As artificial intelligence rapidly scales up and electronic architectures become increasingly constrained by the von Neumann bottleneck, photonic computing is emerging as a transformative solution, offering unprecedented computing speed and parallelism. While most photonic computing platforms are designed as accelerators and rely on costly electro-optic conversions for data input, weight loading and storage, recent developments in photonic in-memory computing offer the potential for brain-like co-located storage and processing, which can effectively avoid the speed and energy limitations imposed by frequent data movement. However, large-scale photonic in-memory computing remains a formidable challenge .
In a new paper published in eLight , a team of scientists, led by Professor Yan-Qing Lu, Professor Peng Chen from Nanjing University, China and Professor Min Gu, Professor Xinyuan Fang from University of Shanghai for Science and Technology, China, have presented a large-scale photonic in-memory computing platform based on liquid-crystal (LC) Poincaré-sphere-connected diffractive neural networks (DNNs), which directly tackles two long-standing bottlenecks in photonic computing : 1) costly electro-optic conversions and data movement; 2) mismatch between limited photonic resources and large computational workloads.
Through light-driven reorientation of LC molecules, the nonvolatile photonic memory, defined by the machine-learning-designed polarization conversion trajectories on the Poincaré spheres, can be synchronously rewritten, enabling massively programmable interconnections among vectorial neurons. This architecture realizes in-situ optical sensing, memory, and computing of vectorial light with minimal data-movement overheads and allows flexible switching between various machine vision tasks with ultralow static power. The scientists further expand it into a 9.44-million-neuron large-scale processor array to support massively parallel computing, single-shot recording, and light-speed retrieval of 576 polarization encoding channels, achieving a 100,000-fold improvement in memory capacity over state-of-the-art photonic in-memory computing. This new platform offers a novel pathway toward scalable photonic memory and establishes a new paradigm for physical neural networks and brain-like artificial intelligence.
These scientists summarize the operational principle and potential impact of this new platform:
“Using photo-rewritable LCs for nonvolatile photonic memory, we propose a new photonic in-memory computing platform based on Poincaré-sphere-connected vectorial DNNs. For the input and output layers, the polarization of light at each position is described as a point on the Poincaré sphere. For the hidden layers, the vectorial neurons formed by space-variant polarization converters are described as numerous trajectories on the Poincaré spheres, with their start and end points indicating the polarization states before and after conversion. Through free-space diffraction of the vectorial light field, the Poincaré spheres in adjacent layers are interconnected across amplitude, phase, and polarization dimensions, thereby linking the input to the output vectorial light.”
“Through light-driven reorientation of LC molecules, the polarization conversion trajectories on the interconnected Poincaré spheres can be programmed in a nonvolatile manner. With large-scale reconfigurable Poincaré spheres, the LC vectorial DNN enables not only higher capacity and adaptability but also specialized functionalities such as polarization perception. Conventionally, sensing, computing, and recording of polarized objects are usually performed in separate electronic modules. In contrast, the vectorial DNN can sense and distinguish them directly in the optical domain, and the computing results can be recorded in additional programmable Poincaré spheres, which holds great promise for real-time all-optical processing of ubiquitous light polarization,” the scientists added.
“The proposed platform substantially enhances the scalability, adaptability, and dimensionality of optical computing. It shows strong potential for applications in machine vision, adaptive optics, multimodal sensing, and quantum informatics, thereby establishing a new framework for intelligent photonics and brain-inspired artificial intelligence,” the scientists reported.
eLight
Light-driven molecular reorientation for large-scale photonic in-memory computing