Integrated photonic deep neural network with end-to-end on-chip backpropagation training

Farshid Ashtiani; Kwangwoong Kim; Mohamad Hossein Idjadi

REVIEW 3 major objections 5 minor 48 references

Integrated photonic deep neural network with end-to-end on-chip backpropagation training

T0 review · 3 major / 5 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 This paper reports a silicon photonic chip that performs the full end-to-end backpropagation training loop in optics, matching an ideal digital model on two classification tasks.

desk verdict The chip does something real, but the paper's central claim is one step ahead of the implementation: the output-layer nonlinearity and cost live in a microcontroller, so it is not full end-to-end on-chip backpropagation as advertised. read the letter →

arxiv 2506.14575 v1 pith:YDUE7VZ4 submitted 2025-06-17 physics.optics

Farshid Ashtiani , Mohamad Hossein Idjadi , Kwangwoong Kim This is my paper · ORCID

classification physics.optics

keywords integratedphotonicneuralnetworkon-chipbackpropagationgradientdescenttrainingsiliconphotonicsReLUactivationfabricationvariationcompensationPINattenuatoroptical

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

The paper reports the first integrated photonic deep neural network, to its knowledge, whose full error-backpropagation training loop runs on a single silicon photonic chip: forward propagation through linear optical weights and opto-electronic ReLU activations, backward propagation of errors through gradient circuits, and weight updates, with no digital computer in the loop beyond microcontroller housekeeping. It claims that this on-chip training reaches accuracy on par with an ideal digital model for two nonlinear classification tasks, XOR and 2D point separation, and that it compensates for fabrication-induced device variations, which digital training cannot. The point matters because scalable photonic neural networks have been limited by training quality: offline digital backpropagation ignores actual device variations, while gradient-free methods lose the versatility of backpropagation. If the claim holds, photonic accelerators can be trained robustly on the chip that runs them.

What carries the argument

The load-bearing mechanism is the opto-electronic implementation of the rectified-linear-unit (ReLU) activation and its gradient in one PIN attenuator: in low-gain mode the device's attenuation-versus-voltage curve acts as a ReLU-like nonlinearity, and in high-gain mode the same curve acts as a step function, which is exactly σ′(z) for ReLU. The backward path uses nominally identical devices, so the gradient circuit computes the Hadamard product of the backpropagated error with the activation gradient. Linear weights are push-pull PIN attenuators feeding balanced photodetectors, giving bipolar weights from -1 to 1. The architecture uses per-layer supply light so each layer's optical power is independent, and the paper also sketches a hardware-reuse version that halves the number of nonlinear blocks and optical inputs by mode-switching each activation between inference and training.

What would settle it

Re-run the 2D point-separation experiment with a fixed, stated 50/150 train/test split, evaluate on the 150 held-out points only, and report mean and standard deviation over the five weight initializations. If the mean held-out accuracy is substantially below the 92.5% shown or below the digital-reference accuracy, the parity claim would collapse.

Watch

Extended reading notes

Core claim

The central discovery is that the missing piece of photonic backpropagation, the nonlinear activation gradient, can be implemented with the same PIN-attenuator intensity modulators used for the activation itself: a single modulator biased at high attenuation gives a ReLU-like response in low-gain mode and a step-function gradient in high-gain mode. With that piece in place, a chip containing an input layer, an 8-neuron hidden layer, and an output layer can compute both the forward pass and the backward errors δ(2) and δ(1), and update all weights by gradient descent entirely on-chip. The authors experimentally train the chip for XOR and for a 50-point subset of a 200-point two-dimensional classification problem, reaching 92.5% accuracy on the shown points, and report that repeated on-chip training runs are more robust across weight initializations than digital training followed by on-chip inference. They take this as evidence that end-to-end on-chip backpropagation compensates for measured device-to-device variations, including weight errors up to about 0.27, without any model of those variations.

Load-bearing premise

The quantitative claim of 92.5% accuracy being on par with the digital model assumes that the 200 evaluated points include points not used in the 50-point training set, but the paper never states how the points were split and reports only the best accuracy across trials, not the average.

