Federal learning method and system based on unified representation and classifier correction

• Main Authors: YE MANG, LIU YUXIA, HUANG WENKE, CHEN JUN
• Format: Patent
• Language: Chinese; English
• Published: 07.05.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention discloses a federated learning method and system based on unified representation and classifier correction, which are used for solving the problem of federated learning under long-tail data distribution. Firstly, a global unified prototype is obtained by aggregating local prototypes (namely category average features extracted by a global model); the prototypes are utilized to adjust the feature space, and the features in the same category are drawn close to the corresponding global unified prototype and pushed away from other categories at the same time. In addition, the classifier deviation is reduced through prototype mixing by using a global prototype. A balanced virtual feature set is generated by fusing a global unified prototype and local features. A classifier is retrained on the feature set, the decision boundary is corrected, and offset is reduced; the method can effectively improve the classification performance of the model under the long-tail data, and reduces the deviation between t