Artificial Intelligence is rapidly transforming our world, but its growing complexity demands massive computational power and energy. Optical Neural Networks (ONNs), which use light instead of electricity to process data, have emerged as a promising solution, offering ultra-low latency and high energy efficiency. However, teaching light to "think" like a computer has a fundamental physical hurdle: light intensity represents brightness and is naturally non-negative.In mathematical terms, modern AI algorithms rely heavily on real numbers—both positive and negative—to distinguish correlations in data. Traditional optical methods struggle with this, often requiring electronic post-processing to handle negative numbers or restricting the AI to "non-negative" mathematics, which severely limits its intelligence and stability.

Now, a research team led by Prof. Jianji Dong at Huazhong University of Science and Technology (HUST) has broken this bottleneck. They employ two MRMs biased at different resonance wavelengths to achieve real-valued optical encoding, together with a dual-MRM activation element driven by the differential photocurrent of photodiodes, which provides optically cascadable real-valued nonlinear activation. Combined with a real-valued Mach–Zehnder interferometer mesh for matrix computation, this architecture realizes a fully real-valued end-to-end ONN.

The researchers experimentally demonstrated a tanh-like nonlinear activation function and validated it on an iris classification task, achieving an accuracy of 98%. They further modeled the generator of a generative adversarial network based on this structure. Notably, this generator can use natural optical noise as its input, effectively eliminating the need for electro-optic and digital-to-analog conversions at the input stage. Through these innovations, the proposed ONN paves the way for optical-to-optical on-chip image generation, demonstrating the distinct advantages of optical computing. The work, entitled “A fully real-valued end-to-end optical neural network for generative model,” was published in Frontiers of Optoelectronics (published on Jan. 26, 2026)
DOI：10.2738/foe.2026.0004