Editorial extensions

If this is right

Photonic neural networks can now be trained with the same gradient-descent backpropagation used for digital networks, so training no longer has to be exiled to a digital computer with a simplified device model.
Because training happens on the hardware that will do inference, fabrication-induced device-to-device variations are absorbed by the optimization itself, making accuracy repeatable across trials.
The same nonlinear activation block can be reconfigured between ReLU and its gradient, and the proposed hardware-reuse scheme cuts optical inputs and nonlinear blocks by half, pointing toward deeper multi-layer networks.
Different activation families, including sigmoid variants and ReLU variants, are implementable with the same intensity-modulator and micro-ring-modulator building blocks, so the on-chip backpropagation approach transfers across architectures.

Reading between the lines

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

If the held-out accuracy is confirmed, which the paper does not explicitly show, the result suggests that on-chip photonic training could outperform any digital-model-based training for analog accelerators whose device variations are significant, because it optimizes the real transfer function rather than an idealized one.
The demonstration still uses a microcontroller for the cost function and some housekeeping; a fully optical end-to-end loop would require moving those last electronic steps into optics or into tightly co-integrated electronics, and the hardware-reuse architecture is a step toward that.
The same PIN-attenuator gradient trick could be tested in other integrated platforms, such as micro-ring-modulator-based activations, and in deeper networks where device variations compound; a natural next experiment is a three-layer task with more than 200 points and a stated train/test split.

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

Short version: this is a solid experimental demonstration with a real novelty claim, but the abstract oversells what the chip actually does. The paper deserves a serious referee, not a desk rejection, but it needs a substantive revision before the central claim can stand.

What is genuinely new: the hidden-layer ReLU and its step-function gradient are implemented on a silicon photonic chip, and the backward path through the hidden layer is computed optically. Compared with Pai et al. (linear weights only, in situ) and Spall et al. (free-space vapor-cell nonlinearity), this is the first integrated demonstration that puts both the nonlinear activation and its gradient on-chip. The loss curves and output separation in Fig. 4 give direct evidence that the gradient updates are actually training the network, and the robustness comparison in Fig. 4f is a fair and useful experiment. That part is credible and worth building on.

Now the soft spots, in order of weight. First, the abstract says "All linear and nonlinear computations are performed on a single photonic chip," but Section 3 and Supplementary Note 2 are explicit that the output ReLU and the cost function are implemented in the microcontroller. That is not a minor wording issue: Eq. (3) defines the output-layer error delta^(2) as (dC/da^(2)) sigma'(z^(2)), so the error that starts backpropagation depends on digital arithmetic. The demonstrated system is best described as opto-electronic hybrid training with an on-chip hidden layer, not full end-to-end on-chip backpropagation. The stress-test note has this right, and it undercuts the dominant novelty claim.

Second, calling a 2-8-1 network "deep" is a stretch. The method might scale, but the demonstrated architecture is a single hidden layer.

Third, the reported 92.5% accuracy is the highest across trials, and the paper never states whether the 200 evaluated points exclude the 50 training points. As written, the inference accuracy could be a training-fit score. This is a real problem for the generalization claim, though secondary to the end-to-end issue.

Fourth, the step-function gradient proxy is used without quantifying how far it deviates from the true ReLU derivative. That is a minor concern; the loss curves show it works in practice, but the approximation error should be measured or bounded.

The underlying hardware result is solid, and the issues are addressable with better reporting and a more honest title/abstract. This is not a takedown; it is a good paper that overclaims around the edges. Send it to peer review, and ask the authors to fix the end-to-end claim, add train/test statistics, and quantify the gradient approximation.

