Pith. sign in

REVIEW 5 major objections 6 minor 17 references

Neuromorphic Photonic Processing and Memory with Spiking Resonant Tunnelling Diode Neurons and Neural Networks

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A resonant tunnelling diode, biased at its negative differential resistance region, can act as an ultrafast event detector and, in simulation, as the neuron of photonic classifiers and tunable optical memories.

desk verdict Useful RTD proof-of-concept with a real experimental demo, but the quantitative claims need held-out tests and controls. read the letter →

arxiv 2507.20866 v1 pith:7GTUC4MI submitted 2025-07-28 physics.comp-ph

Dafydd Owen-Newns , Joshua Robertson , Giovanni Donati , Jose Figueiredo , Edward Wasige , Kathy Ludge , Bruno Romeira , Antonio Hurtado This is my paper · ORCID
classification physics.comp-ph
keywords resonanttunnellingdiodeneuromorphicphotonicsphotonicspikingneuralnetworkedgedetectionextremelearningmachinefadingmemoryMackey-Glasstimeseriesoptoelectronicneuron
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 argues that a single light-sensitive resonant tunnelling diode (RTD) can serve as a multi-modal, optical-and-electronic spiking neuron that detects fast rising edges in time-series data. The authors demonstrate this experimentally on a standard chaotic time series: the diode fired at 16 of 20 target edges and produced no false positives. They then use numerical simulation to extend the same RTD neuron to larger architectures: 20 uncoupled RTDs form a photonic spiking extreme learning machine that classifies a benchmark flower dataset with up to 96.5% accuracy, and 10 coupled RTD-laser neurons form a fading memory whose storage time is set by optical attenuation. If accurate, RTDs would provide a single hardware platform for event detection, classification, and tunable short-term memory at gigahertz rates in telecom wavelength bands.

What carries the argument

The load-bearing object is the RTD spiking neuron modelled as a lumped circuit, with capacitance, inductance, resistance, and the diode's nonlinear current-voltage relation, coupled to two-level laser rate equations to form an RTD-laser neuron. Its defining mechanism is threshold-controlled excitability: near the peak or valley of the current-voltage curve a small perturbation triggers an all-or-nothing spike, and the required perturbation amplitude moves with bias voltage. That single mechanism does triple duty: dual photonic-electronic modulation computes the difference between a raw signal and its delayed copy as a rising-edge detector; the same threshold acts as a binary Heaviside activation in the extreme learning machine; and a spread of bias thresholds across ten coupled devices, combined with attenuated feedback, sets how many cycles a stored spike survives.

What would settle it

Repeat the experiment at the simulated speeds: drive a photo-detecting RTD with 20 picosecond optical pulses at biases across its peak and valley, and compare the measured spike/no-spike boundary with the model's threshold map; then run the chaotic time-series edge-detection with 500 picosecond pulse separation and the ten-step delay. If the boundary, the detected edges, or the roughly 300 picosecond refractory period disagrees with the model, the gigahertz-rate claims do not transfer to hardware.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the excitability of an RTD, rooted in its steep N-shaped current-voltage curve with a negative differential resistance region, can be exploited as a controllable threshold for spiking, and that this threshold can be modulated simultaneously by light and by voltage. Encoding a raw time series in optical pulse amplitude and its delayed copy in bias voltage makes the diode fire precisely when the delayed difference exceeds threshold, so the device computes a temporal derivative in hardware. The same thresholding behaviour, read across a spatial array, implements a Heaviside activation for a photonic extreme learning machine; and when RTD outputs are reinjected through an attenuating all-to-all feedback loop, the number of regenerated cycles becomes a tunable memory depth. The experimental edge-detection result, 16 of 20 edges with no false positives, is the paper's direct evidence for the single-neuron claims; the network results are numerical.

Load-bearing premise

The numerical results rest on the assumption that the fitted lumped-circuit RTD model and the laser rate equations remain quantitatively correct at the simulated gigahertz rates and for ten to twenty coupled devices; if real high-speed spiking dynamics diverge from those equations, the 96.5% classification accuracy and the memory-depth curves are predictions about a model rather than about hardware.

Editorial extensions

If this is right

  • A single RTD neuron can act as an ultrafast, tunable alarm or change-detector for time-series data, with its sensitivity set by either optical power or bias voltage.
  • Because the edge detector operates in both the optical and electronic domains, it can accept a raw optical signal and produce a spiking electronic output without a separate analogue-to-digital conversion step.
  • A 20-neuron RTD array can classify a benchmark dataset at gigahertz rates while training only the output layer, with peak simulated accuracy of 96.5%.
  • A coupled network of ten RTD-laser neurons can store a spike for a controllable number of feedback cycles, with the memory depth set simply by changing optical attenuation.
  • The same RTD hardware can be reconfigured between processing and memory roles because both rely on the same voltage-tunable spiking threshold.

