Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Decentralized Federated Dataset Dictionary Learning for Multi-Source Domain Adaptation

Rebecca Clain, Eduardo Fernandes Montesuma, Fred Ngolè Mboula

TL;DR

The paper tackles decentralized multi-source domain adaptation by removing the central server from FedDaDiL and modeling cross-client shifts with Wasserstein barycenters B(alpha;P) under the distance W_c. It introduces De-FedDaDiL, a fully decentralized algorithm where each client maintains its own atom dictionary P_l and barycentric coordinates alpha_l, exchanging atoms with random peers and updating locally for E epochs. Empirical results on MSDA benchmarks show performance close to or exceeding federated FedDaDiL while reducing communication and eliminating a single point of failure. The work demonstrates consensus dynamics of dictionaries via decreasing barycenter distances, supporting robustness and scalability for private distributed domain adaptation.

Abstract

Decentralized Multi-Source Domain Adaptation (DMSDA) is a challenging task that aims to transfer knowledge from multiple related and heterogeneous source domains to an unlabeled target domain within a decentralized framework. Our work tackles DMSDA through a fully decentralized federated approach. In particular, we extend the Federated Dataset Dictionary Learning (FedDaDiL) framework by eliminating the necessity for a central server. FedDaDiL leverages Wasserstein barycenters to model the distributional shift across multiple clients, enabling effective adaptation while preserving data privacy. By decentralizing this framework, we enhance its robustness, scalability, and privacy, removing the risk of a single point of failure. We compare our method to its federated counterpart and other benchmark algorithms, showing that our approach effectively adapts source domains to an unlabeled target domain in a fully decentralized manner.

Decentralized Federated Dataset Dictionary Learning for Multi-Source Domain Adaptation

TL;DR

The paper tackles decentralized multi-source domain adaptation by removing the central server from FedDaDiL and modeling cross-client shifts with Wasserstein barycenters B(alpha;P) under the distance W_c. It introduces De-FedDaDiL, a fully decentralized algorithm where each client maintains its own atom dictionary P_l and barycentric coordinates alpha_l, exchanging atoms with random peers and updating locally for E epochs. Empirical results on MSDA benchmarks show performance close to or exceeding federated FedDaDiL while reducing communication and eliminating a single point of failure. The work demonstrates consensus dynamics of dictionaries via decreasing barycenter distances, supporting robustness and scalability for private distributed domain adaptation.

Abstract

Decentralized Multi-Source Domain Adaptation (DMSDA) is a challenging task that aims to transfer knowledge from multiple related and heterogeneous source domains to an unlabeled target domain within a decentralized framework. Our work tackles DMSDA through a fully decentralized federated approach. In particular, we extend the Federated Dataset Dictionary Learning (FedDaDiL) framework by eliminating the necessity for a central server. FedDaDiL leverages Wasserstein barycenters to model the distributional shift across multiple clients, enabling effective adaptation while preserving data privacy. By decentralizing this framework, we enhance its robustness, scalability, and privacy, removing the risk of a single point of failure. We compare our method to its federated counterpart and other benchmark algorithms, showing that our approach effectively adapts source domains to an unlabeled target domain in a fully decentralized manner.

Paper Structure

This paper contains 10 sections, 7 equations, 2 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. RELATED WORK
  3. Decentralized Federated Learning
  4. Decentralized Multi-Source Domain Adaptation
  5. PROPOSED APPROACH
  6. Background
  7. Decentralized federated dataset dictionary learning
  8. EXPERIMENTS AND DISCUSSION
  9. Analyzing Consensus Among Clients' Atoms
  10. Conclusion

Figures (2)

  • Figure 1: De-FedDaDiL. Each client initializes atoms and, at each iteration, exchanges atom versions ($\mathcal{P}$) with a randomly selected peer. Clients update their models based on their own and received version, while keeping barycentric coordinates private ($\alpha$). The learning process is collaborative and doesn't require a central server.
  • Figure 2: Max Wasserstein distance between the barycenters of the different domains over each training iteration - Target domain is art

Theorems & Definitions (1)

  • Definition 1