Referee Report

3 major / 5 minor

Summary. The paper reports a silicon photonic chip that implements forward and backward paths for neural-network training, with optical linear weights, optoelectronic ReLU-like activations, and on-chip approximations of the activation gradient. The authors demonstrate gradient-descent backpropagation training on the chip for two nonlinear classification tasks (XOR and 2D point separation), reporting inference accuracy close to that of an ideal digital model and improved robustness against fabrication-induced variations compared with digital training followed by on-chip inference. The central claim is that this is the first integrated photonic deep neural network with full end-to-end on-chip backpropagation training, in which all linear and nonlinear computations are performed on a single chip.

Significance. If the central claim is established, this would be a meaningful advance for integrated photonic neural networks: it replaces offline digital backpropagation or gradient-free training with a physical implementation of the gradient computation, directly addressing device-to-device fabrication variations. The paper's strengths include direct experimental measurement of loss convergence and output separation (Fig. 4a, b, d), characterization of on-chip device variations (Supplementary Note 5, Supplementary Fig. 4), and a head-to-head comparison of on-chip training, digital training plus on-chip inference, and fully digital inference (Fig. 4f). These are concrete, falsifiable demonstrations rather than simulations. However, the scope of the claimed 'end-to-end on-chip' training needs to be re-evaluated in light of the microcontroller-implemented output nonlinearity and cost function, and the reported accuracy needs a clearer evaluation protocol. With those clarifications, the work would still be a solid experimental contribution, though the novelty claim would need to be restated more precisely.

major comments (3)

[Abstract; Section 3; Supplementary Note 2] The abstract states that 'all linear and nonlinear computations are performed on a single photonic chip,' but Section 3 explicitly says that the output-layer ReLU activation σ(2) and the cost function C(ŷ,a(2)) 'are implemented in the microcontroller unit,' and Supplementary Note 2 says that the output weighted sum W(2)a(1) is stored in microcontroller memory before the output nonlinearity and cost are applied. This is not a wording issue: Equation (3), δ(2) = (∂C/∂a(2))σ'(z(2)), is the starting point of the backward pass, so the output-layer error that seeds backpropagation depends on arithmetic performed in the microcontroller. The demonstrated system is therefore a hybrid opto-electronic training scheme with an on-chip hidden layer and a digital output layer/cost segment, not full end-to-end on-chip backpropagation. The authors should either move the output nonlinearity and cost computation onto the photonic chip, or revise the abstract, introduction, and summary claims to describe the system as hybrid on-chip/digital backpropagation with a precise statement of which computations occur where. This distinction is load-bearing because it determines whether the claimed 'first end-to-end on-chip BP training' is accurate.
[Section 4, Fig. 4c-e] The reported 'highest inference accuracy of 92.5%' is not supported by a clear evaluation protocol. The dataset consists of 200 randomly generated points, of which 50 are randomly selected for training, but the manuscript does not state whether the 200 points used for the reported accuracy include the training points. If the accuracy is computed on the full 200-point set including training points, the number is a training-fit score, not an inference accuracy, and the claimed parity with the ideal digital model is not established. Additionally, reporting the highest accuracy over trials rather than the mean and standard deviation makes the result sensitive to the best run. Please specify the train/test split, report mean and standard deviation across trials for both accuracy and loss, and ensure that the digital-model comparison in Fig. 4f uses the identical evaluation protocol.
[Fig. 4f; Supplementary Note 5] The robustness claim rests on comparing five trials of on-chip training with digital training and with the 'reference case' of digital training and inference, but the figure reports only point values with no error bars or statistical measures. Given that on-chip device variations are large (Supplementary Note 5 reports a MAC error standard deviation of 0.3662 and weight variations up to 0.273), the reader needs to know whether the observed difference between the grey curve and the blue curve is statistically meaningful across more than five weight initializations. Please provide per-trial data, mean ± standard deviation, and ideally the number of independent trials used for each curve.

minor comments (5)

