Flexible federated learning method and device for intelligent pharmaceutical small molecule property prediction

Abstract

The application relates to a flexible federated learning method and device for intelligent pharmaceutical small-molecule property prediction, wherein the method comprises the following steps: performing heterogeneous deep learning modeling and federated collaborative training according to private data of each participating institution and a global model of a central server, obtaining a local model, and updating the global model; uploading update data of each local model to the central server; constructing a gradient buffer corresponding to each participating institution in the central server based on the update data; detecting an institution state of each participating institution; when the institution state is an exit state or a joining state, processing institution data of a target participating institution under a preset dynamic weight strategy according to the detected institution state and the gradient buffer. Through the application, the problems of lack of adaptability to institution dynamic entry and exit and data compliance guarantee in the related art are solved, and compliance erasure of institution data of an exiting machine and adaptive smooth fusion of a newly joined institution are realized.

Images