Take Your Pick: Enabling Effective Distributed Learning Within Low-Dimensional Feature Space.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Personalized federated learning (PFL) enables diverse client models for cross-domain applications like autonomous driving and medical diagnosis.
• Current PFL models use a global encoder for universal features and personalized layers, but domain gaps cause irrelevant feature components.
• Existing methods personalizing encoder parameters face challenges due to high dimensionality and nonlinearity.

Purpose of the Study:

• To propose FedPick, a novel PFL framework operating in the low-dimensional feature space.
• To address the challenge of irrelevant universal features in cross-domain PFL by adaptively selecting task-relevant features.
• To provide a more accessible and interpretable PFL implementation.

Main Methods:

• FedPick adaptively selects task-relevant features from global encoder outputs based on local data distribution.
• The framework operates within the lower-dimensional feature space, offering greater intuitiveness and interpretability.
• Feature selection is performed client-wise to tailor universal features to local tasks.

Main Results:

• FedPick effectively selects task-relevant features for each client in cross-domain scenarios.
• Experimental results demonstrate significant improvements in model performance across multiple datasets.
• The proposed method shows superior performance compared to existing PFL techniques.

Conclusions:

• FedPick offers an effective and interpretable solution for personalized federated learning in cross-domain settings.
• Adaptive feature selection in the feature space is a viable alternative to parameter personalization.
• The framework has strong potential for applications requiring robust cross-domain learning.