An asynchronous federated continual learning method for large models

By employing an asynchronous federated continuous learning approach, the client constructs feature prototypes and covariance statistics, asynchronously uploads and fuses feature information, thus solving the problems of training blockage and historical knowledge forgetting in federated learning. This enables efficient collaboration and stable updates of large financial models in heterogeneous and dynamic environments.

CN121859998BActive Publication Date: 2026-06-02SOUTHWESTERN UNIV OF FINANCE & ECONOMICS
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
SOUTHWESTERN UNIV OF FINANCE & ECONOMICS
Filing Date
2026-03-17
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing federated learning methods suffer from training blockages, insufficient model adaptability, and historical knowledge forgetting issues due to synchronous update mechanisms in financial business scenarios, making it difficult to adapt to the heterogeneity and dynamism among multiple institutions.

Method used

An asynchronous federated continuous learning approach is adopted. The client constructs feature prototypes and covariance statistics by freezing the global feature extraction module, asynchronously uploads incremental feature information, and fuses it on the server side. Adaptive fusion weights are constructed by combining the client's quality score, and a covariance-aware Bayesian classifier is built for decision modeling.

Benefits of technology

It improves the efficiency of cross-organizational collaboration, continuously absorbs new business knowledge, suppresses the forgetting of historical knowledge, maintains the consistency and stability of global feature representation, and improves the classification accuracy of new task data and the robustness of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121859998B_ABST
    Figure CN121859998B_ABST
Patent Text Reader

Abstract

The application discloses a large model-oriented asynchronous federated continuous learning method, belongs to the field of federated learning and continuous learning, and is applied to a client and comprises the following steps: receiving initial model parameters of a global model and a global feature extraction module issued by a server, performing feature representation on local business task data based on the module, and constructing corresponding feature prototypes and feature covariance statistics; uploading the feature prototypes and the feature covariance statistics to the server in an asynchronous mode to perform asynchronous federated fusion and update global feature statistics; and receiving the updated global feature statistics fused by the server, wherein the global feature statistics are used for local decision modeling of subsequent tasks. Through the above asynchronous federated aggregation strategy, incremental feature knowledge from different financial institutions and different task stages can be continuously integrated, the problem of forgetting historical knowledge can be effectively alleviated, and the stability and consistency of global feature representation in the long-term evolution process can be maintained.
Need to check novelty before this filing date? Find Prior Art