Reconfigurable nonlinear optical computing device for retina-inspired computing

Bing Xiong; Changzheng Sun; Hongtao Li; Hualong Ren; Jian Wang; Jiyuan Zheng; Lai Wang; Lin Gan; Peiyuan Zhao; Xiangwei Zeng

REVIEW 5 major objections 6 minor 1 references

Reconfigurable nonlinear optical computing device for retina-inspired computing

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A VCSEL biased below its lasing threshold acts as a reconfigurable optical sigmoid activation function at 3–250 μW input power, and simulations show it can match LeakyReLU on MNIST while also reconfiguring to preprocess degraded images.

desk verdict Credible static demonstration of a reconfigurable sigmoid transfer function in a subthreshold VCSEL, but the ONN claims rest on simulations of fitted curves, not dynamic device behavior. read the letter →

arxiv 2502.05410 v1 pith:DKRIMZ3V submitted 2025-02-08 physics.optics physics.app-ph

Xiayang Hua , Jiyuan Zheng , Peiyuan Zhao , Hualong Ren , Xiangwei Zeng , Zhibiao Hao , Changzheng Sun , Bing Xiong

show 6 more authors

Yanjun Han Jian Wang Hongtao Li Lin Gan Yi Luo Lai Wang

This is my paper · ORCID

classification physics.opticsphysics.app-ph PACS 42.55.Px42.65.-k

keywords VCSELopticalneuralnetworknonlinearactivationfunctionsigmoidreconfigurablephotonicssub-thresholdoperationretina-inspiredcomputingMNISTsimulation

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 claims that a vertical-cavity surface-emitting laser (VCSEL) biased below its lasing threshold acts as an optical sigmoid-type nonlinear activation unit that can be reconfigured by adjusting the injection current and the input wavelength. Measured transmission curves are sigmoidal at input powers of only 3–250 μW, which the paper contrasts with the milliwatt-level thresholds of several earlier all-optical activation schemes, and the nonlinear threshold shifts with bias current at 15 μW·mm²/mA. When the fitted device nonlinearity is used as a hidden-layer activation in a simulated network, MNIST accuracy reaches 97.3%, comparable to LeakyReLU and above the 92.1% of a linear network. The same reconfigurability is then applied to preprocess degraded images, recovering accuracy on over-exposed, under-exposed, and noisy inputs. The upshot is a low-power, programmable optical activation element for neural networks that also emulates retinal adaptation.

What carries the argument

The central object is the sub-threshold VCSEL acting as an optical amplifier with an input-dependent gain peak. The load-bearing identity is the positive feedback loop: input power increases stimulated recombination, which suppresses non-radiative recombination and heating, which blue-shifts the gain peak, which reduces the input wavelength's detuning from the peak, which raises gain, until electrical injection saturates. This produces the sigmoid shape. The reconfigurability comes from the two control knobs, bias current density and wavelength detuning $\Delta\lambda$, and the paper's four-parameter sigmoid is the mathematical surrogate used to insert the device into neural-network simulations.

What would settle it

Measure the VCSEL's input-output transmission in real time with a modulated input spanning 3–250 μW at detunings of −0.07 to −0.15 nm and compare against the fitted four-parameter sigmoid; if the dynamic curves deviate from the fits by more than the static fit residual, particularly if the output continues to rise beyond the fitted saturation or the threshold power drifts under thermal cycling, then the activation model used in the simulations does not describe the deployed device.

Watch

Extended reading notes

Core claim

At the center of the paper is a single experimental observation: when a VCSEL is operated just below threshold and probed with continuous-wave light near 1550 nm tuned slightly to the blue side of its gain peak, its reflected and amplified output is a sigmoidal function of input power. The mechanism is carrier competition: stronger injected light consumes more carriers through stimulated amplification, reducing non-radiative recombination and local heating, which blue-shifts the gain peak; the reduced detuning then amplifies the input further until electrical-injection saturation caps the output. The gain-peak shift makes the transfer curve S-shaped, and because the start and steepness of the curve depend on bias current density and wavelength detuning, the function is programmable. The paper fits these curves with a four-parameter sigmoid and shows, by simulation, that the fitted functions work as neural activation units and as scenario-specific image-preprocessing nonlinearities.

Load-bearing premise

The whole network demonstration rests on the assumption that the four-parameter sigmoid curves fitted to a few static measurements of the VCSEL capture how the device actually behaves when placed inside a working optical network, including with modulated signals and over the full 3–250 μW range.

