UCSC-SOE-10-08: Learning Rotations Online

Adam M. Smith, Manfred Warmuth
02/12/2010 09:00 AM
Computer Science
In this paper we show the matrix von Mises-Fisher (vMF) distribution is a reasonable distribution on which to build an online density estimation scheme for rotation matrices. We also consider a special case, the unit circle, for initial experimentation. The vector and matrix vMF distributions admit online algorithms in terms of their expectation parameters as a direct result of their exponential family nature, however proving regret bounds for these algorithms is arrested by the complexity of the cumulant functions for these distributions. In closing, we volunteer hazards and suggestions for immediate future work.

UCSC-SOE-10-08