[Section 2, Eq. (5)] Equation (5) writes the weight update as w(l) ← w(l) − η δ(l)a(l−1), which is dimensionally inconsistent for vector-valued layers; the standard update is w(l) ← w(l) − η δ(l)(a(l−1))^T (or the equivalent outer product in the chosen convention). Please correct this to avoid confusion about what is actually computed on chip.
[Section 2, Eq. (4)] Equation (4) as printed, δ(l) = ((w(l+1))^T δ(l+1) ⊙ σ′(l), has a missing closing parenthesis and the argument of σ′ is omitted; it should read δ(l) = ((w(l+1))^T δ(l+1)) ⊙ σ′(l)(z(l)).
[Section 3, last paragraph of Fig. 3a description] The text refers to 'each output of the hidden layer is connected to a linear weight unit of the output layer (w(1) : 1×8)'; this should be w(2) for consistency with the notation in Section 2.
[Section 2, last paragraph] There is a typo: 'using the the proposed architectures' should read 'using the proposed architectures.' There is also a duplicate reference [13] and [14] for the same Goldberg primer.
[Section 6, Data availability] The statement that data are 'available from the corresponding author upon reasonable request' is weaker than the reproducibility standard expected for a central experimental claim; please deposit the measured training-loss curves, accuracy data, and device-variation statistics in a public repository.

Circularity Check

2 steps flagged · score 4.0 of 10

Definitionally broadened 'all-on-chip' claim and ambiguous training-inclusive accuracy; no full equation-level circularity.

self definitional [Abstract vs. Section 3 and Supplementary Note 2]
"All linear and nonlinear computations are performed on a single photonic chip... The output, after passing through another ReLU activation σ(2), is used to calculate the cost function C(ŷ,a(2)). Note that σ(2) and C(ŷ,a(2)) are implemented in the microcontroller unit."

The headline claim is that the photonic chip performs full end-to-end on-chip backpropagation, but the output-layer nonlinearity and cost function, which seed the backward pass through Eq. (3), are assigned to the microcontroller. Supplementary Note 2 also stores the output weighted sum in microcontroller memory and applies the output nonlinearity and cost there. Thus the error δ(2) = (∂C/∂a(2))σ′(z(2)) is not produced entirely by the photonic chip; the claimed 'all computations on a single photonic chip' holds only if 'on-chip' is redefined to include the MCU. This is a definitional broadening of the central claim rather than an equation-to-equation reduction, but it is load-bearing for the paper's novelty.
fitted input called prediction [Section 4; Fig. 4c-d]
"The dataset consists of 200 randomly generated (x1,x2) points... To train the network, 50 points are randomly selected and the on-chip BP training is run for 40 epochs. Figure 4d shows the training loss and accuracy... Highest inference accuracy of 92.5% (points shown in Fig. 4c) is achieved."

The text never states that the 92.5% accuracy is evaluated on the 150 held-out points; the figure shows all 200 dataset points. If the 50 training points are included in the 200 evaluated points, then the reported 'inference' accuracy is partly the fitted training accuracy, and the claimed parity with the ideal digital model is not an out-of-sample result. Because no train/test split is specified, the central quantitative claim cannot be distinguished from a training-fit score. This is a partial fitted-input-called-prediction issue rather than a fully demonstrated forced reduction.

full rationale

This is primarily an experimental demonstration, not a formal derivation, and the standard backpropagation equations (1)-(5) are not themselves circular. No load-bearing self-citation chain is present: reference [47] supports only a supplementary MRM variant and is not needed for the main IM-based chip results. The two flagged issues are: (i) the abstract's all-computations-on-chip claim is internally contradicted by Section 3 and Supplementary Note 2, which place the output ReLU and cost function in the microcontroller; and (ii) the reported 92.5% inference accuracy is ambiguous because the paper does not state that evaluation is restricted to held-out points, so the number may include the 50 training points. Both issues undercut the headline claim, but the underlying measurements of optical weights, hidden-layer ReLU, and its gradient remain independent experimental evidence, so the paper is partially circular rather than fully reducible to its inputs.

