Federated domain adaptation for healthcare

Jiang, Enyi

Permalink

https://hdl.handle.net/2142/120432 Copy

Description

Title

Federated domain adaptation for healthcare

Author(s)

Jiang, Enyi

Issue Date

2023-05-02

Director of Research (if dissertation) or Advisor (if thesis)

Koyejo, Oluwasanmi

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

Federated Learning, Domain Adaptation, ML for Healthcare

Abstract

Federated domain adaptation (FDA) describes the setting where a set of source clients seek to optimize the performance of a target client. To be effective, FDA must address some of the distributional challenges of Federated learning (FL). For instance, FL systems exhibit distribution shifts across clients. Further, labeled data are not always available among the clients. To this end, we propose and compare novel approaches for FDA, combining the few labeled target samples with the source data when auxiliary labels are available to the clients. The in-distribution auxiliary information is included during local training to boost out-of-domain accuracy. Also, during fine-tuning, we devise a simple yet efficient gradient projection method (FedGP) to detect the valuable components from each source client model by comparing them with the target direction. The extensive experiments on healthcare datasets show that our proposed framework outperforms the state-of-the-art unsupervised FDA methods with limited additional time and space complexity. Additionally, we find that common techniques such as FedAvg and fine-tuning fail with a large domain shift. To better investigate the effectiveness of FedGP under various extents of domain shift, we perform extensive semi-synthetic and real-world experiments on general-purposed datasets compared with several baselines. Our results indicate a bias-variance trade-off between source and target domains when combining source and target gradients. FedGP maintains a better trade-off between source gradients' bias (the domain shift between source and target domains) and the target gradient's variance from limited labeled data. Our experiments illustrate the effectiveness of the proposed method in practice.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120432

Copyright and License Information

Copyright 2023 Enyi Jiang

Owning Collections