Hardware Beyond Backpropagation: a Photonic Co-Processor for Direct Feedback Alignment
Abstract
The scaling hypothesis motivates the expansion of models past trillions of parameters as a path towards better performance. Recent significant developments, such as GPT-3, have been driven by this conjecture. However, as models scale-up, training them efficiently with backpropagation becomes difficult. Because model, pipeline, and data parallelism distribute parameters and gradients over compute nodes, communication is challenging to orchestrate: this is a bottleneck to further scaling. In this work, we argue that alternative training methods can mitigate these issues, and can inform the design of extreme-scale training hardware. Indeed, using a synaptically asymmetric method with a parallelizable backward pass, such as Direct Feedback Alignement, communication needs are drastically reduced. We present a photonic accelerator for Direct Feedback Alignment, able to compute random projections with trillions of parameters. We demonstrate our system on benchmark tasks, using both fully-connected and graph convolutional networks. Our hardware is the first architecture-agnostic photonic co-processor for training neural networks. This is a significant step towards building scalable hardware, able to go beyond backpropagation, and opening new avenues for deep learning.
Cite
@article{arxiv.2012.06373,
title = {Hardware Beyond Backpropagation: a Photonic Co-Processor for Direct Feedback Alignment},
author = {Julien Launay and Iacopo Poli and Kilian Müller and Gustave Pariente and Igor Carron and Laurent Daudet and Florent Krzakala and Sylvain Gigan},
journal= {arXiv preprint arXiv:2012.06373},
year = {2020}
}
Comments
6 pages, 2 figures, 1 table. Oral at the Beyond Backpropagation Workshop, NeurIPS 2020