Backpropagation is the algorithm that built modern AI, and it is a poor fit for optics. A paper from the University of Shanghai for Science and Technology offers one of the cleanest attempts yet to resolve that tension. The team’s Bio-inspired Backpropagation-Free Optical Neural Network, or B-ONN, trains a photonic network without ever sending an error signal backward through the optical system: each layer learns locally, from a target constructed in the optical domain. Published in Light: Science & Applications in July 2026 and newly covered this week, it designs training around what physical hardware can actually do.
Why backpropagation breaks on optics
Backpropagation computes, for every weight, how much a small change would reduce the final error. The chain rule turns that into a backward pass: the output error is propagated layer by layer back toward the input, multiplied at each layer by the transpose of that layer’s weight matrix. On a digital computer this is bookkeeping. On an optical system it is a physical demand: the backward path must be the precise reciprocal of the forward path.
Physical optics resists that demand in several ways at once: fabrication errors, small misalignments, phase fluctuation with temperature and vibration, and components that are simply not reciprocal — light does not travel through them the same way in both directions. When the backward path is imperfect, the gradients are wrong, and training degrades or fails. The usual workaround is to train on a simulation and transfer the weights to hardware, but that reintroduces the simulation-to-reality gap this desk has written about before: a network trained on a perfect model of the optics meets imperfect optics and underperforms. Schemes that shuttle signals between the optical and electrical domains during training add conversion costs that erode the advantages of computing in light.
Earlier escapes each paid a price, and the paper is candid about them. Direct Feedback Alignment bypasses the reciprocal path with fixed random projections, but those projections struggle with convolutional spatial structure, and the approach still relies on digital forward inference. The Forward-Forward Algorithm’s single scalar objective per layer gives no pixel-level guidance to layers with millions of degrees of freedom. Physical Local Learning requires sequential characterization of each layer; physical-reciprocity methods match backpropagation accuracy only in spatially symmetric systems. The field has been trading one constraint for another.
How B-ONN replaces the backward path
B-ONN’s answer is to give every layer its own learning problem and solve it with optics rather than a backward pass. Each layer carries a forward optical layer, which transforms the incoming wavefront through a trainable phase mask, and an error layer built around a trainable error convolution kernel, denoted E — the core innovation. In backpropagation, the error at one layer is computed from the error at the next by multiplying through the transposed weights, the step that demands precise optical conjugation between forward and backward paths. In B-ONN, the error kernel is an independently learnable parameter matrix that adapts during training to approximate the inverse of the forward transformation. It does not need to equal the transposed weights or need a reciprocal path; it needs to approach the inverse well enough to generate useful local targets, and it learns to do so as training proceeds.
The mechanics are straightforward. After a forward pass, each layer’s error layer receives the local error signal from the next layer and uses target propagation to construct the target optical field for the current layer — the field the layer should have produced. Optical interference computes the difference between the layer’s actual output and that target, and that local error independently guides updates to both the forward parameters and the error kernel. Because each layer’s target is generated locally, the layers are decoupled: no chain of errors runs back through the network. The paper calls the result learning equivalent to global backpropagation in that it reaches comparable accuracy, not in that it computes the same gradients — the same principle as biological synapses, which adjust on locally available information.
The team built the algorithm’s operations into three dedicated optical units. The Optical Complex-Amplitude Encoding Unit (OCEU) generates the complex error fields using a single phase-only spatial light modulator (SLM), decomposing a complex field into two complementary phase-only components holographically synthesized on the SLM. The Optical Local Error Unit (OLEU) performs the optical subtraction, built as a Mach–Zehnder interferometer with a π-phase shift in the reference arm so destructive interference yields the difference directly in the optical domain. The Optical Parallel Gradient Unit (OPGU), a beam splitter plus CMOS camera at the detection port, computes gradients for all pixels in parallel: phase-shifting interferometry extracts the complex gradient from interference patterns, so one CMOS exposure captures gradient information for the whole spatial plane.
What the numbers actually show
On MNIST, B-ONN achieved 93.25 percent accuracy; on Fashion-MNIST, 82.28 percent. In a three-layer configuration, the gap to a traditional backpropagation-trained optical network (T-ONN) was less than 0.5 percentage points — effectively parity. The more striking result comes from varying depth from three to nine layers. In shallower configurations (three to six layers), the backpropagation network held a marginal lead, as expected. But a crossover occurred at seven layers, where B-ONN began to outperform T-ONN, reaching a peak of 95.10 percent at nine layers. The paper attributes this to gradient instability — the vanishing and exploding gradients that plague deep physical systems — afflicting the global method while the decoupled local method maintains stable gradients regardless of depth.
