MACRA: Efficient Knowledge Sharing among Personalized Models against Label Shift in Mobile AI Applications

Runtime adaptation frameworks in mobile applications aim to calibrate model parameters to sustain high performance in evolving deployment scenarios, but label shifts alter label distributions across personalized models, rendering existing adaptation methods deficient. Knowledge sharing offers a promising solution by enabling personalized models to share relevant information, reducing redundant computations during independent fine-tuning. However, a major challenge lies in how to balance the communication cost with adaptation accuracy. This paper presents MACRA, which employs an adaptive, selective model merging mechanism based on arithmetic parameter combinations to facilitate knowledge sharing among mobile devices in response to label shifts. MACRA makes three contributions: First, it identifies and selects the most valuable parameters that reflect the shift, enabling high inference accuracy across different shifts. Second, it automatically identifies two types of label shifts and incorporates a merged model cache and check mechanism to ensure the high-quality updates across devices and minimize unnecessary computations. Third, it dynamically estimates the maximal feasible compression rate per shift, achieving optimal communication efficiency with minimal accuracy loss. Experiments on six human-centered datasets across four modalities show that MACRA compresses the transmitted parameters by up to 1/3000 of the original model, achieving 5x higher transmission compression rate than the best baseline, while improving the overall task accuracy by 13.5% and reducing the latency by 61% across different label shifts.