Large-scale models, also referred to as pretrained models, have attracted significant attention due to their outstanding performance and robust generalization. However, the high demands for data quality, stringent data privacy and security requirements, and limited computational and communication resources have imposed restrictions on the further development of large-scale models. Federated Learning (FL) is a popular machine learning framework that can effectively address this issue. However, when collaboratively training large-scale models using FL, local training and the transmission of large-scale parameters impose significant computational and communication burdens on mobile devices. This paper proposes a light-weight and high-efficiency federated learning framework (FedLH) for large-scale models. This framework divides large-scale models into semantic block-based submodels, allowing clients to transmit these submodels to the server for heterogeneous aggregation. This approach enables both communication and computational efficiency. With the proposed framework, each device can learn personalized, structured sparse models that can efficiently run on terminal devices. Experimental results demonstrate that FedLH outperforms other baseline algorithms by significantly reducing the number of training parameters and transmitted data. It also exhibits strong generalization and scalability.

Temporal appearance misalignment is a crucial problem in video person re-identification. The same part of person (e.g. head or hand) appearing on different locations in video sequence weakens its discriminative ability, especially when we apply standard temporal aggregation such as 3D convolution or LSTM. To address this issue, we propose Self-Separated network (SSN) to seek out the same parts in different images. As the name implies, SSN, if trained in an unsupervised strategy, guarantees the selected parts distinct. With a few samples of labeled parts to guide SSN training, this semi-supervised trained SSN seeks out the parts that are human-understandable within a frame and stable across a video snippet. Given the distinct and stable person parts, rather than performing aggregation on features, we then apply 3D convolution across different frames for person re-identification. This SSN + 3D pipeline, dubbed SSN3D, is proved to be efficient through extensive experiments on both synthetic and real data.