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

The central claim rests on standard backpropagation theory and on experimental assumptions about the PIN attenuator's response and the gradient proxy. No free parameters are fitted to the benchmark accuracy beyond the training hyperparameters and manual bias settings. The ledger shows that the paper adds a physical implementation of known algorithms rather than postulating new physical entities.

free parameters (3)

learning_rate = 0.05 (fixed across all epochs).
Fixed learning rate for on-chip weight updates (Supplementary Note 2); no tuning procedure or sensitivity analysis is reported.
training_epochs = 40 epochs, with stopping when MSE stabilizes.
Training is run for 40 epochs with a stopping rule based on visual MSE stabilization; this hand-set hyperparameter affects the demonstrated convergence.
activation_bias_voltages_and_gain = V_forward, V_backward, and amplifier gain are set by hand.
Operating points for the PIN-attenuator ReLU and gradient circuits are adjusted manually to approximate the ideal curves (Fig. 3g-h, Supplementary Fig. 1).

assumptions (4)

standard math The standard backpropagation update rules (Eqs. 1-5) are valid for the physical network.
Invoked in Section 2 and Supplementary Note 2; assumes the measured forward and backward transformations behave like differentiable layers with the approximate on-chip gradient.
domain assumption The PIN attenuator transmission is monotonic and repeatable enough to implement weights and ReLU-like activations.
Figures 3e-h and Supplementary Fig. 4 show measured monotonic responses with device-to-device variation; the training compensates for this variation (Fig. 4f).
ad hoc to paper A high-gain step function is an acceptable surrogate for the derivative of the ReLU-like activation in backpropagation.
Supplementary Note 2 uses the switched-gain PIN attenuator as sigma-prime; the error introduced by this proxy is not quantified.
ad hoc to paper Implementing the output-layer activation and cost function in the microcontroller does not violate the 'end-to-end on-chip' claim.
Section 3 states sigma(2) and C are in the microcontroller, which conflicts with the abstract's claim that all nonlinear computations are on the photonic chip; this boundary is an assumption, not a demonstrated fact.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Integrated photonic deep neural network with end-to-end on-chip backpropagation training." pith.science (2026). https://pith.science/paper/YDUE7VZ4

@misc{pith2026250614575,
  author       = {Pith},
  title        = {Pith review of: Integrated photonic deep neural network with end-to-end on-chip backpropagation training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDUE7VZ4}},
  note         = {Machine review of arXiv:2506.14575}
}

read the original abstract

Integrated photonic neural networks (PNNs) have demonstrated significant potential to complement the digital electronic counterparts [1-3]. Nevertheless, robust and repeatable performance of scalable integrated PNNs is directly tied to the quality of their training. Error backpropagation (BP), which relies on nonlinear activation gradient computation, is the mainstream algorithm to train digital neural networks due to its scalability, versatility, and implementation efficiency [4]. Consequently, its adoption is highly desirable for the training of scalable PNNs. Despite such benefits and due to the lack of scalable on-chip activation gradient [5], PNNs have mostly been trained using a digital computer to run BP, which is inadequate in addressing device variations, or through gradient-free algorithms that do not fully benefit from the versatility of BP training. Here, we report the demonstration of an integrated photonic deep neural network with end-to-end on-chip gradient-descent BP training. All linear and nonlinear computations are performed on a single photonic chip, leading to scalable and robust training despite the considerable--but typical--fabrication-induced device variations. Two nonlinear data classification tasks are demonstrated in which the chip performance matches that of the ideal digital model, both in accuracy and robustness. Integrating the advantages of BP training with PNNs allows for generalization to various PNN architectures, paving the way for scalable and reliable next-generation photonic computing systems.

Figures

Figures reproduced from arXiv: 2506.14575 by the authors.

