|Mittagsseminar Talk Information|
Date and Time: Thursday, March 03, 2011, 12:15 pm
Duration: This information is not available in the database
Location: CAB G51
Speaker: Martin Jaggi
We consider parameterized convex optimization problems over the unit simplex, that depend on one additional parameter. We provide a simple scheme for maintaining an ε-approximate solution along the entire parameter path, using as few updates ("warm-start") as possible. Surprisingly the necessary number of updates is O(1/ε), independent of the dimension of the problem.
This idea can also be extended to optimization problems over matrices. We will in particular consider machine learning applications, where it is often difficult to select a good parameter. One consequence is that we can efficiently compute the solution path for matrix factorization problems.
Joint work with Joachim Giesen and Soeren Laue.
Automatic MiSe System Software Version 1.4803M | admin login