Optical Neural Networks Should Be Evaluated by Function, Not Only Physics
Organizations: Massachusetts Institute of Technology, Cambridge, MA, USA. · NTT Research Inc., Sunnyvale, CA, USA.
Abstract
Optical neural networks (ONNs) are attracting significant attention as alternative platforms for machine learning (ML). In an ONN, an optical module often takes over part of the computation inside a digital ML pipeline. How much of the task does the optical module itself perform, and how much does it contribute to any claimed system-level advantages, such as throughput or energy? This paper offers a simple, function-oriented evaluation methodology. Whatever its internal physics, the optical module appears to the surrounding digital pipeline as a mapping between real-valued tensors. We therefore evaluate it in three steps: (i) define its boundary in the pipeline, as real-valued tensors in and out; (ii) replace it with standard digital modules to identify those that match the optical system's performance on the same task and metric; and (iii) state system-level claims, such as speed or energy, per inference against the performance-matched digital module. We summarize steps (i) and (ii) in an Expressivity Card and demonstrate its use on published ONN systems. This helps isolate the optical module's contribution and provides a concrete digital reference for assessing its claimed system-level advantages, without the need to understand the physics. While we focus on optical systems as the most prominent example, the same approach applies to any physical analog computing platform used in ML, wherever a physical module takes over part of a digital pipeline.
Figures & tables
| Module | Params (Mod.) | Params (Total) | FID |
| Optical module | |||
| Optical (reported, hardware) | 160K * | 3.3M | 131 |
| Optical (simulated) | 160K * | 3.3M | |
| Digital Substitutes (Linear) | |||
| Linear | 1.05M | 4.2M | |
| Digital Substitutes (MLP-2) | |||
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Paper | Tasks | System-level claims | Digital baselines on the tasks | Optics isolated |
|---|---|---|---|---|
| Xu et al. 2024, Science Xu et al. [2024] | Image classification (CIFAR-10, Omniglot, etc.); generation of music and stylized images | Energy and area efficiency ( H , P ): diffractive propagation counted as dense matrices between adjacent layers; vs. an NVIDIA H100 GPU and other optical systems. Latency ( N ): inference time per image. | Whole (cited): VGG-16 and other electronic networks, by their reported accuracy on CIFAR-10. | No |
| Chen et al. 2025, Nature Chen et al. [2025a] | Image generation (MNIST, CelebA, Van Gogh-style art, etc.) | Operations ( F ): the digital encoder only, vs. all-digital generators. Operations ( S ): the decoder equated to the cost of simulating it (reply to reviewers). Energy ( D ): per image, vs. diffusion models. Latency ( N ): light transit through the decoder. | Whole : fully connected generators; diffusion models with as many parameters as the encoder; the teacher diffusion model. System’s digital part: a neural-network encoder. | No |
| Chen et al. 2025, Science Chen et al. [2025b] | Semantic image generation, style transfer, denoising, etc. | Throughput, energy efficiency ( H , P ): linear diffractive layers counted as one dense matrix from all input to all output pixels; vs. A100 specifications. Time, energy per frame ( D ): measured, vs. StyleGAN on an A100 at comparable quality. | Whole (other input): separately trained digital models, such as Stable Diffusion, StyleGAN, VGG-19 style transfer, and NeRF. | No |
| Xia et al. 2024, Nat. Photon. Xia et al. [2024] | Image classification and reconstruction; keypoint and pedestrian detection | Digital operations and parameters ( F ): the system’s digital decoder only, vs. CNNs on the same tasks. Response time ( N ): per video frame, compared with human reaction time. | Whole : CNNs on keypoint and pedestrian detection. Removed : linear regression with the same number of modes. Sim. : MLPs trained to approximate the optical mapping. System’s digital part: an MLP or a linear layer. | No |