**Figure 1.** Backpropagation training of a neural network. The neural network consists of N layers. w (l) , z (l) and σ (l) are the weight matrix, the weighted-sum output, and the nonlinear activation function, respectively. The forward path generates an output by passing the input through neural layers and the output error is calculated using the cost function C(.). The backward path computes the error for each layer (δ (l) ), … view at source ↗

**Figure 2.** Photonic nonlinear activation functions and gradients using IM and MRM. a, One IM is used to implement sigmoid function and two cascaded IMs approximate the gradient function. b, The through port of an add-drop MRM that is biased at resonance is used for sigmoid and the drop port for the gradient when the ring is biased off-resonance using Vbackward. ReLU and it gradient (step function) can be implemented by switchi… view at source ↗

**Figure 3.** Silicon photonic deep neural network chip. a, Schematic of the neural network architecture implemented on the silicon chip. The forward path consists of an input layer, a hidden layer and an output layer. The backward path calculates the errors corresponding to the hidden and output layers (i.e., δ (1), δ(2)). b, Silicon photonic chip photograph showing different network layers. c, Wirebonded and packaged chip with … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Nonlinear data classification demonstration. a, Measured (blue triangles) and expected (red circles) XOR operation after on-chip BP training. The process is repeated multiple times and the cyan and yellow shaded regions show the variations in the output levels. b, MSE …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages

[1]

J.et al.Photonics for artificial intelligence and neuromorphic computing.Nature Photonics15, 102–114 (2021)

Shastri, B. J.et al.Photonics for artificial intelligence and neuromorphic computing.Nature Photonics15, 102–114 (2021)

2021
[2]

McMahon, P. L. The physics of optical computing.Nature Reviews Physics5, 717–734 (2023)

work page 2023
[3]

Huang, C.et al.Prospects and applications of photonic neural networks.Advances in Physics: X 7, 1981155 (2022)

work page 2022
[4]

& Liao, Q

Poggio, T., Banburski, A. & Liao, Q. Theoretical issues in deep networks.Proceedings of the National Academy of Sciences117, 30039–30045 (2020)

work page 2020
[5]

M., Tait, A

Buckley, S. M., Tait, A. N., McCaughan, A. N. & Shastri, B. J. Photonic online learning: a perspective.Nanophotonics12, 833–845 (2023)

work page 2023
[6]

How ai is shaping scientific discovery (2023)

National Academies of Sciences, Engineering, and Medicine. How ai is shaping scientific discovery (2023). URL . Accessed: 2025-05-08

work page 2023
[7]

Kudithipudi, D.et al.Neuromorphic computing at scale.Nature637, 801–812 (2025)

work page 2025
[8]

Maslej, N.et al.The ai index 2024 annual report. Tech. Rep., AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA (2024)

work page 2024

Show all 48 references

[9]

& Hinton, G

LeCun, Y., Bengio, Y. & Hinton, G. Deep learning.nature521, 436–444 (2015)

2015
[10]

M.Neural networks for pattern recognition(Oxford university press, 1995)

Bishop, C. M.Neural networks for pattern recognition(Oxford university press, 1995)

1995
[11]

LeCun, Y., Bengio, Y.et al.Convolutional networks for images, speech, and time series.The handbook of brain theory and neural networks3361, 1995 (1995)

1995
[12]

& Kim, E

Daniali, M. & Kim, E. Perception over time: Temporal dynamics for robust image understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5656–5665 (2023)

2023
[14]

A primer on neural network models for natural language processing.Journal of Artificial Intelligence Research57, 345–420 (2016)

Goldberg, Y. A primer on neural network models for natural language processing.Journal of Artificial Intelligence Research57, 345–420 (2016)

2016
[15]

& Manning, C

Hirschberg, J. & Manning, C. D. Advances in natural language processing.Science349, 261–266 (2015)

2015
[16]

N., Kingsbury, B

Arisoy, E., Sainath, T. N., Kingsbury, B. & Ramabhadran, B. Deep neural network language models. InProceedings of the NAACL-HLT 2012 Workshop: Will We Ever Really Replace the N-gram Model? On the Future of Language Modeling for HLT, 20–28 (2012)

2012
[17]

Ashtiani, F., Geers, A. J. & Aflatouni, F. An on-chip photonic deep neural network for image classification.Nature606, 501–506 (2022)

2022
[18]

Chen, Y.et al.All-analog photoelectronic chip for high-speed vision tasks.Nature623, 48–57 (2023)

2023
[19]

