Variational Approximations of Bayesian Inference
Event details
| Date | 18.11.2011 |
| Hour | 10:15 |
| Speaker | Pr. M. Seeger, Laboratory for Probabilistic Machine Learning, EPFL. |
| Location |
ME C2 405
|
| Category | Conferences - Seminars |
Bayesian decision-making is driven by queries to the posterior distribution, obtained by conditioning a probabilistic model and prior knowledge on acquired data. This process is computationally challenging in general and remains out of reach for large-scale coupled models of images or image sequences. We give an overview of recent variational techniques, with which the Bayesian computational challenge can be met approximately. Intractable posterior integrations are relaxed to variational optimization problems. Using novel decoupling techniques and double loop algorithms, these problems can be mapped to standard large scale optimization primitives such as penalized least squares estimation and Gaussian variance computations. We present results of our methodology for the problem of optimization magnetic resonance imaging acquisitions by Bayesian sequential optimal design.
Links
Practical information
- General public
- Free