[2312.12616] On-line Variational Sequential Monte Carlo


Obtain a PDF of the paper titled On-line Variational Sequential Monte Carlo, by Alessandro Mastrototaro and Jimmy Olsson

Obtain PDF

Summary:Being probably the most classical generative mannequin for serial information, state-space fashions (SSM) are basic in AI and statistical machine studying. In SSM, any type of parameter studying or latent state inference usually entails the computation of complicated latent-state posteriors. On this work, we construct upon the variational sequential Monte Carlo (VSMC) technique, which supplies computationally environment friendly and correct mannequin parameter estimation and Bayesian latent-state inference by combining particle strategies and variational inference. Whereas commonplace VSMC operates within the offline mode, by re-processing repeatedly a given batch of knowledge, we distribute the approximation of the gradient of the VSMC surrogate ELBO in time utilizing stochastic approximation, permitting for on-line studying within the presence of streams of knowledge. This leads to an algorithm, on-line VSMC, that’s able to performing effectively, totally on-the-fly, each parameter estimation and particle proposal adaptation. As well as, we offer rigorous theoretical outcomes describing the algorithm’s convergence properties because the variety of information tends to infinity in addition to numerical illustrations of its wonderful convergence properties and usefulness additionally in batch-processing settings.

Submission historical past

From: Alessandro Mastrototaro [view email]
Tue, 19 Dec 2023 21:45:38 UTC (955 KB)
Fri, 2 Feb 2024 16:24:14 UTC (2,904 KB)

Supply hyperlink


Please enter your comment!
Please enter your name here