ICML 2020poster4 citations
Low-loss connection of weight vectors: distribution-based approaches
Abstract
Recent research shows that sublevel sets of the loss surfaces of overparameterized networks are connected, exactly or approximately. We describe and compare experimentally a panel of methods used to connect two low-loss points by a low-loss curve on this surface. Our methods vary in accuracy and complexity. Most of our methods are based on ”macroscopic” distributional assumptions and are insensitive to the detailed properties of the points to be connected. Some methods require a prior training of a ”global connection model” which can then be applied to any pair of points. The accuracy of the method generally correlates with its complexity and sensitivity to the endpoint detail.
BibTeX
@InProceedings{pmlr-v119-anokhin20a,
title = {Low-loss connection of weight vectors: distribution-based approaches},
author = {Anokhin, Ivan and Yarotsky, Dmitry},
booktitle = {Proceedings of the 37th International Conference on Machine Learning},
pages = {335--344},
year = {2020},
editor = {III, Hal Daumé and Singh, Aarti},
volume = {119},
series = {Proceedings of Machine Learning Research},
month = {13--18 Jul},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v119/anokhin20a/anokhin20a.pdf},
url = {https://proceedings.mlr.press/v119/anokhin20a.html},
abstract = {Recent research shows that sublevel sets of the loss surfaces of overparameterized networks are connected, exactly or approximately. We describe and compare experimentally a panel of methods used to connect two low-loss points by a low-loss curve on this surface. Our methods vary in accuracy and complexity. Most of our methods are based on ”macroscopic” distributional assumptions and are insensitive to the detailed properties of the points to be connected. Some methods require a prior training of a ”global connection model” which can then be applied to any pair of points. The accuracy of the method generally correlates with its complexity and sensitivity to the endpoint detail.}
}