That depth result is the paper’s most consequential claim: photonic networks that matter will be deep, and if the standard training method degrades with depth on physical hardware while the local method does not, the training algorithm determines whether scaling is possible at all. It should still be read with the usual caution: MNIST and Fashion-MNIST are small, well-understood tasks, and parity on them is necessary, not sufficient. The paper reports no results on larger datasets, and readers should not extrapolate the depth crossover to transformer-scale models without evidence.
Robustness: the result that matters most
If the benchmark parity is the headline, the robustness results are the argument. The team trained both B-ONN and T-ONN under idealized, noise-free conditions, then tested them under real hardware imperfections. Under additive Gaussian phase noise — modeling fabrication surface roughness, etching-depth variation, and SLM phase quantization errors — B-ONN maintained approximately 75 percent accuracy at σ ≈ 0.4π, while the backpropagation network degraded sharply once noise passed σ > 0.2π. Under lateral misalignment — mechanical positioning errors and optical-axis drift from vibration or thermal expansion — B-ONN held accuracy above 75 percent within ±3 pixels of shift, while T-ONN fell from 90 percent to below 60 percent with only one to two pixels of shift.
The mechanism is that B-ONN learns smooth phase profiles. Where a backpropagation-trained network develops high-frequency, sharply localized structures that control light precisely under ideal conditions but shatter under perturbation, B-ONN’s phase maps are globally smooth — a built-in low-pass filter corresponding to wide, flat minima in the loss landscape rather than narrow optima. Crucially, this robustness emerges without noise-augmented training: the network was never trained on noisy examples. It is, in the paper’s phrase, architecture-inherent.
This is the result that connects most directly to deployment. A system requiring sub-pixel alignment and a thermally stabilized, vibration-isolated environment is a laboratory instrument; one tolerating three pixels of misalignment and substantial phase noise could plausibly be manufactured and shipped. The numbers do not prove deployability — they are simulations of specific noise models, not field measurements — but they shift the question from “can we build the optics precisely enough?” to “can the training method absorb the imprecision we know we will have?”
From the bench to a printed chip
The paper goes beyond simulation in two hardware demonstrations. First, a programmable SLM system: a 532 nm laser source, an amplitude-modulated SLM encoding the input, a 4f relay demagnifying the wavefront by 0.75× onto a phase-modulated SLM carrying the trained weights, and a CMOS camera. On a four-class handwritten digit task (digits 0 through 3), the network converged within five epochs, reaching 99.5 percent accuracy in simulation and 95 percent in the physical experiment. The honest caveat is the sample count: 4,157 samples were used for numerical validation, but only 20 representative samples were physically tested. The 95 percent physical figure proves the trained network runs on real optics; it is not a statistically robust benchmark of generalization.
Second, a chip-scale integrated diffractive network, fabricated by nano-printing (two-photon polymerization), with two diffractive layers. Trained on the same four-class digit subset, it reached 94 percent accuracy on the validation set after 20 epochs, with measured optical energy distributions matching simulation in routing light to the correct target quadrants. Here the fabrication story and the algorithmic story converge: two-photon polymerization rounds off sharp step edges into continuous slopes, which would degrade a network depending on sharp phase features. B-ONN’s smooth phase profiles, marked in the paper’s microscopy images as “Continuous Phase Plateau” regions, are precisely the kind of profile this process produces naturally. The algorithm anticipates the manufacturing constraint rather than fighting it — the paper’s most persuasive single idea.
The honest ledger
Several limitations deserve equal weight, because they define what this paper does not establish. First, the demonstrations are small. Four-class digit recognition on a two-layer printed chip and an SLM bench system is a long way from a useful AI workload. The paper names optical Transformer networks and optical generative models as future directions; those are aspirations, not results, and no evidence here bears on whether local target propagation scales to attention mechanisms or multimodal generation.
Second, there is no system-level energy or latency measurement. The paper describes the optical training loop as eliminating the latency and energy bottlenecks of electronic backpropagation, but it reports no measured power consumption, no inference latency, and no comparison of the full pipeline — laser, SLMs, camera, control electronics — against a digital baseline. Until such measurements exist, the energy case is a direction of travel, not a result.
