A federated width learning method and system for data heterogeneity

Abstract

The application provides a data-isomerism-oriented federated width learning method and system, comprising the following steps: each client performs local training on a width feature extraction module based on a federated learning model; after the local training is completed, the client uploads the optimized width feature extraction module parameters to a server; the server receives the width feature extraction module parameters uploaded by all clients, performs weighted average aggregation based on the sample quantity of each client, obtains a global width feature extraction module, and returns the global width feature extraction module to each client; each client uses the global width feature extraction module to perform feature extraction on local data, and converts the extracted features into decomposition factors through matrix decomposition; the server performs a summation operation on the received decomposition factors to obtain global aggregated decomposition factors; each client uses the global decomposition factors received from the server to calculate local model weights, and after the weight calculation is completed, the client uploads the local weights to the server.

Images