Automating multi-task learning on optical neural networks with weight sharing and physical rotation.

Area of Science:

• Artificial Intelligence
• Optical Computing
• Machine Learning

Background:

• The rise of AI democratization necessitates efficient multi-task learning (MTL) solutions.
• Diffractive Optical Neural Networks (DONNs) offer low energy consumption and high speed for MTL.
• Current DONN implementations for MTL are hindered by manual reconfiguration and system duplication.

Purpose of the Study:

• To introduce LUMEN-PRO, an automated framework for MTL using DONNs.
• To overcome the limitations of manual layer reconfiguration and physical system duplication in DONN-based MTL.
• To enhance the energy efficiency, accuracy, and memory footprint of MTL on DONNs.

Main Methods:

• Developed an automated MTL framework (LUMEN-PRO) for arbitrary backbone DONNs.
• Implemented task automation using a set of tasks on a single DONN architecture.
• Leveraged physical optical system rotability to replace task-specific layers with rotated shared layers, optimizing memory.

Main Results:

• LUMEN-PRO achieved a high-accuracy multi-task DONN model with a significantly smaller memory footprint than existing MTL methods.
• The framework demonstrated up to 49.58% higher accuracy and 4x better cost efficiency compared to single-task and existing DONN approaches.
• Achieved the memory lower bound for MTL, matching single-task model memory efficiency and surpassing other systems in per-operator efficiency.

Conclusions:

• LUMEN-PRO provides a flexible and automated solution for energy-efficient MTL on DONNs.
• The framework significantly reduces memory requirements and enhances computational efficiency for optical AI systems.
• LUMEN-PRO represents a substantial advancement in realizing practical and efficient MTL with DONNs.