Editorial extensions

If this is right

A sub-threshold VCSEL can serve as a low-power, continuous-wave optical activation unit, potentially removing the need for electrical converters at each nonlinear node.
Because the same physical device can be reprogrammed by changing bias current and detuning, one hardware platform can implement different activation curves in different layers or at different times.
The demonstrated tuning sensitivity of 15 μW·mm²/mA means the nonlinear threshold can be shifted by tens of microwatts with modest current changes, enabling adaptation of the activation point.
In the paper's simulations, fitted device nonlinearities support 97.3% MNIST accuracy and recover 19.2 to 41.8 percentage points of accuracy on degraded images, indicating the device could serve as an optical preprocessing front end.
Continuous-wave operation makes the scheme compatible with conventional, non-spiking optical neural network architectures rather than only pulsed spiking networks.

Reading between the lines

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

The gain-peak blue-shift mechanism implies the nonlinearity is thermally and carrier coupled; if the bias current is modulated faster than the thermal relaxation time, the transfer curve may lag or drift, so dynamic operation needs separate characterization.
Because the sigmoid parameters are controlled by current, a VCSEL array could be programmed per pixel as an adaptive retina-like front end, potentially performing contrast enhancement or noise gating before the network's learned weights.
The simulations substitute fitted static curves for the device; the natural next test is to pass real modulated optical signals through the VCSEL and train on the measured transfer, including saturation and noise, which would test whether the 97.3% figure survives hardware-in-the-loop.
The four-parameter sigmoid's linear term suggests part of the incident light bypasses the nonlinear gain; an optimized cavity or anti-reflection design could reduce this leakage and sharpen the activation.

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.

Desk Editor's Note

read the letter

Quick take: this is a credible static demonstration of a reconfigurable sigmoid-like transfer function in a subthreshold VCSEL, with a plausible thermal mechanism and a useful tuning range. The MNIST and retina-preprocessing results are simulations driven by fitted curves, not hardware demonstrations, so the "neural network activation" claim is only partly supported.

What's new: using a subthreshold VCSEL as an all-optical sigmoid activation, with current-controlled reconfigurability and low input power (3–250 µW). Prior work used SOAs, FP lasers, MRRs, and so on; this VCSEL approach is distinct. The systematic measurements of gain-peak blue shift with input power (Fig. 2d–f) and the tuning sensitivity of 15 µW·mm²/mA are solid, reproducible contributions. The paper is clearly written and the comparison table is useful.

Soft spots, in order of severity. First, no dynamic characterization. The nonlinearity is thermal in origin: input power reduces non-radiative heating, blue-shifts the gain, and creates positive feedback. That mechanism has a thermal time constant, so the static CW transfer curve may not hold for modulated signals. The paper does not report bandwidth, pulse response, or any measurement under time-varying input. The MNIST and preprocessing simulations use the static fits, so they demonstrate the fitted function, not the device in a real network. This is a common gap in this literature, but here it is load-bearing because the paper explicitly claims suitability for ONNs.

Second, missing error bars on the transfer curves and fit parameters. The fits in Fig. 4 look good, but quantitative claims about threshold and sensitivity would be stronger with uncertainty estimates.

Third, Table 1 lists "3–250 µW" as the nonlinear threshold range, but the text and Fig. 3 show thresholds between 60 and 135 µW for the tested detunings; 3 µW is the start of the input scan, not a threshold. The abstract's phrasing is also ambiguous. This should be cleaned up.

Fourth, the preprocessing improvement (41.76%, 19.2%, 25.89%) involves post-hoc selection of specific nonlinear functions for each degradation type. That is fine as an illustration, but it is not a systematic study of adaptability; a reader cannot tell how sensitive the gains are to the choice of function.

None of these are fatal to the core finding. The static nonlinearity is real, the physics explanation is coherent, and the idea is worth pursuing. Who it is for: people working on optical activation functions and photonic neural networks will want to read it. My recommendation: send to peer review, but require dynamic characterization or a clear downgrade of the integration claims. If they add a bandwidth measurement, even a rough one, the paper becomes much stronger.

Referee Report

5 major / 6 minor

Summary. The paper proposes a subthreshold-biased VCSEL as a reconfigurable optical nonlinear activation unit. Quasi-static CW measurements show sigmoidal input-output transfer curves whose threshold and steepness change with bias current and wavelength detuning; the curves are fitted to a four-parameter sigmoid (Eq. 2). The fitted functions are then used in simulations as activation functions in a fully connected MNIST classifier, reporting 97.3% accuracy comparable to LeakyReLU, and as image-preprocessing nonlinearities that improve recognition of under-exposed, over-exposed, and noisy MNIST images. The authors claim low input power (3–250 μW), high tuning sensitivity (15 μW·mm²/mA), and reconfigurability suitable for retina-inspired adaptive computing.

