I. Introduction
The rapid development of device-edge-cloud collaborative computing techniques has revolutionized the field of artificial intelligence (AI)
https://aws.amazon.com/cn/ [1]
. Through the power of multiple device connections, edge computing infrastructures, and centralized cloud resources, this innovative paradigm enables AI models to overcome the limitations of traditional computing architectures [2]. Therefore, a large number of AI tasks arise in our daily lives, including smart healthcare, financial management, and autonomous driving. Since diverse data provide rich information for model selection to obtain good desired results, the feedback data generated by these tasks can potentially enhance the generalization performance of AI models [3]. In addition, the generalization capability of pre-trained models plays a crucial role in boosting the quality of AI serviceshttps://openai.com/research/ai-and-compute
. Currently, domain adaptation (DA) [4], [5], transfer learning (TL) [6], [7] and out-of-distribution (OOD) generalization [8], [9] are usually employed to improve the knowledge transferability and generalization performance of AI models.