Photonic convolutional neural network with robustness against wavelength deviations.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Science

Background:

• Photonic convolutional neural networks (PCNNs) show promise for high-speed computation.
• Fabrication variations in integrated multiwavelength laser arrays (MLAs) can lead to imperfect wavelength spacing, impacting PCNN performance.
• The robustness of PCNNs to non-ideal wavelength spacing is not well-understood.

Purpose of the Study:

• To experimentally investigate the practicality of integrated multiwavelength laser arrays (MLAs) for PCNN applications.
• To evaluate the performance of PCNNs with non-ideal wavelength spacing.
• To assess the potential of scalable MLAs for low-cost optical computing.

Main Methods:

• Experimental and numerical investigation of PCNN performance with non-ideal wavelength spacing.
• Utilizing integrated multiwavelength laser arrays (MLAs) as the light source for PCNN.
• Benchmarking PCNN accuracy on the MNIST handwritten digit classification task.

Main Results:

• PCNNs exhibit tolerance to wavelength deviation, maintaining robust photonic recognition accuracy.
• Experimental photonic prediction accuracy of 91.2% achieved for MNIST classification using MLAs.
• The system operates at speeds on the order of tera operations per second.

Conclusions:

• Scalable MLAs are a viable alternative light source for PCNNs, supporting low-cost optical computing.
• The robust performance and capabilities of PCNNs driven by MLAs can advance photonic neural network applications.
• This research broadens the application scope of photonic neural networks in next-generation data computing.