Picking Models for Heterogeneous Clients: A Server-Client Feature Contrastive Learning Design

Conventional deep learning applications for clients generally adapt a global model trained on large-scale server data to different clients, or rely on meta-learning methods to be client-adaptive. However, cross-client data heterogeneity and the building of meta-training resources on server result in degraded client performance and excessive server overhead. In this work, we propose ContrastPick, a design friendly to both sides of the server-client scenario, to provide proper deep learning models to different clients. Given a large-scale dataset and a collection of candidate models from a pretrained supernet on server, we separate the server dataset into smaller subsets and then evaluate the candidates on such subsets. We introduce a contrastive approach to learn a dataset encoder to minimize the similarity between separated subset pairs, thereby matching a client dataset to a suitable subset with high-performing candidate models. With ContrastPick, we can efficiently and effectively serve models to clients without introducing massive re-training and storage overhead caused by meta-training resources. Extensive experimental results on 4 datasets composed of heterogeneous clients demonstrate that ContrastPick can achieve superior performance with significantly reduced overhead compared with baselines.