meProp: Sparsified Back Propagation for Accelerated Deep Learning with Reduced Overfitting
Xu Sun, Xuancheng Ren, Shuming Ma, Houfeng Wang
Abstract
We propose a simple yet effective technique for neural network learning. The forward propagation is computed as usual. In back propagation, only a small subset of the full gradient is computed to update the model parameters. The gradient vectors are sparsified in such a way that only the top-$k$ elements (in terms of magnitude) are kept. As a result, only $k$ rows or columns (depending on the layout) of the weight matrix are modified, leading to a linear reduction ($k$ divided by the vector dimension) in the computational cost. Surprisingly, experimental results demonstrate that we can update only 1–4\% of the weights at each back propagation pass. This does not result in a larger number of training iterations. More interestingly, the accuracy of the resulting models is actually improved rather than degraded, and a detailed analysis is given.
BibTeX
@InProceedings{pmlr-v70-sun17c,
title = {me{P}rop: Sparsified Back Propagation for Accelerated Deep Learning with Reduced Overfitting},
author = {Xu Sun and Xuancheng Ren and Shuming Ma and Houfeng Wang},
booktitle = {Proceedings of the 34th International Conference on Machine Learning},
pages = {3299--3308},
year = {2017},
editor = {Precup, Doina and Teh, Yee Whye},
volume = {70},
series = {Proceedings of Machine Learning Research},
month = {06--11 Aug},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v70/sun17c/sun17c.pdf},
url = {https://proceedings.mlr.press/v70/sun17c.html},
abstract = {We propose a simple yet effective technique for neural network learning. The forward propagation is computed as usual. In back propagation, only a small subset of the full gradient is computed to update the model parameters. The gradient vectors are sparsified in such a way that only the top-$k$ elements (in terms of magnitude) are kept. As a result, only $k$ rows or columns (depending on the layout) of the weight matrix are modified, leading to a linear reduction ($k$ divided by the vector dimension) in the computational cost. Surprisingly, experimental results demonstrate that we can update only 1–4\% of the weights at each back propagation pass. This does not result in a larger number of training iterations. More interestingly, the accuracy of the resulting models is actually improved rather than degraded, and a detailed analysis is given.}
}