For decades, engineers have dreamed of using light instead of electrons to perform the matrix multiplications at the heart of neural network inference. Photons can carry information with virtually no ohmic heating, and optical systems naturally excel at the linear operations that dominate AI workloads. Yet training such networks has remained stubbornly reliant on digital simulations. The reason is backpropagation: that elegant reverse-mode differentiation algorithm requires perfect knowledge of every operation in the forward pass, a precise digital model of the system, and the ability to propagate error signals backward through the exact same physical path. In an analog optical processor—where light scatters, interferes, and drifts with temperature—this is nearly impossible to implement in situ.

The new work borrows inspiration from the most sophisticated learning machine known: the human brain. Biological neural circuits do not solve credit assignment by shipping error gradients backward along the same axons that carried forward signals. Instead, they rely on local plasticity rules, where synaptic weights adjust based only on the activity of the neurons they directly connect. The team, led by researchers at Tsinghua University, translated this principle into the language of photonics by devising a physical learning rule that operates entirely locally within an optical chip. Their method uses only the forward-propagating light fields to compute weight updates, needing no digital twin and no reverse pass.

At the core of the system is a cascaded array of programmable Mach-Zehnder interferometers, a common building block for integrated photonic processors. During training, the network is fed pairs of augmented data streams: a “positive” sample that carries the correct label implicitly mixed into the optical signal, and a “negative” sample where the label is scrambled. A simple optical power measurement at the output of each layer produces an analog goodness signal. The local update rule then nudges each phase shifter—the photonic equivalent of a synaptic weight—in a direction that increases the difference in goodness between the positive and negative passes. This procedure, inspired by the forward-forward algorithm recently proposed by Geoffrey Hinton, requires no computation of gradients and no error backflow, making it perfectly matched to the physics of light.

The implications for energy efficiency are staggering. The team demonstrated that their optical chip can be trained on a canonical image classification task while consuming less than one-thousandth of the electrical power that would be required to simulate the same training process on a digital computer. Because the physical system itself acts as the optimizer, the entire power budget collapses to the energy of the laser source and the low-power actuators tuning the phase shifters. There is no need to shuttle enormous weight matrices between memory and a power-hungry GPU; the learning is physically embedded in the propagation of photons through silicon.

Crucially, the training method is tolerant to the manufacturing imperfections and environmental noise that have plagued previous optical AI systems. Since the local update rule is derived directly from measured optical intensities, any static fabrication errors or slow thermal drifts are automatically compensated during learning. The network essentially calibrates itself while it trains, a property that the authors demonstrate by showing that their chip maintains accuracy even when deliberately detuned from its initial operating point. This robustness could finally make wafer-scale optical neural networks manufacturable without requiring perfect lithography.

The researchers also showed that their approach scales gracefully. By cascading multiple layers of interferometers, they built a three-layer optical network with over two thousand tunable parameters and trained it to recognize handwritten digits with accuracy approaching that of digital networks of the same size. The training dynamics exhibited the same kind of layerwise learning that makes deep networks so effective, with lower layers automatically learning edge-detector-like features and higher layers assembling those into class-specific patterns. That this emergent hierarchical representation arises from a purely local, non-backpropagating rule is a profoundly interesting result for both physics and neuroscience.

While the current demonstration still uses a digital computer to apply the simple update rule after reading the optical goodness signals, the path to an all-optical closed-loop system is clear. The next step will be to integrate on-chip photodetectors and analog feedback circuits so that the weight adjustments themselves become part of the optical domain. Such a self-contained learning system could operate entirely at the speed of light, opening the door to real-time adaptive processors for applications ranging from lidar signal interpretation to ultrafast image recognition in autonomous vehicles. The marriage of bio-inspired learning algorithms and photonic hardware may well become the blueprint for sustainable AI.

Subject of Research: Bio-inspired backpropagation-free training for optical neural networks using local learning rules in integrated photonic chips.

Article Title: Bio-inspired backpropagation-free training for optical neural networks

Article References:

Li, T., Dong, Y., Tu, K. et al. Bio-inspired backpropagation-free training for optical neural networks. Light Sci Appl 15, 305 (2026). https://doi.org/10.1038/s41377-026-02394-3

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s41377-026-02394-3

Keywords: optical neural networks, backpropagation-free learning, bio-inspired computing, neuromorphic photonics, energy-efficient AI, forward-forward algorithm, integrated photonics