Bio-inspired backpropagation-free training for optical neural networks

Optical neural networks(ONNs), with their potential for light-speed parallel and ultra-low-power operations, are seen as the next-generation hardware to overcome the bottlenecks of traditional computing. However, a critical hurdle to their practical application is the efficient training of large-scale networks on imperfect physical hardware. Conventional methods rely on the backpropagation algorithm, which requires a perfect reciprocal signal path within the optical system. In reality, any minor fabrication error or environmental disturbance can break this condition, leading to training failure. Moreover, this reliance forces frequent conversions between optical and electrical signals, sacrificing the intrinsic advantages of optical computation.

In a new paper published in Light: Science & Applications, a research team led by Academician Min Gu from the School of Artificial Intelligence Science and Technology, University of Shanghai for Science and Technology, China, has developed a Bio-inspired Backpropagation-Free Optical Neural Network (B-ONN), which realizes robust and scalable optical deep learning by decoupling network training from the physical constraints of photonic hardware. Based on this framework, they demonstrated a fully functional optical chip that achieves in-situ training with remarkable robustness against fabrication variations and environmental perturbations. More importantly, the chip-based B-ONN maintains high classification accuracy even under significant optical misalignment and phase noise, conditions that would otherwise cause severe failure in conventional backpropagation-based ONNs. This work establishes a framework that enables practical, scalable, and fault-tolerant optical computing. By overcoming the stringent physical symmetry requirements of conventional approaches, it makes high-speed, low-power optical intelligence viable for real-world environments.

The physical deployment of B-ONN comprises three dedicated units: the Optical Complex-Amplitude Encoding Unit (OCEU) for generating complex error fields, the Optical Local Error Unit (OLEU) for performing optical subtraction, and the Optical Parallel Gradient Unit (OPGU) for massively parallel gradient computation. This system enables the network to be trained and inferenced in the optical domain, supporting various successive cognitive tasks with a single photonic device in real time. The researchers explain its operational principle:

“We designed an in-situ, locally trainable network for three functions in one: (1) to train each layer independently, avoiding global backpropagation (2) to perform high-accuracy optical inference via diffractive propagation and (3) to maintain robustness under physical imperfections through smooth, hardware-compatible phase profiles.”

 “Since the optical inference operates at the speed of light and the training is parallelized in space, the system’s throughput and scalability are fundamentally high,” they added.

“The presented framework can be applied not only to realize high-speed, low-power optical computing in real-world environments, but also opens new opportunities for ultra-fast vision processing, next-generation optical interconnects, and smart photonic sensing” the scientists forecast. They further envision that, as optical training schemes mature, this platform will serve as a foundation for more complex all-optical intelligence, including optical Transformer networks for sequential modeling and even optical generative models capable of multi‑modal information synthesis—paving the way toward truly scalable, brain‑inspired photonic computing systems.

---

---