2025)

work page arXiv 2025
[41]

U., Oguz, I., Psaltis, D

Yildirim, M., Dinc, N. U., Oguz, I., Psaltis, D. & Moser, C. Nonlinear processing with linear optics. Nature Photonics 18, 1076–1082 (2024)

2024
[42]

Wanjura, C. C. & Marquardt, F. Fully non-linear neuromorphic computing with linear wave scattering. In AI and Optical Data Sciences VI , vol. 13375, 101–104 (SPIE, 2025)

2025
[43]

& Feng, L

Wu, T., Li, Y., Ge, L. & Feng, L. Field-programmable photonic nonlinearity. Nature Photonics 1–8 (2025)

2025
[44]

Al-Kayed, N. et al. Programmable 200 gops hopfield-inspired photonic ising machine. Nature 648, 576–584 (2025)

2025
[45]

Hua, S. et al. An integrated large-scale photonic accelerator with ultralow latency. Nature 640, 361–367 (2025). 14

2025
[46]

Ahmed, S. R. et al. Universal photonic artificial intelligence acceleration. Nature 640, 368–374 (2025)

2025
[47]

Choi, M. et al. Transferable polychromatic optical encoder for neural networks. Nature Communications 16, 5623 (2025)

2025
[48]

Shen, Y. et al. Deep learning with coherent nanophotonic circuits. Nature Photonics 11, 441 (2017)

2017
[49]

W., Minkov, M., Shi, Y

Hughes, T. W., Minkov, M., Shi, Y. & Fan, S. Training of photonic neural networks through in situ backpropagation and gradient measurement. Optica 5, 864–871 (2018)

2018
[50]

Momeni, A. et al. Training of physical neural networks. Nature 645, 53–61 (2025)

2025
[51]

& Sharma, U

Bahri, Y., Dyer, E., Kaplan, J., Lee, J. & Sharma, U. Explaining neural scaling laws. Proceedings of the National Academy of Sciences 121, e2311878121 (2024)

2024
[52]

Imaging and computing with disorder

Gigan, S. Imaging and computing with disorder. Nature Physics 18, 980–985 (2022)

2022
[53]

Saade, A. et al. Random projections through multiple optical scattering: Approximating kernels at the speed of light. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 6215–6219 (IEEE, 2016)

2016
[54]

W., Del Hougne, P., De Rosny, J., Lerosey, G

Matthès, M. W., Del Hougne, P., De Rosny, J., Lerosey, G. & Popoff, S. M. Optical complex media as universal reconfigurable linear operators. Optica 6, 465–472 (2019)

2019
[55]

Ohana, R. et al. Kernel computations from large-scale random features obtained by optical processing units. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 9294–9298 (IEEE, 2020)

2020
[56]

& Krzakala, F

Launay, J., Poli, I., Boniface, F. & Krzakala, F. Direct feedback alignment scales to modern deep learning tasks and architectures. Advances in neural information processing systems 33, 9346–9360 (2020)

2020
[57]

Wang, Z. et al. Streamlined optical training of large-scale modern deep learning architectures with direct feedback alignment. arXiv preprint arXiv:2409.12965 (2024)

work page arXiv 2024
[58]

& Weinberger, K

Huang, G., Liu, Z., Van Der Maaten, L. & Weinberger, K. Q. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , 4700–4708 (2017)

2017
[59]

& Alemi, A

Szegedy, C., Ioffe, S., Vanhoucke, V. & Alemi, A. Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence , vol. 31 (2017)

2017
[60]

& Gigan, S

Rafayelyan, M., Dong, J., Tan, Y., Krzakala, F. & Gigan, S. Large-scale optical reservoir computing for spatiotemporal chaotic systems prediction. Physical Review X 10, 041037 (2020)

2020
[61]

& Gigan, S

Dong, J., Rafayelyan, M., Krzakala, F. & Gigan, S. Optical reservoir computing using multiple light scattering for chaotic systems prediction. IEEE Journal of Selected Topics in Quantum Electronics 26, 1–12 (2019)

2019
[62]

& Soriano, M

Van der Sande, G., Brunner, D. & Soriano, M. C. Advances in photonic reservoir computing. Nanophotonics 6, 561–576 (2017)

2017
[63]

Zoph, B. & Le, Q. V. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578 (2016)

work page Pith review arXiv 2016
[64]

Elsken, T., Metzen, J. H. & Hutter, F. Neural architecture search: A survey. Journal of Machine Learning Research 20, 1–21 (2019)

2019
[65]

Chitty-Venkata, K. T. & Somani, A. K. Neural architecture search survey: A hardware perspective. ACM Computing Surveys 55, 1–36 (2022)

2022
[66]

Wright, L. G. et al. Deep physical neural networks trained with backpropagation. Nature 601, 549–555 (2022)

2022
[67]

Popoff, S. M. et al. Measuring the transmission matrix in optics: An approach to the study and control<? format?> of light propagation in disordered media. Physical review letters 104, 100601 (2010)

2010
[68]

Yang, J. et al. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data 10, 41 (2023)

2023
[69]

C., Chandelier, E

Feltes, B. C., Chandelier, E. B., Grisci, B. I. & Dorn, M. Cumida: An extensively curated microarray database for benchmarking and testing of machine learning approaches in cancer research. Journal of Computational Biology 26, 376– 386 (2019)

2019
[70]

Brongersma, M. L. et al. The second optical metasurface revolution: moving from science to technology. Nature Reviews Electrical Engineering 2, 125–143 (2025)

2025
[71]

Rocha, J. C. et al. Fast and light-eﬀicient wavefront shaping with a mems phase-only light modulator. Optics Express 32, 43300–43314 (2024)

2024
[72]

& Wang, L

Gao, L., Liang, J., Li, C. & Wang, L. V. Single-shot compressed ultrafast photography at one hundred billion frames per second. Nature 516, 74–77 (2014)

2014
[73]

Nakagawa, K. et al. Sequentially timed all-optical mapping photography (stamp). Nature Photonics 8, 695–700 (2014)

2014
[74]

Github repository, https://github.com/comediaLKB/trainable-reconfigurable-optical-network (2026)

2026