Decentralized Black-Box Variational Inference for Bayesian Learning on Sensor Networks

In large sensor networks, often information from data collected by all individual sensors is desired without aggregating the data in any one place. Particularly when the sensors in question are on autonomous platforms and make timely decisions based on this information, a full quantification of unce...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Conference record - Asilomar Conference on Signals, Systems, & Computers S. 150 - 155
Hauptverfasser: Cadena, Jose, Ray, Priyadip, Goldhahn, Ryan
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 31.10.2021
Schlagworte:
autonomous sensor networks
Bayes methods
Data aggregation
decentralized Bayesian inference
Estimation
Inference algorithms
Markov processes
Monte Carlo methods
Uncertainty
variational Bayes
ISSN:2576-2303
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In large sensor networks, often information from data collected by all individual sensors is desired without aggregating the data in any one place. Particularly when the sensors in question are on autonomous platforms and make timely decisions based on this information, a full quantification of uncertainty is more useful than a point estimate. Fully decentralized Bayesian inference in large autonomous sensor networks remains a challenging research problem, due to the inherent latency of Markov Chain Monte Carlo (MCMC) inference. To extend fully decentralized Bayesian inference to applications with stringent time constraints, we propose a distributed black-box variational inference algorithm. In our proposed approach, each sensor learns an approximation to the centralized posterior distribution via exchanging messages with a random subset of sensors in the network; thus, sensors collaboratively perform Bayesian learning without the need for expensive data aggregation. We demonstrate in simulation that the proposed approach can achieve a good approximation to the posterior orders of magnitude faster than a distributed MCMC based approach with a meagre increase in compute power at individual nodes.
ISSN:2576-2303
DOI:10.1109/IEEECONF53345.2021.9723409