Significance. If the static measurement and the transfer-function fidelity under realistic operating conditions both hold, the device would be a useful addition to the menu of low-power reconfigurable all-optical nonlinear elements for optical neural networks. The paper's strengths are that the nonlinearity is directly measured rather than assumed, the physical mechanism is stated in terms of a rate-equation model, and the simulation results are compared against a standard nonlinearity (LeakyReLU). The central limitation is that the device-level demonstration is quasi-static CW only, whereas the neural-network and preprocessing claims assume that the fitted sigmoid remains valid under time-varying inputs and in a multi-device network; that assumption is not tested. The reported simulation numbers are therefore demonstrations of the fitted function rather than of the deployed device, which limits the strength of the 'suitable for neural-network activation' claim.

major comments (5)

[Results, Eq. (2), Figs. 4–6] The central claim that the device can serve as a neural-network activation unit rests on the four-parameter sigmoid fit of Eq. (2) to quasi-static CW measurements, but no dynamic characterization is reported. Because the invoked mechanism is thermal—the input-induced reduction of non-radiative heating blue-shifts the gain peak—the transfer function is expected to be history-dependent with a thermal time constant. At modulation rates faster than that time constant, the instantaneous response could differ from the fitted sigmoid, and cascaded stages would accumulate the deviation. Please provide dynamic measurements (e.g., modulated input around the operating point at relevant rates) or explicitly restrict the claim to CW/quasi-static operation and adjust the abstract and discussion accordingly.
[Abstract, Fig. 3, Table 1] The claimed operating range of 3–250 μW is not supported by the data shown in the manuscript. The only explicit nonlinear-threshold values reported in the text are 60 μW and 135 μW (for bias current densities of 165 and 160 mA/mm² at Δλ = −0.11 nm), and no transfer curve is shown spanning the full 3–250 μW range. Please report how the 3 μW and 250 μW endpoints were obtained, for which bias/detuning conditions, and provide a table of all measured thresholds with uncertainties.
[Figs. 3–4, Eq. (2)] The manuscript does not report measurement uncertainty, number of repetitions, or fit-quality metrics for the transfer curves. Without error bars, and without fit parameters and residuals for each detuning, the sigmoid characterization and the claimed reconfigurability (e.g., threshold tuning from 60 to 135 μW) cannot be quantitatively assessed. Please add error bars to Figs. 3–4 and report the fitted values of A, B, C, D with uncertainties for every condition shown in Fig. 4.
[Fig. 5, simulation setup] The MNIST simulation lacks essential details needed to evaluate the 97.3% accuracy claim: the training/test split, number of epochs, learning rate, batch size, optimizer, initialization, and number of independent runs are not given. In addition, the simulation uses one fitted curve for all 500 nonlinear hidden units, whereas a physical implementation would involve many devices with slightly different transfer functions. Please provide the full simulation protocol and, ideally, a sensitivity analysis that varies the fitted parameters across the units.
[Fig. 6, preprocessing demonstration] The procedure for selecting the three preprocessing nonlinear functions is not described. If the functions were chosen after evaluating accuracies on the contaminated test images, the reported improvements (65.97%→91.86%, 19.96%→61.72%, 77.80%→97.00%) would be optimistically biased. Please specify the selection rule (e.g., validation-set search) and state whether the test set was used at any point in choosing the preprocessing functions.

minor comments (6)

[Eq. (1)] Equation (1), the rate-equation model, is garbled in the manuscript and does not render legibly; please reproduce it cleanly so that the physical mechanism can be checked.
[Eq. (2)] Equation (2), the four-parameter sigmoid, is also garbled in the displayed text. Although the roles of A, B, C, and D are described, the explicit functional form must be visible for the fitting results to be verifiable.
[Fig. 4 caption] The caption says that panels (a)–(f) show Δλ values of −0.07, −0.09, −0.11, −0.13, and −0.15 nm, which is only five detunings for six panels; please reconcile the number of panels and the listed detunings.
[Fig. 2 caption and text] The subscripts distinguishing the gain-peak wavelength and the spontaneous-emission-peak wavelength are missing in the caption and in the body text (e.g., "λ" and "λ" appear without subscripts), making the description of the blue-shift effect harder to follow.
[Fig. 6d and text] The text alternates between "Over-exposed" and "Over-fixed" ("Over-fixed: 19.96%-61.72%"), and "Under-exposed" and "Under-fixed"; please use consistent terminology throughout.
[Table 1] The table column for the input model shows "CW" and "pulse" but the comparison text would benefit from a consistent definition of "nonlinear threshold power" and a note on how it was extracted in each cited work; also, some entries (e.g., "PD+MZI ~ 0.1mW ~ Yes") contain placeholder tildes.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: measured transfer curves are fitted and then applied as simulated activations, a conditional modeling step rather than a self-referential derivation.

full rationale

The paper's central claim is an experimental demonstration: a subthreshold VCSEL under optical injection yields a measured, reconfigurable sigmoid-shaped transmission curve. The nonlinearity is established by spectrometer and power measurements (Figs. 2 and 3), and the laser rate-equation argument in Eq. (1) is invoked only to explain the observed blue-shift mechanism, not to derive the transfer curve. The subsequent MNIST and image-preprocessing demonstrations use the four-parameter sigmoid of Eq. (2) fitted to those static measurements (Fig. 4) as the activation function in simulations. This is a conditional modeling step, not circularity: the simulation result (97.3% accuracy) depends on the fitted curve but is not equal to the fit by construction, and the paper explicitly labels the curves as "fitting functions" and the demonstrations as simulations. No load-bearing self-citation, uniqueness theorem, or ansatz-smuggling-through-citation is present; the cited rate equation and prior works are standard background. The lack of dynamic or end-to-end device validation is an extrapolation or completeness concern, not a circularity. Therefore no circular step is identified.

