This paper introduces FreqX, a new interpretability method based on Signal Processing and Information Theory, designed to provide comprehensive and reliable feature attribution for deep learning models across diverse modalities, with significant speed improvements over existing methods.
The proposed method, FreqX, addresses critical challenges in machine learning interpretability, especially within complex setups like Personalized Federal Learning (PFL), where issues like Non-IID data, device heterogeneity, and unclear contributions hinder deployment. FreqX leverages insights from Signal Processing and Information Theory to generate explanations that contain both attribution information and crucial concept information. Experiments demonstrate that FreqX achieves superior explanatory results while executing at least 10 times faster than state-of-the-art baselines that only provide concept information.