Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bayesian Federated Learning for Continual Training

Usevalad Milasheuski, Luca Barbieri, Sanaz Kianoush, Monica Nicoli, Stefano Savazzi

TL;DR

The paper tackles continual learning under non-stationary data distributions in a privacy-preserving federated setting by developing a Bayesian Federated Learning (BFL) framework. It uses Stochastic Gradient Langevin Dynamics (SGLD) to sample from the posterior and exploits the previous day’s posterior as the prior for the next day, forming Posterior-aided Continual Learning (P-CL). The approach is demonstrated on a radar-based human sensing use case with a TD-MIMO FMCW radar network, showing that P-CL delivers competitive accuracy while improving calibration (lower ECE) and reducing convergence iterations compared to transfer learning and full retraining. This indicates practical potential for uncertainty-aware, continual adaptation in safety-critical, privacy-conscious distributed systems; future work explores more expressive priors and privacy-preserving extensions.

Abstract

Bayesian Federated Learning (BFL) enables uncertainty quantification and robust adaptation in distributed learning. In contrast to the frequentist approach, it estimates the posterior distribution of a global model, offering insights into model reliability. However, current BFL methods neglect continual learning challenges in dynamic environments where data distributions shift over time. We propose a continual BFL framework applied to human sensing with radar data collected over several days. Using Stochastic Gradient Langevin Dynamics (SGLD), our approach sequentially updates the model, leveraging past posteriors to construct the prior for the new tasks. We assess the accuracy, the expected calibration error (ECE) and the convergence speed of our approach against several baselines. Results highlight the effectiveness of continual Bayesian updates in preserving knowledge and adapting to evolving data.

Bayesian Federated Learning for Continual Training

TL;DR

The paper tackles continual learning under non-stationary data distributions in a privacy-preserving federated setting by developing a Bayesian Federated Learning (BFL) framework. It uses Stochastic Gradient Langevin Dynamics (SGLD) to sample from the posterior and exploits the previous day’s posterior as the prior for the next day, forming Posterior-aided Continual Learning (P-CL). The approach is demonstrated on a radar-based human sensing use case with a TD-MIMO FMCW radar network, showing that P-CL delivers competitive accuracy while improving calibration (lower ECE) and reducing convergence iterations compared to transfer learning and full retraining. This indicates practical potential for uncertainty-aware, continual adaptation in safety-critical, privacy-conscious distributed systems; future work explores more expressive priors and privacy-preserving extensions.

Abstract

Bayesian Federated Learning (BFL) enables uncertainty quantification and robust adaptation in distributed learning. In contrast to the frequentist approach, it estimates the posterior distribution of a global model, offering insights into model reliability. However, current BFL methods neglect continual learning challenges in dynamic environments where data distributions shift over time. We propose a continual BFL framework applied to human sensing with radar data collected over several days. Using Stochastic Gradient Langevin Dynamics (SGLD), our approach sequentially updates the model, leveraging past posteriors to construct the prior for the new tasks. We assess the accuracy, the expected calibration error (ECE) and the convergence speed of our approach against several baselines. Results highlight the effectiveness of continual Bayesian updates in preserving knowledge and adapting to evolving data.

Paper Structure

This paper contains 6 sections, 9 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Distributed Bayesian Optimization
  3. Continual Bayesian FL Methodology
  4. Continual learning use case: human sensing
  5. Accuracy vs. Confidence results and analysis
  6. Conclusions

Figures (3)

  • Figure 1: Bayesian FL for human sensing. Each robot cell is equipped with a radar measuring the angle $\alpha$ and the distance $l$ to the operator (range-azimuth map).
  • Figure 2: Continual learning example. Black crosses represent the Regions of Interest (ROIs) for the current day, while the gray ones represent the targets for the past day to highlight the misalignment between days.
  • Figure 3: Confidence histograms (top) and reliability diagrams (bottom) for the Transfer Learning (left), Model Retraining (middle) and Posterior-aided Continual Learning (right). The blue color represents the model outputs (confidence), whereas the red color represents the calibration gap, which is the difference between accuracy and confidence for a given bin.