Light-in-the-loop: using a photonics co-processor for scalable training of neural networks

TL;DR

This paper introduces an optical co-processor that accelerates neural network training by performing error projection optically, potentially reducing training costs for large, complex models in applications like autonomous systems.

Contribution

It presents the first optical co-processor capable of speeding up neural network training through optical error projection using direct feedback alignment.

Findings

Successfully trained a handwritten digit recognition network using the optical co-processor.

Demonstrated optical error projection as a viable alternative to backpropagation.

Showed potential for scalable, energy-efficient neural network training.

Abstract

As neural networks grow larger and more complex and data-hungry, training costs are skyrocketing. Especially when lifelong learning is necessary, such as in recommender systems or self-driving cars, this might soon become unsustainable. In this study, we present the first optical co-processor able to accelerate the training phase of digitally-implemented neural networks. We rely on direct feedback alignment as an alternative to backpropagation, and perform the error projection step optically. Leveraging the optical random projections delivered by our co-processor, we demonstrate its use to train a neural network for handwritten digits recognition.