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

The paper rests on standard VCSEL physics for the mechanism, but the quantitative claims come from measured transfer curves and their four-parameter fits. The neural-network simulations add the fitted parameters as free inputs; no new physical entities are invented.

free parameters (4)

A (sigmoid amplitude) = not reported per condition
Used in Eq. (2) to fit measured transfer curves; values vary with detuning and current density but are not tabulated.
B (sigmoid steepness) = not reported per condition
Controls the sharpness of the transition; fitted to data.
C (nonlinear threshold center) = not reported per condition
Sets the input power at the sigmoid midpoint; fitted to data.
D (linear term coefficient) = not reported per condition
Accounts for unamplified reflected light; fitted to data.

assumptions (3)

domain assumption Carrier rate equation (1) with terms for injection, spontaneous, radiative, non-radiative, Auger, and stimulated recombination.
Invoked in Results to explain the nonlinear behavior; standard for VCSELs but not derived here.
domain assumption The gain-peak blue shift with increasing input power is dominated by a reduction in local temperature caused by reduced non-radiative recombination.
The paper's explanation of the positive feedback mechanism; no direct temperature measurement is provided.
ad hoc to paper The four-parameter sigmoid in Eq. (2) accurately represents the device's transfer function under all operating conditions used in the simulations.
The neural-network simulations and preprocessing demonstrations rely on this fit being valid; only single-point fits are shown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reconfigurable nonlinear optical computing device for retina-inspired computing." pith.science (2026). https://pith.science/paper/DKRIMZ3V

@misc{pith2026250205410,
  author       = {Pith},
  title        = {Pith review of: Reconfigurable nonlinear optical computing device for retina-inspired computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKRIMZ3V}},
  note         = {Machine review of arXiv:2502.05410}
}

read the original abstract

Optical neural networks are at the forefront of computational innovation, utilizing photons as the primary carriers of information and employing optical components for computation. However, the fundamental nonlinear optical device in the neural networks is barely satisfied because of its high energy threshold and poor reconfigurability. This paper proposes and demonstrates an optical sigmoid-type nonlinear computation mode of Vertical-Cavity Surface-Emitting Lasers (VCSELs) biased beneath the threshold. The device is programmable by simply adjusting the injection current. The device exhibits sigmoid-type nonlinear performance at a low input optical power ranging from merely 3-250 {\mu}W. The tuning sensitivity of the device to the programming current density can be as large as 15 {\mu}W*mm2/mA. Deep neural network architecture based on such device has been proposed and demonstrated by simulation on recognizing hand-writing digital dataset, and a 97.3% accuracy has been achieved. A step further, the nonlinear reconfigurability is found to be highly useful to enhance the adaptability of the networks, which is demonstrated by significantly improving the recognition accuracy by 41.76%, 19.2%, and 25.89% of low-contrast hand-writing digital images under high exposure, low exposure, and high random noise respectively.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references

[1]

2BeijingNationalResearchCenterforInformationScienceandTechnology(BNRist)

Reconfigurable nonlinear optical computing device for retina-inspired computing XiayangHua1, JiyuanZheng*1,2✉,PeiyuanZhao3,HualongRen4,XiangweiZeng4,ZhibiaoHao1,2, ChangzhengSun1,2,BingXiong1,YanjunHan1,JianWang1,HongtaoLi1,LinGan1,YiLuo1,2and LaiWang*1,2✉ 1DepartmentofElectronicEngineering,TsinghuaUniversity,Beijing100084,China. 2BeijingNationalResearchC...

2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

2BeijingNationalResearchCenterforInformationScienceandTechnology(BNRist)

Reconfigurable nonlinear optical computing device for retina-inspired computing XiayangHua1, JiyuanZheng*1,2✉,PeiyuanZhao3,HualongRen4,XiangweiZeng4,ZhibiaoHao1,2, ChangzhengSun1,2,BingXiong1,YanjunHan1,JianWang1,HongtaoLi1,LinGan1,YiLuo1,2and LaiWang*1,2✉ 1DepartmentofElectronicEngineering,TsinghuaUniversity,Beijing100084,China. 2BeijingNationalResearchC...

2019