Tuning Personalized Models by Two-Phase Parameter Decoupling with Device-to-Device Communication
Atsuya Muramatsu, Hideya Ochiai, Hiroshi Esaki · 2024
Edge machine learning in IoT applications has attracted attentions along with the development of AI chips. In such a decentralized scenario, collaboration at the edge is required to improve the performance of machine learning models. In collaborative edge machine learning, personalized models are beneficial for participants due to the data heterogeneity problem. This paper explores personalization methods for image recognition tasks in wireless ad hoc federated learning (WAFL), which is a framework of edge-based collaborative learning. We propose Two-Phase Parameter Decoupling, which is a combination of the original WAFL and parameter decoupling. We used UTokyo Building Recognition Dataset to evaluate our proposed approaches. Two-Phase Parameter Decoupling attained 2.3% higher personalization accuracy than the original WAFL.