| Yildirim et al. 2024, Nat. Photon. Yildirim et al. [2024] | Image classification (MNIST, Fashion-MNIST, Imagenette) | – | Whole : LeNet-5 and a lighter CNN, with comparable accuracy on all three tasks. System’s digital part: pooling and a linear layer. | Weaker yes |
| Wang et al. 2024, Nat. Comput. Sci. Wang et al. [2024a] | Classification of images (sign language, CIFAR-10, etc.) and graphs; regression | Throughput, energy efficiency ( S , P ): operations needed to emulate the fixed optical matrix digitally; vs. V100 GPU specifications. Digital operations ( F ): the linear readout only, vs. LeNet-5 at similar accuracy. | Whole : LeNet-5 (images); classical regressors (not neural networks). Whole (cited): graph models of other papers. Removed : a simulated linear random projection. Sim. : a simplified model of the optics. System’s digital part: a linear readout. | Weaker yes (partly) |
| Paper | Tasks | System-level claims | Digital baselines on the tasks | Optics isolated |
|---|---|---|---|---|
| Chen et al. 2023, Nature Chen et al. [2023b] | Image classification (MNIST, Fashion-MNIST, KMNIST, ImageNet classes); traffic video | Throughput, energy efficiency ( H , P ): the diffractive masks counted as dense matrices; vs. GPU specifications. The same metrics ( D , P ): counted as the operations of a LeNet of the same accuracy. Energy and latency per frame ( D ): measured, vs. fully connected networks and LeNets on an A100 GPU. | Whole : fully connected networks and LeNets on the same images (not on KMNIST). Removed : the analog electronic layer alone. System’s electronic part: a binary-weighted fully connected layer. | Weaker yes (partly) |
| Wang et al. 2023, Nat. Photon. Wang et al. [2023] | Image classification at the sensor (QuickDraw, cell organelles, etc.) | Power ( D ): the whole sensor, estimated at a projected high frame rate; vs. a high-speed camera with an FPGA running a digital network. | In place : a digital nonlinear encoder with the same 4-dimensional bottleneck and decoder (QuickDraw); a pretrained ResNet with the same decoder (other training data). Removed : a digital linear encoder; direct imaging, downsampled. | Yes (partly) |
| Peng et al. 2026, Nature Peng et al. [2026] | Object detection, segmentation, 3D reconstruction, video, in natural scenes | Parameters, operations, energy ( F ): those of the digital back end only, vs. state-of-the-art digital models, some more accurate. Energy and latency ( D ): end to end, vs. a conventional camera with a digital model. | Whole : lightweight and state-of-the-art digital vision models, some pretrained on external data (other training data). System’s digital part: a neural-network back end. | No |
| Huang et al. 2024, Sci. Adv. Huang et al. [2024] | Classification of hand-drawn figures, human actions, and blood cells | Operations ( H ): the share done by the optics, counted as those of the convolutions it implements. Latency ( F ): the digital post-processing only, vs. a digital network of the same structure and similar performance. | – | No |
| Zhou et al. 2023, Sci. Adv. Zhou et al. [2023b] | Spike-sequence classification, action recognition, motion tracking; nanosecond flashing patterns | Latency ( S ): vs. the same network run on a CPU. Latency ( P ), an operation-level comparison: vs. one matrix–vector multiplication of similar size on a CPU and a GPU, and the cycle times of FPGAs and event cameras. | – | No |
| Yan et al. 2024, Sci. Adv. Yan et al. [2024] | Depth perception; obstacle avoidance by a mobile robot | Power ( D ): measured per meta-pixel, vs. that of stereo vision and lidar. Latency , qualitative: depth read out at the speed of light, without the post-processing of stereo vision and lidar. | In place : a U-Net trained on the camera’s diffraction images (one task). Whole : commercial 3D sensors (a stereo camera, a Kinect, a lidar) on obstacle avoidance. | Yes (partly) |
| Company | Focus | Total Raised | Valuation | Latest Event |
| Tier 1: Public / Acquired | ||||
| Lightelligence 1 | Computing/Interconnect | $765M+ | $10.4B (mkt cap) | IPO (Apr 2026) |
| Celestial AI 2 | Interconnect | $520M | Acq. $5.5B | Acquired (Dec 2025) |
| Tier 2: Unicorns ( $1B valuation) | ||||
| Ayar Labs 3 | Interconnect | $870M | $3.75B | Series E (Mar 2026) |
| Lightmatter 4 | Computing/Interconnect | $850M | $4.4B | Series D (Oct 2024) |