PACVB-package:
Variational Bayes (VB) Approximation of Gibbs Posteriors with Hinge Losses
Description
The package computes a Gaussian approximation of a Gibbs posterior for convexified risks.
We use a hinge loss, and a hinge variation of the AUC loss. The implementation follows the lines of Alquier et al. [2015]. The authors obtain PAC-Bayesian bounds for the variational approximation of an exponential weighted average estimator. The optimal bound is obtain through a gradient descent on a convex objective.
Details
|
Package: |
| PACVB |
|
Type: |
| Package |
|
Version: |
| 1.0 |
|
Date: |
| 2015-07-29 |
|
License: |
| GPL (>= 2) |
References
Alquier, P., Ridgway, J., and Chopin, N. On the properties of variational approximations of Gibbs posteriors. arXiv preprint, 2015