Perfectly Perform Machine Learning Task with Imperfect Optical Hardware Accelerator

TL;DR

This paper introduces an optical hardware platform for machine learning that compensates for device imperfections through calibration, enabling accurate GEMM computations and successful ML task deployment.

Contribution

It presents a novel optical ML hardware with a calibration method to mitigate device imperfections, facilitating scalable and accurate GEMM operations for ML applications.

Findings

Calibration enables accurate optical GEMM despite device non-uniformity

Optical hardware achieves prediction accuracy comparable to GPUs

Characterization of optical system under various configurations

Abstract

Optical architectures have been emerging as an energy-efficient and high-throughput hardware platform to accelerate computationally intensive general matrix-matrix multiplications (GEMMs) in modern machine learning (ML) algorithms. However, the inevitable imperfection and non-uniformity in large-scale optoelectronic devices prevent the scalable deployment of optical architectures, particularly those with innovative nano-devices. Here, we report an optical ML hardware to accelerate GEMM operations based on cascaded spatial light modulators and present a calibration procedure that enables accurate calculations despite the non-uniformity and imperfection in devices and system. We further characterize the hardware calculation accuracy under different configurations of electrical-optical interfaces. Finally, we deploy the developed optical hardware and calibration procedure to perform a ML task of predicting the intersubband plasmon frequency in single-wall carbon nanotubes. The obtained prediction accuracy from the optical hardware agrees well with that obtained using a general purpose electronic graphic process unit.