Xu, X.et al.11 tops photonic convolutional accelerator for optical neural networks.Nature589, 44–51 (2021)

2021
[20]

Nature Photonics18, 1335–1343 (2024)

Bandyopadhyay, S.et al.Single-chip photonic deep neural network with forward-only training. Nature Photonics18, 1335–1343 (2024)

2024
[21]

R.et al.Universal photonic artificial intelligence acceleration.Nature640, 368–374 (2025)

Ahmed, S. R.et al.Universal photonic artificial intelligence acceleration.Nature640, 368–374 (2025)

2025
[22]

& Fang, L

Zhou, T., Wu, W., Zhang, J., Yu, S. & Fang, L. Ultrafast dynamic machine vision with spatiotem- poral photonic computing.Science Advances9, eadg4391 (2023). 8

2023
[23]

Nature589, 52–58 (2021)

Feldmann, J.et al.Parallel convolutional processing using an integrated photonic tensor core. Nature589, 52–58 (2021)

2021
[24]

Wang, T.et al.Image sensing with multilayer nonlinear optical neural networks.Nature Photonics 17, 408–415 (2023)

2023
[25]

Mennel, L.et al.Ultrafast machine vision with 2d material neural network image sensors.Nature 579, 62–66 (2020)

2020
[26]

Nature Electronics4, 837–844 (2021)

Huang, C.et al.A silicon photonic–electronic neural network for fibre nonlinearity compensation. Nature Electronics4, 837–844 (2021)

2021
[27]

Miscuglio, M.et al.Massively parallel amplitude-only fourier neural network.Optica7, 1812–1819 (2020)

2020
[28]

& Fischer, I

Argyris, A., Bueno, J. & Fischer, I. Photonic machine learning implementation for signal recovery in optical communications.Scientific reports8, 8487 (2018)

2018
[29]

F., Shastri, B

Wang, B., De Lima, T. F., Shastri, B. J., Prucnal, P. R. & Huang, C. Multi-wavelength pho- tonic neuromorphic computing for intra and inter-channel distortion compensations in wdm optical communication systems.IEEE Journal of Selected Topics in Quantum Electronics29, 1–12 (2022)

2022
[30]

Gradient-based learning applied to document recognition.Proceedings of the IEEE86, 2278–2324 (1998)

1998
[31]

G.et al.Deep physical neural networks trained with backpropagation.Nature601, 549–555 (2022)

Wright, L. G.et al.Deep physical neural networks trained with backpropagation.Nature601, 549–555 (2022)

2022
[32]

& Lvovsky, A

Spall, J., Guo, X. & Lvovsky, A. I. Hybrid training of optical neural networks.Optica9, 803–811 (2022)

2022
[33]

Shen, Y.et al.Deep learning with coherent nanophotonic circuits.Nature photonics11, 441–446 (2017)

2017
[34]

J.et al.Silicon photonic architecture for training deep neural networks with direct feedback alignment.Optica9, 1323–1332 (2022)

Filipovich, M. J.et al.Silicon photonic architecture for training deep neural networks with direct feedback alignment.Optica9, 1323–1332 (2022)

2022
[35]

Y.et al.Monadic pavlovian associative learning in a backpropagation-free photonic network

Tan, J. Y.et al.Monadic pavlovian associative learning in a backpropagation-free photonic network. Optica9, 792–802 (2022)

2022
[36]

Xue, Z.et al.Fully forward mode training for optical neural networks.Nature632, 280–286 (2024)

2024
[37]

Pai, S.et al.Experimentally realized in situ backpropagation for deep learning in photonic neural networks.Science380, 398–404 (2023)

2023
[38]

& Lvovsky, A

Spall, J., Guo, X. & Lvovsky, A. I. Training neural networks with end-to-end optical backpropaga- tion.Advanced Photonics7, 016004–016004 (2025)

2025
[39]

A.Neural networks and deep learning, vol

Nielsen, M. A.Neural networks and deep learning, vol. 25 (Determination press San Francisco, CA, USA, 2015)

2015
[40]

