In optical computing, a pressing challenge is the efficient implementation of real-valued optical matrix-vector multiplication (MVM). While optical computing offers advantages such as high bandwidth, low latency, and energy efficiency, traditional optical matrix computing methods have been designed for complex-valued matrices, resulting in a significant redundancy of resources when dealing with real-valued matrices. This redundancy consumes extra energy and leads to an expanded chip footprint, raising concerns about space efficiency and scalability in large-scale optical neural networks (ONNs) and optimization problem solvers.
Efforts to address this issue have been made, with solutions such as a pseudo-real-value MZI mesh. The pseudo-real-value MZI mesh aimed to reduce the number of phase shifters required for real-valued matrices but introduced complexities related to coherent detection and additional reference paths, potentially introducing sources of error and layout intricacies.
In response to these challenges, a novel and simplified solution has emerged as a Real-Valued MZI Mesh for incoherent optical MVM. This innovative approach reduces the scale of phase shifters required to N^2 while maintaining an optical depth of N + 1. Instead of detecting the complex value of the output optical field, this method employs an extra port to perform optical power subtraction, yielding a real-valued output. This not only streamlines the hardware requirements but also simplifies the detection process, overcoming the limitations of previous solutions.
To assess the performance and viability of the proposed Real-Valued MZI Mesh, extensive numerical evaluations were conducted utilizing particle swarm optimization (PSO). The results of these evaluations demonstrated the mesh’s exceptional performance in benchmark tasks, highlighting its potential as an efficient solution for real-valued optical MVM in ONNs. Furthermore, error analyses revealed its resilience to fabrication errors, enhancing its reliability for practical applications.
Additionally, the study introduced a matched on-chip nonlinear activation function, further emphasizing the mesh’s suitability for large-scale ONNs. With its space efficiency, energy efficiency, scalability, and robustness to fabrication errors, the Real-Valued MZI Mesh emerges as a promising solution to all the challenges posed by real-valued optical matrix computing. As the field of optical computing continues to evolve, this innovative approach holds significant promise for the future of large-scale ONNs and combination optimization problem solvers, offering a more efficient and practical path forward.
Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 30k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.
If you like our work, you will love our newsletter..
Niharika is a Technical consulting intern at Marktechpost. She is a third year undergraduate, currently pursuing her B.Tech from Indian Institute of Technology(IIT), Kharagpur. She is a highly enthusiastic individual with a keen interest in Machine learning, Data science and AI and an avid reader of the latest developments in these fields.
Credit: Source link
