Optical multi-task learning using multi-wavelength diffractive deep neural networks.

Area of Science:

• Neuromorphic photonic computing
• Optical artificial intelligence (AI)
• Diffractive deep neural networks (D2NNs)

Background:

• Existing photonic neural networks are limited to single tasks, hindering parallel processing.
• Task competition in current architectures degrades performance when attempting multitask learning.
• There is a need for integrated optical systems capable of handling multiple AI tasks simultaneously.

Purpose of the Study:

• To propose a novel optical multitask learning system using multiwavelength D2NNs.
• To demonstrate parallel processing of multiple AI tasks within a single monolithic system.
• To enhance computing throughput and accuracy in photonic AI.

Main Methods:

• Designed multiwavelength diffractive deep neural networks (D2NNs) with joint optimization.
• Encoded multitask inputs into distinct multiwavelength channels.
• Utilized datasets including MNIST, FMNIST, KMNIST, and EMNIST for classification tasks.

Main Results:

• Multiwavelength D2NNs significantly outperform single-wavelength D2NNs in multitask accuracy for the same network size.
• The proposed system effectively alleviates task competition, enabling high-accuracy parallel processing.
• Larger multiwavelength D2NNs achieve accuracy comparable to individually trained single-wavelength networks.

Conclusions:

• Multiwavelength D2NNs offer a viable solution for high-throughput optical multitask learning.
• This approach paves the way for wavelength-division multiplexing in neuromorphic photonic computing.
• The developed system advances the development of general AI systems capable of parallel task execution.