Name
Abstract | ||
|---|---|---|
It is known that the Frank-Wolfe (FW) algorithm, which is affine covariant, enjoys faster convergence rates than O (1/K) when the constraint set is strongly convex. However, these results rely on norm-dependent assumptions, usually incurring non-affine invariant bounds, in contradiction with FW's affine covariant property. In this work, we introduce new structural assumptions on the problem (such as the directional smoothness) and derive an affine invariant, norm-independent analysis of Frank-Wolfe. We show that our rates are better than any other known convergence rates of FW in this setting. Based on our analysis, we propose an affine invariant backtracking line-search. Interestingly, we show that typical backtracking line-searches using smoothness of the objective function present similar performances than its affine invariant counterpart, despite using affine dependent norms in the step size's computation. |
Year | Venue | DocType |
|---|---|---|
2021 | INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 139 | Conference |
Volume | ISSN | Citations |
139 | 2640-3498 | 0 |
PageRank | References | Authors |
0.34 | 0 | 4 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Thomas Kerdreux | 1 | 0 | 0.68 |
| Lewis Liu | 2 | 0 | 1.35 |
| Simon Lacoste-Julien | 3 | 1138 | 62.72 |
| Damien Scieur | 4 | 0 | 1.35 |