Cloud-Device Collaborative Adaptation to Continual Changing Environments in the Real-World

1 st round 10 th round 5 th round (b) Comparison of our method with others on continuous domain shifts Visual Prompts Uncertain samples Student model (Small model) + Model parameters Uncertainty Device Cloud Teacher model (Large model) VPLU strategy Changing environment Uncertainty Guided Sampling Data flow Visual prompts Uncertain sample Up-link Down-link Processed prompts (a) The problem and our main idea Figure 1. (a) Models deployed on devices are preferably lightweight. However, device models will suffer from severe performance degradation when facing continual distribution shift data. Our main idea is to improve the continual domain adaptation capability of the device model by performing our proposed Cloud-Device Collaborative Adaptation paradigm. (b) We compare our method with previous works [13, 28, 29] . Our method surpasses the state-of-the-art approach and exhibits a solid ability when facing continual distribution shifts.