Channel response-aware photonic neural network accelerators for high-speed inference through bandwidth-limited optics

Area of Science:

• Photonics
• Artificial Intelligence
• Computer Engineering

Background:

• Photonic neural network accelerators (PNNAs) offer low-energy solutions for deep learning (DL) demands.
• Integrating high-speed photonic circuits into analogue neuromorphic computing requires specialized DL training methods.

Purpose of the Study:

• To develop a novel channel response-aware (CRA) deep learning architecture.
• To address implementation challenges of high-speed compute rates on bandwidth-limited photonic devices.

Main Methods:

• Incorporated device frequency response into the DL training procedure.
• Validated the CRA architecture via software simulations and experimental implementation.
• Implemented the output layer of a neural network on an integrated SiPho coherent linear neuron (COLN) for MNIST image classification.

Main Results:

• The CRA model achieved experimental accuracies of 98.5% (20 GMAC/sec/axon), 97.3% (25 GMAC/sec/axon), and 92.1% (32 GMAC/sec/axon).
• The CRA model outperformed the baseline model by 7.9%, 12.3%, and 15.6% at respective compute rates.
• Demonstrated effective classification of MNIST dataset images.

Conclusions:

• The channel response-aware deep learning architecture effectively addresses bandwidth limitations in photonic neural network accelerators.
• This approach enables higher accuracy for high-speed deep learning tasks on photonic hardware.
• The findings pave the way for more efficient and powerful photonic computing systems.