Date of Publication :7th October 2025

Abstract: Edge devices with constrained computational resources increasingly require efficient on-device inference capabilities while maintaining high accuracy when deployed in new environments. Domain adaptation methods can mitigate performance degradation in new domains, but current approaches either sacrifice accuracy for memory efficiency or require excessive memory during adaptation. We present UniMEC, a unified framework that jointly optimizes domain adaptation and memory efficiency through a novel architecture that incorporates learning to-branch mechanisms and lite residual modules. Unlike previous methods that address adaptation and efficiency sequentially, UniMEC optimizes these objectives simultaneously through a unified loss function and progressive training scheme. Our approach introduces a branching architecture that dynamically allocates computational resources to the most critical pathways while maintaining knowledge transfer between a large teacher model and a compact edge model. Extensive experiments across standard domain adaptation benchmarks (Office-31 and Office- Home) demonstrate that UniMEC achieves up to 10-15% higher target domain accuracy while reducing memory requirements by 20-25% compared to state-of-the-art methods. Further, our unified training approach significantly reduces the overall adaptation time, making it practical for real-world edge deployment scenarios.

Reference :