Lecture
This lecture discusses the development of an all-analog photoelectronic chip designed for high-speed vision tasks. The instructor begins by highlighting the ubiquity of vision tasks in applications such as autonomous driving, robotics, and medical diagnostics. They address the challenges posed by classical computation, particularly in terms of energy consumption and hardware requirements, which can create bottlenecks in performance. The lecture introduces the proposed ACCEL framework, which combines optical and electrical analog computing to enhance efficiency. The optical analog computing component aims to compress data and reduce the number of photodiodes and analog-to-digital converters (ADCs) needed, thereby mitigating speed and energy consumption issues. The instructor explains the system's architecture, including the use of diffractive optical components and the integration of a binary fully connected neural network for real-time training. The lecture concludes with a performance comparison, demonstrating the advantages of the hybrid system in terms of speed, accuracy, and robustness against fabrication errors and misalignment.