Colloquium (as part of Raytheon Day)

Particle Flow for Filtering and Bayesian Deep Learning

Particle flow is an algorithm to implement Bayes’ rule efficiently in high dimensions.  It has been applied for many nonlinear and non-Gaussian estimation problems as well as for Bayesian deep learning.  Particle flow is many orders of magnitude faster than standard particle filters for high dimensional estimation problems.  The basic idea of particle flow is to turn Bayes’ rule into a flow of particles that represent the posteriori probability density.  We do this by deriving a law of motion of the particles, as in physics.  The law of motion is the solution of a highly underdetermined linear partial differential equation.  We solve this PDE exactly using a method of Gromov.  Particle flow is also embarrassingly parallelizable, because we can exploit parallelism over particles.  In contrast, standard particle filters cannot exploit such parallelism because they resample particles.  We never resample particles, because we avoid particle degeneracy by moving our particles to the correct regions of state space.

Location:  In person: UofA Campus: ENR2 building, Room S395 or  Zoom:  Password: Math


2 p.m. Friday


ENR2 Room S395 and on Zoom