Faster Neural Network Training with Approximate Tensor Operations
Abstract
We propose a novel technique for faster deep neural network training which systematically applies sample-based approximation to the constituent tensor operations, i.e., matrix multiplications and convolutions. We introduce new sampling techniques, study their theoretical properties, and prove that they provide the same convergence guarantees when applied to SGD training. We apply approximate tensor operations to single and multi-node training of MLP and CNN networks on MNIST, CIFAR-10 and ImageNet datasets. We demonstrate up to 66% reduction in the amount of computations and communication, and up to 1.37x faster training time while maintaining negligible or no impact on the final test accuracy.
Cite
Text
Adelman et al. "Faster Neural Network Training with Approximate Tensor Operations." Neural Information Processing Systems, 2021.Markdown
[Adelman et al. "Faster Neural Network Training with Approximate Tensor Operations." Neural Information Processing Systems, 2021.](https://mlanthology.org/neurips/2021/adelman2021neurips-faster/)BibTeX
@inproceedings{adelman2021neurips-faster,
title = {{Faster Neural Network Training with Approximate Tensor Operations}},
author = {Adelman, Menachem and Levy, Kfir and Hakimi, Ido and Silberstein, Mark},
booktitle = {Neural Information Processing Systems},
year = {2021},
url = {https://mlanthology.org/neurips/2021/adelman2021neurips-faster/}
}