S. Kumar, M. Katidis, K. Musa, Z. Li, F. Long, C. Qu, Y.P. Huang
Stevens Institute of Technology,
United States
Keywords: Optical computing, Neural networks, pattern recognition
Summary:
Traditional computing architectures, such as the Von Neumann architecture, face significant bottlenecks in computational speed, data processing efficiency, and energy consumption. In response, optical computing platforms have garnered renewed interest as a promising alternative. These platforms leverage the inherent parallelism, massive interconnectivity, and energy efficiency of light to dramatically improve computational performance while reducing energy demands. The 2024 Nobel Prize in Physics was awarded for foundational discoveries and inventions that have advanced machine learning with artificial neural networks (ANNs). However, most studies on ANNs rely on a black-box approach that focuses on optimizing hyperparameters, with limited investigation into the underlying physical processes. In optical settings, photonic neural networks (PNNs) are particularly noteworthy for their ability to perform rapid optical matrix-vector multiplications, offering a significant advantage for computational speed and efficiency. In this study, we demonstrate a PNN in a simple experimental setup that integrates a spatial light modulator, lens-based Fourier transform, and coupling into a single-mode fiber with digital feedback. The network, comprising about 100 neurons, successfully stores and retrieves patterns, achieving performance near the critical capacity while exhibiting robustness against random phase-flipping errors. Additionally, the network's capacity can be further enhanced through higher-order nonlinearity using a dense associative network. Our results highlight the potential of PNNs for practical applications, including real-time image processing for autonomous driving, advanced AI systems with fast memory retrieval, and other scenarios demanding efficient, high-speed data processing. Ref: "Robust Pattern Retrieval in an Optical Hopfield Neural Network”, M. Katidis, K. Musa, S. Kumar, Z. Li, C. Qu, F. Long, Y. Huang, Accepted Opt. Lett. (2024)