MLFL Wiki |
Main /
## Sparse Variational Learning For Stereo VisionVariational inference techniques are commonly used to learn probabilistic models in problem domains such as text and image processing. Like belief propagation, the variational approximation can be viewed in a message passing framework. However, when the state space of discrete variables grows very large, iteratively calculating updates from messages can be prohibitively slow. In this talk I will describe an additional variational approximation that makes messages sparse. Our sparse message passing retains similar theoretical bounds of variational techniques, while preserving an appropriate amount of uncertainty for gradient-based parameter learning. Our work is motivated by an application in depth estimation from stereo images, and we show that our technique is orders of magnitude faster than the typical variational approach, and yields more accurate models than prior work using a point estimate for parameter learning. |

Page last modified on April 15, 2008, at 03:32 PM