Reading between the lines

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

  • Editorial extension: the dual-modulation edge detector could be integrated into a photonic circuit with a fixed optical delay line, making the temporal-difference operation passive and potentially faster than the electronic delay used in the experiment.
  • Editorial extension: the binary 'node significance' training result suggests that a hardware RTD array could be reconfigured with simple switches rather than analogue weight memories, since only a handful of binary output weights are needed.
  • Editorial extension: the attenuation-tuned fading memory cell could serve as a short-term buffer in a reservoir computer, although the paper itself only demonstrates storage and tuning, not readout of stored information into a downstream task.
  • Editorial extension: the experimental demonstration encoded each time step with 200 ns-long pulses while the simulations use 20 ps pulses; pushing the experiment toward the simulated gigahertz rates would reveal whether parasitic electrical effects, rather than the diode physics, set the practical speed limit.
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

5 major / 6 minor

Summary. This paper proposes and numerically/experimentally investigates spiking resonant tunnelling diode (RTD) neurons for neuromorphic photonic processing. It reports (i) an experimental demonstration of rising-edge detection in a Mackey-Glass time series using dual photonic-electronic modulation of a single RTD (16/20 edges, no false positives), (ii) numerical simulations of a 20-neuron RTD-based extreme learning machine for Iris classification (up to 96.5% accuracy), and (iii) numerical simulations of a 10-neuron RTD-laser fading memory cell with attenuation-tunable memory depth. The paper uses a lumped-circuit RTD model from prior work and standard laser rate equations.

Significance. The work is potentially significant because it demonstrates that a single RTD can perform a spiking computation on real telecom-wavelength optical input, and it proposes scalable architectures (ELM, memory cell) that exploit RTD's speed. Strengths include use of a previously fitted model (Table 1), a data DOI, and a clear physical description of the dual-modulation idea. However, the experimental evidence is a single trace with a post-hoc threshold, and the numerical results lack reproducibility details (random matrix seed, laser parameters). If the experimental result is confirmed with controls and the simulations are made reproducible, the paper would constitute a useful step toward ultrafast photonic spiking hardware. As it stands, the evidence supports a proof-of-concept rather than the strong performance claims in the abstract.

major comments (5)
  1. [§3.1, Figs. 5–6] The experimental rising-edge detection is a single demonstration on one Mackey-Glass segment. The delayed-difference threshold (0.32) is identified by inspecting the target edges on that same trace, and the optical/electrical encoding parameters (2 mW average power, -0.3 V modulation) are tuned on the same segment. With no held-out segment, no repeated trials, and no control condition (e.g., a scrambled delayed signal), the 16/20 detection rate with zero false positives cannot be distinguished from threshold fitting. Please report the spike classification criterion (how a transient in the DC-filtered trace is scored as a spike), repeat the experiment on a fresh MG segment with fixed parameters, and add a control that disrupts the delayed-difference relation.
  2. [§3.1, Fig. 4] The numerical edge-detection also selects the threshold (0.45) from the same time series. The sentence 'Once the RTD spiking threshold was correctly tuned, the event-based rising edge detection system was able to correctly identify all target features' is circular if the tuning uses the known target locations. Please specify a threshold-selection rule that does not use the target labels, or reframe the demonstration as a parameter exploration rather than a blind detection.
  3. [§3.2.1, Figs. 7–8] The ELM classification results are not reproducible as reported. The random input weight matrix (Nf x N) is neither specified nor fixed by a seed; the accuracy curves in Fig. 8(c,d) show single runs; and no error bars or repeated random realizations are given. The claim of 'consistently above 93%' is not supported without a spread of results. Please provide the random matrix or a reproducible seed, report mean ± std over multiple realizations, and compare against a linear baseline on the raw four features.
  4. [§3.2.2, Eqs. (4)–(5) and Fig. 10] The laser rate-equation parameters (J, η, γm, γl, γnr, τph, N0) are not listed, although the RTD parameters are. The memory cell also assumes that the all-to-all optical coupling is equivalent to each neuron receiving the mean of all outputs through a single attenuator; this is an ansatz that should be justified or relaxed. Without parameter values and a coupling-model check, the quantitative memory-depth curves (Fig. 10c) are not reproducible.
  5. [§3.2.2, Fig. 10] The bias distribution B_n = 0.9 − 0.185 n^2 is introduced without derivation or sensitivity analysis. Because the memory depth is controlled by the competition between bias thresholds and feedback attenuation, the claim of tunability should be supported by showing the dependence of memory depth on the bias distribution, not only on attenuation. Also, define how memory depth is measured (number of cycles until no neuron spikes, or something else) and how the 131 ps spike activation delay in the autaptic case is computed.
minor comments (6)
  1. [Title and author list] The title contains a typo: 'R es-' should be 'Res-', and the author name has an odd spacing 'Jos´ e F igueiredo'.
  2. [§5, Experimental Methods] The sentence 'A bias point in the peak was chosen... shown in Figure 7 a)' should refer to Figure 1(a), not Figure 7.
  3. [§3.1, Fig. 5] The caption says 'Pulse widths of 20 ps are separated by 500 ps,' but the main text says '2 ns-long (negative) square pulses were used every 200 ns.' Please clarify the actual experimental pulse parameters.
  4. [Table 1] Units are not specified for the parameters in Table 1; state that all values are in SI units or list units per parameter.
  5. [Eq. (7)] In the Heaviside step function, T is not defined in the text; specify that T corresponds to the RTD spike threshold set by bias and optical amplitude.
  6. [§3.2.1] There is a typo: 'we use use here' should be 'we use here'.

