Nonlinear Filtering with Population Codes.
Year: 2015
Type: preprint
Abstract: In this paper we develop a method for the nonlinear filtering of dynamic population codes with unknown latent dynamics. This method is based on a parameterized function which updates beliefs while keeping the dimension of the belief space constant. In order to optimize this function, we construct a graphical model representation of the belief-latent state dynamics, and show that a certain marginalization of the graphical model has the form of a conditional exponential family harmonium. This marginalization allows us to maximize the likelihood of observations, and thereby optimize the parameterized function, by the method of contrastive divergence minimization. By defining the parameterized function as a multilayer perceptron, we show how the stochastic contrastive divergence gradient may be computed with backpropagation. We conclude by demonstrating the effectiveness of this approach on a filtering problem based on a stochastic pendulum.
Source: arXiv (Cornell University)
Author: Sacha Sokoloski
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2207308817