Neuromorphic Photonic Operating Systems

点击翻译按钮获取中文摘要

The advance of modern computing is made possible through the continued exploration and integration of alternative architectures and alternative physics. Neuromorphic photonics combines photonic device physics with distributed processing architectures. By multiplexing network connections through on-chip waveguides as opposed electrical wiring and digital busses, neuromorphic photonic systems can operate on nanosecond timescales compared to the millisecond timescales found in neuromorphic electronic counterparts. Ultrafast neural networks promise to expand, not just improve, the applications of machine learning, prediction, and control, specifically for real-time problems in which the computation has a latency deadline. Ultrafast real-time computing challenges exist and are currently unsolved in cognitive radios, high-energy physics experiments, superconducting quantum computers, fusion energy reactors, and potentially other areas. Technologies that enable advances in any of these applications would immensely impact what we are capable of as a society. Recent leaps in silicon photonic manufacturing have created an unprecedented opportunity to produce large-scale, low-cost photonic processors in volume. Photonic information processing research has also accelerated considerably; however, research pipelines are silo-ed and vertical. Each research lab must become proficient in everything from laying out waveguides to understanding machine learning to performing experimental demonstrations. Research in all forms of computing, on the other hand, is sub-disciplined into hardware, operating system (OS), and software. The OS provides a standard interface that enables complex ideas in hardware and software to develop independently and work together within one machine. With an emerging potential for photonic information processors to grow in scale and complexity, standardized neuromorphic photonic operating systems (NPOS) are urgently needed. The proposed research will address these needs by creating, validating, sharing, and applying NPOSs. The key components of NPOS are a kernel, virtual machine, hardware-specific drivers, and hardware-independent programming interface. NPOS will be validated on an experimental photonic testbed that is 100% remote controlled. Remote capabilities mean that testbed access can expand smoothly from the local lab network, then to a collaborator intranet, then, ultimately, to the public internet as a cloud service. Finally, the program will leverage NPOS to explore new application areas, focusing on high-impact, real-time applications that cannot be performed by state-of-the-art electronics. Through the program, 2 PhD, 3 MSc, and 3 undergraduate students will be trained in experimental integrated photonics and neural network programming. Training will position them to become leaders in Canada's world class photonics ecosystem and rapidly growing sectors of machine learning and unconventional computing.