Circularity Check

2 steps flagged · score 6.0 of 10

Edge-detection results reduce to the delayed-difference threshold used to define the targets; ELM and memory sections are independent.

  1. self definitional [Section 3.1, Figure 4 and surrounding text]
    "A threshold of 0.45 delayed difference was found to detect the sharpest rising edges in the MG times-series (see Figure 4(b)). This threshold value (0.45) was then implemented empirically by correctly choosing the optical pulse amplitude and voltage modulation levels that together produced the matching spiking output."

    The target features are defined by the paper as intervals where Dτ(t)=y(t)-y(t-τ) exceeds 0.45 (the grey regions in Figure 4b). The RTD model is then programmed by choosing optical amplitude and bias modulation so that it spikes exactly when that same delayed difference exceeds the threshold. Thus the numerical result that all target features are detected is a restatement of the threshold definition; the detector and the ground truth are the same quantity by construction.

  2. fitted input called prediction [Section 3.1, Figures 5-6 and accompanying text]
    "Here, the grey shaded regions reveal time steps where rising edges (where the delayed difference exceeds 0.32), occur in the time-series... having detected the majority of the rising edges in the time-series (16 out of 20) in total, without false positive detections."

    The optical input encodes y(t) and the electrical bias encodes y(t-τ); spiking occurs when the combined photonic-electronic drive crosses a threshold set by the chosen modulation levels and bias range. The target edges are defined as the same delayed-difference condition (Dτ>0.32). The reported 16/20 and zero false positives therefore measure how well the physical RTD implements the pre-defined Dτ threshold on a single tuned MG segment, not an independent edge-detection benchmark; the threshold and encoding were co-selected rather than fixed a priori on held-out data.

full rationale

The non-circular parts of the paper are the ELM classification and the coupled-RTD memory simulations. These use the previously published lumped-circuit RTD model and laser rate equations as a forward model, apply random fixed input weights, train only the output layer on a training split, and evaluate on a held-out test split; the reported 96.5% and 93% accuracies are genuine numerical results. The self-citations to the RTD model [22] and node-significance algorithm [35] are not load-bearing in a circular way because those prior results are independently parameterized and published. The circularity is concentrated in Section 3.1: both the numerical and experimental edge-detection demonstrations define the target rising edges by a delayed-difference threshold and then set the RTD's optical/electrical operating point to spike on that same threshold. The detection output is therefore the same comparison used to label the ground truth, so the cited detection rates are partially forced by construction. This does not eliminate the value of the experimental hardware demonstration, which shows a physical RTD can approximate the threshold operation, but it means the specific 16/20 and zero-false-positive claims are not an independent predictive benchmark.