Menon, A., Mehrotra, K., Mohan, C. K. & Ranka, S. Characterization of a class of sigmoid functions with applications to neural networks.Neural networks9, 819–835 (1996)

1996
[41]

& Pasiliao Jr, E

Banerjee, C., Mukherjee, T. & Pasiliao Jr, E. An empirical study on generalizations of the relu activation function. InProceedings of the 2019 ACM Southeast Conference, 164–167 (2019)

2019
[42]

Abadi, M.et al.TensorFlow: Large-scale machine learning on heterogeneous systems (2015). URL . Software available from tensorflow.org

2015
[43]

InOptical Fiber Communication Conference, T3H–3 (Optica Publishing Group, 2020)

Rakowski, M.et al.45nm cmos-silicon photonics monolithic technology (45clo) for next-generation, low power and high speed optical interconnects. InOptical Fiber Communication Conference, T3H–3 (Optica Publishing Group, 2020). 9

2020
[44]

& Aflatouni, F

Omirzakhov, K., Hao, H., Pirmoradi, A. & Aflatouni, F. Energy efficient monolithically integrated 256 gb/s optical transmitter with autonomous wavelength stabilization in 45 nm cmos soi.IEEE Journal of Solid-State Circuits(2024)

2024
[45]

Chan, D. W. U.et al.C-band 67 ghz silicon photonic microring modulator for dispersion- uncompensated 100 gbaud pam-4.Optics Letters47, 2935–2938 (2022)

2022
[46]

A.et al.60gb/s waveguide-coupled o-band gesi quantum-confined stark effect electro- absorption modulator

Srinivasan, S. A.et al.60gb/s waveguide-coupled o-band gesi quantum-confined stark effect electro- absorption modulator. InOptical Fiber Communication Conference, Tu1D–3 (Optica Publishing Group, 2021)

2021
[47]

& Idjadi, M

Ashtiani, F. & Idjadi, M. H. On-chip nonlinear activation and gradient functions for photonic backpropagation training and inference. In2023 IEEE Photonics Conference (IPC), 1–2 (2023)

2023
[48]

Kingma, D. P. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[49]

Y.et al.Review of silicon photonics technology and platform development.Journal of Lightwave Technology39, 4374–4389 (2021)

Siew, S. Y.et al.Review of silicon photonics technology and platform development.Journal of Lightwave Technology39, 4374–4389 (2021). 6 Data availability The data supporting findings of this study is available from the corresponding author upon reasonable request. 7 Author con...

2021

Pith tools

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

[1] [1]

J.et al.Photonics for artificial intelligence and neuromorphic computing.Nature Photonics15, 102–114 (2021)

Shastri, B. J.et al.Photonics for artificial intelligence and neuromorphic computing.Nature Photonics15, 102–114 (2021)

2021

[2] [2]

McMahon, P. L. The physics of optical computing.Nature Reviews Physics5, 717–734 (2023)

work page 2023

[3] [3]

Huang, C.et al.Prospects and applications of photonic neural networks.Advances in Physics: X 7, 1981155 (2022)

work page 2022

[4] [4]

& Liao, Q

Poggio, T., Banburski, A. & Liao, Q. Theoretical issues in deep networks.Proceedings of the National Academy of Sciences117, 30039–30045 (2020)

work page 2020

[5] [5]

M., Tait, A

Buckley, S. M., Tait, A. N., McCaughan, A. N. & Shastri, B. J. Photonic online learning: a perspective.Nanophotonics12, 833–845 (2023)

work page 2023

[6] [6]

How ai is shaping scientific discovery (2023)

National Academies of Sciences, Engineering, and Medicine. How ai is shaping scientific discovery (2023). URL . Accessed: 2025-05-08

work page 2023

[7] [7]

Kudithipudi, D.et al.Neuromorphic computing at scale.Nature637, 801–812 (2025)

work page 2025

[8] [8]

Maslej, N.et al.The ai index 2024 annual report. Tech. Rep., AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA (2024)

work page 2024