Variance Reduction for Faster Non-Convex Optimization

Zeyuan Allen-Zhu, Elad Hazan
Proceedings of The 33rd International Conference on Machine Learning, PMLR 48:699-707, 2016.

Abstract

We consider the fundamental problem in non-convex optimization of efficiently reaching a stationary point. In contrast to the convex case, in the long history of this basic problem, the only known theoretical results on first-order non-convex optimization remain to be full gradient descent that converges in O(1/\varepsilon) iterations for smooth objectives, and stochastic gradient descent that converges in O(1/\varepsilon^2) iterations for objectives that are sum of smooth functions. We provide the first improvement in this line of research. Our result is based on the variance reduction trick recently introduced to convex optimization, as well as a brand new analysis of variance reduction that is suitable for non-convex optimization. For objectives that are sum of smooth functions, our first-order minibatch stochastic method converges with an O(1/\varepsilon) rate, and is faster than full gradient descent by Ω(n^1/3). We demonstrate the effectiveness of our methods on empirical risk minimizations with non-convex loss functions and training neural nets.

Cite this Paper


BibTeX
@InProceedings{pmlr-v48-allen-zhua16, title = {Variance Reduction for Faster Non-Convex Optimization}, author = {Allen-Zhu, Zeyuan and Hazan, Elad}, booktitle = {Proceedings of The 33rd International Conference on Machine Learning}, pages = {699--707}, year = {2016}, editor = {Balcan, Maria Florina and Weinberger, Kilian Q.}, volume = {48}, series = {Proceedings of Machine Learning Research}, address = {New York, New York, USA}, month = {20--22 Jun}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v48/allen-zhua16.pdf}, url = { http://proceedings.mlr.press/v48/allen-zhua16.html }, abstract = {We consider the fundamental problem in non-convex optimization of efficiently reaching a stationary point. In contrast to the convex case, in the long history of this basic problem, the only known theoretical results on first-order non-convex optimization remain to be full gradient descent that converges in O(1/\varepsilon) iterations for smooth objectives, and stochastic gradient descent that converges in O(1/\varepsilon^2) iterations for objectives that are sum of smooth functions. We provide the first improvement in this line of research. Our result is based on the variance reduction trick recently introduced to convex optimization, as well as a brand new analysis of variance reduction that is suitable for non-convex optimization. For objectives that are sum of smooth functions, our first-order minibatch stochastic method converges with an O(1/\varepsilon) rate, and is faster than full gradient descent by Ω(n^1/3). We demonstrate the effectiveness of our methods on empirical risk minimizations with non-convex loss functions and training neural nets.} }
Endnote
%0 Conference Paper %T Variance Reduction for Faster Non-Convex Optimization %A Zeyuan Allen-Zhu %A Elad Hazan %B Proceedings of The 33rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2016 %E Maria Florina Balcan %E Kilian Q. Weinberger %F pmlr-v48-allen-zhua16 %I PMLR %P 699--707 %U http://proceedings.mlr.press/v48/allen-zhua16.html %V 48 %X We consider the fundamental problem in non-convex optimization of efficiently reaching a stationary point. In contrast to the convex case, in the long history of this basic problem, the only known theoretical results on first-order non-convex optimization remain to be full gradient descent that converges in O(1/\varepsilon) iterations for smooth objectives, and stochastic gradient descent that converges in O(1/\varepsilon^2) iterations for objectives that are sum of smooth functions. We provide the first improvement in this line of research. Our result is based on the variance reduction trick recently introduced to convex optimization, as well as a brand new analysis of variance reduction that is suitable for non-convex optimization. For objectives that are sum of smooth functions, our first-order minibatch stochastic method converges with an O(1/\varepsilon) rate, and is faster than full gradient descent by Ω(n^1/3). We demonstrate the effectiveness of our methods on empirical risk minimizations with non-convex loss functions and training neural nets.
RIS
TY - CPAPER TI - Variance Reduction for Faster Non-Convex Optimization AU - Zeyuan Allen-Zhu AU - Elad Hazan BT - Proceedings of The 33rd International Conference on Machine Learning DA - 2016/06/11 ED - Maria Florina Balcan ED - Kilian Q. Weinberger ID - pmlr-v48-allen-zhua16 PB - PMLR DP - Proceedings of Machine Learning Research VL - 48 SP - 699 EP - 707 L1 - http://proceedings.mlr.press/v48/allen-zhua16.pdf UR - http://proceedings.mlr.press/v48/allen-zhua16.html AB - We consider the fundamental problem in non-convex optimization of efficiently reaching a stationary point. In contrast to the convex case, in the long history of this basic problem, the only known theoretical results on first-order non-convex optimization remain to be full gradient descent that converges in O(1/\varepsilon) iterations for smooth objectives, and stochastic gradient descent that converges in O(1/\varepsilon^2) iterations for objectives that are sum of smooth functions. We provide the first improvement in this line of research. Our result is based on the variance reduction trick recently introduced to convex optimization, as well as a brand new analysis of variance reduction that is suitable for non-convex optimization. For objectives that are sum of smooth functions, our first-order minibatch stochastic method converges with an O(1/\varepsilon) rate, and is faster than full gradient descent by Ω(n^1/3). We demonstrate the effectiveness of our methods on empirical risk minimizations with non-convex loss functions and training neural nets. ER -
APA
Allen-Zhu, Z. & Hazan, E.. (2016). Variance Reduction for Faster Non-Convex Optimization. Proceedings of The 33rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 48:699-707 Available from http://proceedings.mlr.press/v48/allen-zhua16.html .

Related Material