Note: Your custom schedule will not be saved unless you create a new account or login to an existing account.
| Paper: | SPTM-P5.7 |
| Session: | Adaptive Systems and Signal Processing |
| Time: | Wednesday, May 19, 15:30 - 17:30 |
| Presentation: |
Poster |
| Topic: |
Signal Processing Theory and Methods: Detection, Estimation, and Class. Thry & Apps. |
| Title: |
KALMAN FILTERING IN STOCHASTIC GRADIENT ALGORITHMS: CONSTRUCTION OF A STOPPING RULE |
| Authors: |
Barbara Bittner; Université Nice Sophia Antipolis/CNRS | | |
| | Luc Pronzato; Université Nice Sophia Antipolis/CNRS | | |
| Abstract: |
Stochastic gradient algorithms are widely used in signal processing. Whereas stopping rules for deterministic descentalgorithms can easily be constructed, using for instance the normof the gradient of the objective function, the situation is morecomplicated for stochastic methods since the gradient needs firstto be estimated. We show how a simple Kalman filter can be used toestimate the gradient, with some associated confidence, and thusconstruct a stopping rule for the algorithm. The construction isillustrated by a simple example. The filter might also be used toestimate the Hessian, which would open the way to a possibleacceleration of the algorithm. Such developments are brieflydiscussed. |
| |
| Back | |