Third, the training loop is not purely optical. Standard digital optimizers such as Adam are still employed to accelerate numerical convergence and stabilize temporal dynamics — a reasonable engineering choice, since the gradients themselves are computed optically and locally, which is the hard part. But “all-optical training” should be understood as optical gradient computation with digital optimization support, not as a loop with no electronics at all.
Fourth, there is a space cost, and the paper states it plainly. B-ONN incurs approximately three times the physical space overhead of a conventional approach, in exchange for an 18-fold reduction in dynamic electrical memory traffic for a nine-layer architecture. That trade — more optical hardware, less electrical memory movement — is attractive if memory traffic is the bottleneck, but it is a trade, not a free lunch. Each layer’s error unit adds physical components, so monolithic photonic integration and advanced 3D packaging will be needed to compress the cascaded modules into a practical footprint. The space overhead is also why the chip demonstration is two layers.
Finally, a note on timing and provenance. The paper was published on 6 July 2026; the trade coverage that surfaced it this week carried no accuracy numbers, datasets, or device specifications at all. Every number in this article comes from the peer-reviewed paper itself, not from the coverage. A press summary can make a result sound both more and less than it is — more, by omitting the small sample counts and missing energy data; less, by omitting the accuracy figures that make the result checkable.
Where this fits, and what to watch
B-ONN lands in a field converging on the same insight from several directions. The Koç University work covered on this desk earlier in October attacks the simulation-to-reality gap from the hardware side: learn from measurements of the actual device, through Bayesian optimization and hardware-in-the-loop protocols, so the system adapts to its own imperfections. B-ONN attacks the same gap from the algorithm side: design a training rule that does not require the device to be a faithful realization of a simulation. Both share the premise that photonic hardware will never be as precisely reproducible as digital electronics, and that training methods must treat the physical device as ground truth. They are complementary, not competing.
The MIT Julia Lab–Neurophos collaboration covered on this desk last week sits at a different layer of the same stack: the compiler and algorithm layer. B-ONN is, in effect, a contribution to that layer from the research side — a training algorithm designed for the device’s physics rather than ported from digital practice.
Three developments would indicate B-ONN is becoming more than a laboratory result: a chip-scale demonstration with more than two layers and a task beyond four-class digit recognition, ideally with the local learning loop itself running on the integrated device rather than an SLM bench; a system-level energy figure — the full pipeline, laser to camera to readout, against a digital baseline; and adoption by other groups, whether on a different photonic platform, at greater depth, or combined with hardware-in-the-loop calibration of the kind the Koç work demonstrates. B-ONN stops asking optics to imitate the backward pass and instead asks what optics can compute natively — interference, subtraction, parallel gradients — and builds the learning rule from those primitives. That is the kind of design thinking the field will need to reach systems that work outside the laboratory.
Sources and further reading
This article is based on the peer-reviewed paper: Li, T., Dong, Y., Tu, K., Zhang, Y., Li, S., Duan, Y., Sun, P., Gu, M., and Wang, J., “Bio-inspired backpropagation-free training for optical neural networks,” Light: Science & Applications 15, 305 (2026), published 6 July 2026, DOI 10.1038/s41377-026-02394-3. All authors are affiliated with the University of Shanghai for Science and Technology (School of Artificial Intelligence Science and Technology and Institute of Photonic Chips); corresponding authors Yuchao Zhang, Min Gu, and Jing Wang. Accuracy figures (93.25% MNIST, 82.28% Fashion-MNIST, <0.5% gap to T-ONN, 95.10% peak at nine layers), robustness figures (~75% at σ ≈ 0.4π phase noise, >75% within ±3 pixels misalignment, T-ONN falling below 60% at 1–2 pixels), hardware results (99.5% simulation / 95% physical on the SLM system, 94% chip-scale validation), and the space–memory trade-off (~3× space overhead, 18-fold memory-traffic reduction at nine layers) are from the paper. Trade coverage via bioengineer.org (26 October 2026) carried no numerical results and was not used as a source for any figure. For the complementary hardware-in-the-loop approach, see this desk’s coverage of the Koç University multichannel and fiber-laser systems; for the compiler and algorithm layer, see the MIT Julia Lab–Neurophos collaboration coverage. Background on diffractive deep neural networks and on backpropagation-free methods (Direct Feedback Alignment, Forward-Forward, Physical Local Learning) is as reviewed in the paper’s introduction.