Motonobu Kanagawa
Wednesday 19th November 2014
Time: 4pm
4th Floor Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
Monte Carlo Filtering using Kernel Embedding of Distributions
Kernel embedding of distributions is a nonparametric method for
representing and estimating distributions using reproducing kernels,
which has various applications in machine learning and statistics. In
this talk, I will present a filtering algorithm for a state-space model
based on kernel embedding. Specifically, I will focus on the following
setting: (i) the observation model is unknown even in parametric form,
but state-observation examples are given as training data, while (ii) a
good model for state-transition is known. The proposed filter has
potential applications in fields that involve complex observation
processes, such as robotics and brain-computer interface. As an
illustrative application, the proposed filter is applied to the robot
localization problem, which is a fundamental task in robotics.