Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks
Abstract
Deep neural networks are commonly developed and trained in 32-bit floating point format. Significant gains in performance and energy efficiency could be realized by training and inference in numerical formats optimized for deep learning. Despite advances in limited precision inference in recent years, training of neural networks in low bit-width remains a challenging problem. Here we present the Flexpoint data format, aiming at a complete replacement of 32-bit floating point format training and inference, designed to support modern deep network topologies without modifications. Flexpoint tensors have a shared exponent that is dynamically adjusted to minimize overflows and maximize available dynamic range. We validate Flexpoint by training AlexNet, a deep residual network and a generative adversarial network, using a simulator implemented with the neon deep learning framework. We demonstrate that 16-bit Flexpoint closely matches 32-bit floating point in training all three models, without any need for tuning of model hyperparameters. Our results suggest Flexpoint as a promising numerical format for future hardware for training and inference.
Keywords
Cite
@article{arxiv.1711.02213,
title = {Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks},
author = {Urs Köster and Tristan J. Webb and Xin Wang and Marcel Nassar and Arjun K. Bansal and William H. Constable and Oğuz H. Elibol and Scott Gray and Stewart Hall and Luke Hornof and Amir Khosrowshahi and Carey Kloss and Ruby J. Pai and Naveen Rao},
journal= {arXiv preprint arXiv:1711.02213},
year = {2017}
}
Comments
14 pages, 5 figures, accepted in Neural Information Processing Systems 2017