Assumptions & free parameters 5 free parameters · 8 assumptions · 0 invented entities

The central numerical claims rest on a previously published RTD model plus several new chosen parameters (thresholds, bias distribution, unspecified laser parameters, unseeded random matrix). The experimental edge detection rests on an assumed monotonic bias-to-threshold relation that is demonstrated in-situ. No new physical entities are postulated.

free parameters (5)
  • RTD circuit model parameters (R, L, C, κ, A, b, c, d, n1, n2, H) = R=10 Ω, L=126 nH, C=0.002 pF, κ=0.1e-6 A, A=-5.5e-5 A, b=0.033 V, c=0.113 V, d=-2.8e-6 V, n1=0.185, n2=0.045, H=18e-5 A
    These values are taken from Table 1 and were fitted in ref [22] to an experimental I-V curve; the paper's numerical results (threshold map, edge detection, ELM, memory) all depend on them.
  • Laser rate-equation parameters (J, η, γm, γl, γnr, τph, N0) = not specified in the paper
    Eqs. 4-5 model the RTD-laser output used in the memory simulations, but the parameter values are not given, so the simulation is not self-contained.
  • Numerical edge-detection threshold on delayed difference = 0.45 (simulation), 0.32 (experiment)
    The threshold is chosen after inspecting the target time series, then the optical amplitude and bias modulation are set to implement it, so the detection performance is not independent of the evaluation target.
  • Bias distribution for memory cell (Bn = 0.9 - 0.185 n^2) = 0.75 V to 0.9 V claimed; formula inconsistent
    The distribution is chosen by hand to set different spiking thresholds across the 10 neurons; the printed formula does not reproduce the claimed range, indicating a typo or unstated indexing.
  • ELM random input weight matrix (Nf x N) = not provided, no seed
    The reported classification accuracies depend on a random linear mixing matrix that is simulated but neither printed nor seeded, so the exact 96.5% figure is not reproducible.
assumptions (8)
  • standard math Kirchhoff's circuit laws govern the lumped RTD circuit (Eqs. 1-2).
    Used to write the voltage-current dynamics of the neuron; standard circuit analysis.
  • domain assumption The RTD I-V characteristic is captured by the analytical f(V) in Eq. 3 with the fitted parameters of Table 1.
    The numerical model of the RTD spiking neuron is imported from ref [22]; if f(V) does not capture the device dynamics at the simulated speeds and biases, the numerical results are invalid.
  • domain assumption The RTD current directly drives a two-level semiconductor laser with coupling efficiency η (Eqs. 4-5).
    This RTD-laser coupling model is assumed for all memory simulations; no experimental verification is provided in this paper.
  • domain assumption An optical input S0(t) couples linearly to the circuit current with conversion factor κ.
    The photodetection layer of the RTD is modeled as a current source proportional to optical power (Eq. 1); this is the basis of all optical inputs.
  • ad hoc to paper The all-to-all optical coupling in the memory cell can be represented as each neuron receiving the equally weighted mean of all outputs through a single attenuator.
    Figure 10(a) assumes fan-in/fan-out optical connections are mean-field and lossless except for the global attenuator; deviations (e.g., unequal coupling, crosstalk, phase effects) are not modeled.
  • domain assumption A delay-difference threshold can be set through simultaneous optical amplitude and electrical bias modulation.
    The edge-detection method assumes the RTD's firing threshold depends monotonically on the bias as shown in Fig. 2(a), which is confirmed for the experimental device only at the tested operating point.
  • standard math The extreme learning machine with binary Heaviside activations is a valid classification architecture for Iris.
    The ELM paradigm (random fixed hidden layer, trained linear readout) is used as the algorithmic basis; the RTD threshold is mapped to the Heaviside step in Eq. 7.
  • domain assumption The node-significance algorithm from ref [35] works as described.
    Used for the binary weight training; cited to the authors' own prior work, not re-derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neuromorphic Photonic Processing and Memory with Spiking Resonant Tunnelling Diode Neurons and Neural Networks." pith.science (2026). https://pith.science/paper/7GTUC4MI

@misc{pith2026250720866,
  author       = {Pith},
  title        = {Pith review of: Neuromorphic Photonic Processing and Memory with Spiking Resonant Tunnelling Diode Neurons and Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GTUC4MI}},
  note         = {Machine review of arXiv:2507.20866}
}
read the original abstract

