bayesian-filtering

Natural AI is Artificial Intelligence based on physics principles. The Free Energy Principle allows for designing agents that learn in real-time, are low-power and are explainable. They are powered by Bayesian mechanics, updating beliefs over states of the world and actions as data streams in. At BIASlab and LazyDynamics, we have developed a toolbox for scalable automatic Bayesian inference called RxInfer.

We propose a variational message passing inference algorithm for online system identification using polynomial NARMAX models. We show empirically that our variational Bayesian estimator outperforms an online recursive least-squares estimator, most notably in small sample size settings and low noise regimes, and performs on par with an iterative least-squares estimator trained offline.

A talk for the Optimization / Machine Learning Competence group at Sioux Technologies. I presented the derivations to make a Bayesian filter for a standard Gaussian linear dynamical system, and outlined variational Bayesian inference as an extension. I briefly showed some of BIASlab’s research.

A paper on online system identification, where I cast the nonlinear stochastic differential equation of a Duffing oscillator to a probabilistic graphical model and use variational message passing to infer dynamical parameters of the system.

Message passing on factor graphs typically relies on a scheduling procedure, in which a central algorithm or compiler figures out which nodes should pass messages where at what time. This is not a biologically plausible mechanism. I explore the possibility of passing messages without a scheduler, where the nodes merely “react” to incoming messages.