Ultra-compact multi-task processor based on in-memory optical computing.

Area of Science:

• Neuromorphic Engineering
• Optical Computing
• Artificial Intelligence

Background:

• On-chip optical neural networks offer high parameter transduction and passive computing but face scalability and multitasking limitations.
• Transfer learning principles are explored to embed most parameters into fixed optical components and fewer into adjustable electrical components.

Purpose of the Study:

• To introduce a novel network architecture for multi-task processing using in-memory optical computing.
• To enhance computational density and energy efficiency in on-chip neuromorphic hardware.

Main Methods:

• Fabrication of two ultra-compact, in-memory, diffraction-based chips with over 60,000 parameters/mm².
• Implementation of a deep neural network model and a hard parameter sharing algorithm.
• Utilization of a deep regression algorithm for modeling physical propagation processes.

Main Results:

• The fabricated chips successfully performed multifaceted classification and regression tasks.
• Achieved accuracies comparable to electrical networks.
• Reduced power-intensive digital computation by 90%.

Conclusions:

• The developed in-memory optical computing framework shows strong potential for next-generation AI platforms.
• This approach significantly enhances energy efficiency and computational density for on-chip neuromorphic hardware.