Neuromorphic computing-modelled after the functionality and efficiency of biological neural systems-offers promising new directions for advancing artificial intelligence and computational models. Photonic techniques for neuromorphic computing hardware are attracting increasing research interest, thanks to their potentials for ultra high bandwidths, low-crosstalk and high parallelism. Among these, approaches based upon resonant tunnelling diodes (RTDs) have recently gained attention as potential building blocks for next-generation light-enabled neuromorphic hardware, due to their capacity to replicate key neuronal behaviours such as excitable spiking and refractoriness, added to their potentials for high operational speeds, energy efficiency and compact footprints. In particular, their ability to function as opto-electronic spiking neurons makes them strong candidates for integration into novel event based neuromorphic computing systems. This work demonstrates the application of optically-triggered spiking RTD neurons to a multiplicity of applications and architectures, these include systems based upon single elements for multi-modal (photonic-electronic) fast rising edge-detection in time-series data, the construction of a two-layer feedforward artificial photonic spiking neural network (pSNN) using RTD neurons as the nonlinear nodes delivering excellent performance in complex dataset classification tasks, and a pSNN comprised of multiple coupled light-sensitive RTD spiking neurons that supports performance as an adjustable neuromorphic optical spiking memory system with a tunable storage time of spiking patterns.

Figures

Figures reproduced from arXiv: 2507.20866 by the authors.

Figure 1
Figure 1. (a) I-V curve of the experimental RTD showing a NDR regio [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Numerical analysis of multi-modal (photonic-electronic) s [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Operation principle of RTD event-based time-series rising e [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Rising edge detection of the Mackey-Glass time-series per [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Experimentally measured RTD time-series showing rising edg [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Rising edge detection of the Mackey-Glass time-series per [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Schematic of the simulated RTD neural network and analys [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Iris flower dataset classification with a simulated array of RT [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Autaptic spiking memory system with a simulated RTD-laser n [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Neuromorphic fading memory cell constructed of 10 RTD [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Experimental setup used to probe the photonic-electr [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

  1. [27]

    M. C. Mackey, L. Glass, Science 1977, 197, 4300 287

    work page 1977

  2. [28]

    Huang, Q.-Y

    G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Neurocomputing 2006, 70, 1-3 489

    work page 2006

  3. [29]

    Borghi, S

    M. Borghi, S. Biasi, L. Pavesi, Scientific reports 2021, 11, 1 15642

    work page 2021

  4. [30]

    J. R. Rausell-Campo, A. Hurtado, D. P´ erez-L´ opez, J. Capmany Francoy, Laser & photonics reviews 2025, 19, 9

    work page 2025

  5. [31]

    Owen-Newns, J

    D. Owen-Newns, J. Robertson, M. Hejda, A. Hurtado, IEEE Journal of Selected Topics in Quantum Electronics 2023, 29, 2: Optical Computing 1

    work page 2023

  6. [32]

    Picco, L

    E. Picco, L. Jaurigue, K. L¨ udge, S. Massar, Communications engineering 2025, 4, 1 3

    work page 2025

  7. [33]

    Biasi, R

    S. Biasi, R. Franchi, L. Cerini, L. Pavesi, APL photonics 2023, 8, 9 096105

    work page 2023

  8. [34]

    R. A. Fisher, Annals of Eugenics 1936, 7, 2 179

    work page 1936

Show all 17 references
  1. [35]

    Owen-Newns, J

    D. Owen-Newns, J. Robertson, M. Hejda, A. Hurtado, Intelligent Computing 2023, 2 0031

    2023

  2. [36]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Neural computation 1997, 9, 8 1735

    1997

  3. [37]

    J. M. Martins, S. V. Gurevich, J. Javaloyes, Phys. Rev. Appl. 2024, 22 024050

    2024

  4. [38]

    Romeira, R

    B. Romeira, R. Av´ o, J. M. L. Figueiredo, S. Barland, J. Javaloy es, Scientific Reports 2016, 6 19510

    2016

  5. [39]

    Terrien, V

    S. Terrien, V. A. Pammi, N. G. R. Broderick, R. Braive, G. Beaud oin, I. Sagnes, B. Krauskopf, S. Barbay, Physical review research 2020, 2, 2 023012

    2020

  6. [40]

    Terrien, V

    S. Terrien, V. A. Pammi, B. Krauskopf, N. G. R. Broderick, S. B arbay, Physical review. E 2021, 103, 1 012210

    2021

  7. [41]

    S. Boyd, L. Chua, IEEE transactions on circuits and systems 1985, 32, 11 1150

    1985

  8. [42]

    Dambre, D

    J. Dambre, D. Verstraeten, B. Schrauwen, S. Massar, Scientific Reports 2012, 2 514

    2012

  9. [43]

    Jaeger, GMD-German National Research Institute for Computer Scien ce (2002) 2002

    H. Jaeger, GMD-German National Research Institute for Computer Scien ce (2002) 2002. 19

    2002

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.