Escaping Saddle Points with Adaptive Gradient Methods
© 2019 International Machine Learning Society (IMLS). Adaptive methods such as Adam and RMSProp are widely used in deep learning but are not well understood. In this paper, we seek a crisp, clean and precise characterization of their behavior in nonconvex settings. To this end, we first provide a no...
| Hauptverfasser: | , , , , |
|---|---|
| Weitere Verfasser: | |
| Format: | Artikel |
| Sprache: | English |
| Veröffentlicht: |
2021
|
| Online Zugang: | https://hdl.handle.net/1721.1/137532 |
| _version_ | 1826199754871144448 |
|---|---|
| author | Staib, Matthew Reddi, Sashank Kale, Satyen Kumar, Sanjiv Sra, Suvrit |
| author2 | Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science |
| author_facet | Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science Staib, Matthew Reddi, Sashank Kale, Satyen Kumar, Sanjiv Sra, Suvrit |
| author_sort | Staib, Matthew |
| collection | MIT |
| description | © 2019 International Machine Learning Society (IMLS). Adaptive methods such as Adam and RMSProp are widely used in deep learning but are not well understood. In this paper, we seek a crisp, clean and precise characterization of their behavior in nonconvex settings. To this end, we first provide a novel view of adaptive methods as preconditioned SGD, where the preconditioner is estimated in an online manner. By studying the preconditioner on its own, we elucidate its purpose: it rescales the stochastic gradient noise to be isotropic near stationary points, which helps escape saddle points. Furthermore, we show that adaptive methods can efficiently estimate the aforementioned preconditioner. By gluing together these two components, we provide the first (to our knowledge) second-order convergence result for any adaptive method. The key insight from our analysis is that, compared to SGD, adaptive methods escape saddle points faster, and can converge faster overall to second-order stationary points. |
| first_indexed | 2024-09-23T11:25:34Z |
| format | Article |
| id | mit-1721.1/137532 |
| institution | Massachusetts Institute of Technology |
| language | English |
| last_indexed | 2024-09-23T11:25:34Z |
| publishDate | 2021 |
| record_format | dspace |
| spelling | mit-1721.1/1375322022-10-01T03:31:18Z Escaping Saddle Points with Adaptive Gradient Methods Staib, Matthew Reddi, Sashank Kale, Satyen Kumar, Sanjiv Sra, Suvrit Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science © 2019 International Machine Learning Society (IMLS). Adaptive methods such as Adam and RMSProp are widely used in deep learning but are not well understood. In this paper, we seek a crisp, clean and precise characterization of their behavior in nonconvex settings. To this end, we first provide a novel view of adaptive methods as preconditioned SGD, where the preconditioner is estimated in an online manner. By studying the preconditioner on its own, we elucidate its purpose: it rescales the stochastic gradient noise to be isotropic near stationary points, which helps escape saddle points. Furthermore, we show that adaptive methods can efficiently estimate the aforementioned preconditioner. By gluing together these two components, we provide the first (to our knowledge) second-order convergence result for any adaptive method. The key insight from our analysis is that, compared to SGD, adaptive methods escape saddle points faster, and can converge faster overall to second-order stationary points. 2021-11-05T16:11:50Z 2021-11-05T16:11:50Z 2019 2021-04-12T17:37:59Z Article http://purl.org/eprint/type/ConferencePaper https://hdl.handle.net/1721.1/137532 Staib, Matthew, Reddi, Sashank, Kale, Satyen, Kumar, Sanjiv and Sra, Suvrit. 2019. "Escaping Saddle Points with Adaptive Gradient Methods." 36th International Conference on Machine Learning, ICML 2019, 2019-June. en http://proceedings.mlr.press/v97/staib19a.html 36th International Conference on Machine Learning, ICML 2019 Article is made available in accordance with the publisher's policy and may be subject to US copyright law. Please refer to the publisher's site for terms of use. application/pdf Proceedings of Machine Learning Research |
| spellingShingle | Staib, Matthew Reddi, Sashank Kale, Satyen Kumar, Sanjiv Sra, Suvrit Escaping Saddle Points with Adaptive Gradient Methods |
| title | Escaping Saddle Points with Adaptive Gradient Methods |
| title_full | Escaping Saddle Points with Adaptive Gradient Methods |
| title_fullStr | Escaping Saddle Points with Adaptive Gradient Methods |
| title_full_unstemmed | Escaping Saddle Points with Adaptive Gradient Methods |
| title_short | Escaping Saddle Points with Adaptive Gradient Methods |
| title_sort | escaping saddle points with adaptive gradient methods |
| url | https://hdl.handle.net/1721.1/137532 |
| work_keys_str_mv | AT staibmatthew escapingsaddlepointswithadaptivegradientmethods AT reddisashank escapingsaddlepointswithadaptivegradientmethods AT kalesatyen escapingsaddlepointswithadaptivegradientmethods AT kumarsanjiv escapingsaddlepointswithadaptivegradientmethods AT srasuvrit escapingsaddlepointswithadaptivegradientmethods |