Date: 16 January 2009
Time: 12:00 p.m.
Place: Newell Simon Hall 1507
Type: Thesis Proposal
Who: Stanislav Funiak
Topic: Distributed Algorithms for Probabilistic Inference and Learning
Abstract:
Probabilistic inference and learning problems arise naturally in distributed systems such as sensor networks, teams of mobile robots, and recommendation systems. In these systems, the data resides at multiple distributed locations, and the network nodes need to collaborate, in order to perform the inference or learning task.
This thesis has three thrusts. First, we propose distributed implementations of several state-of-the-art centralized inference algorithms. Our solutions address challenges, such as effective MAP estimation, scheduling of messages in loopy belief propagation, and assumed density filtering.
Many algorithms for probabilistic inference are described by graphical models, such region graphs or junction trees. These graphical models, together with the update schedule, entirely determine the behavior of the inference algorithm in a centralized settings. Yet, in distributed settings, the graphical model crucially interacts with the physical network and determines properties, such as robustness or communication complexity. In this thesis, we propose a unified view where the graphical model and its placement is optimized jointly to match both the network and the probabilistic model. In this manner, our distributed algorithms will not only attain accurate solutions, but will also have a low message complexity.
Recent advances in peer-to-peer networks offer interesting opportunities for learning latent variable models for collaborative filtering. Peer-to-peer networks simplify many aspects of distributed learning, but open an interesting challenge of supporting recommendation queries with stale local models. We propose a pull-based approach that updates the model parameters, in order to minimize its regret with respect to the optimal set of recommendations.
We demonstrate our algorithms on real-world applications in large-scale modular robot localization, camera networks and movie recommendation systems. We demonstrate that our algorithms scale to large networks and provide improved robustness and convergence properties.
Full Text - http://www.cs.cmu.edu/~sfuniak/pubs/proposal.pdf
No comments:
Post a Comment