The invention presents a federated learning method combining partial transfer learning and hypernetwork-based personalization for non-image data. A global server shares transformer attention block parameters, while nodes locally generate classifier blocks, preserving personalization. Unlike traditional methods using embedding layers, this approach personalizes projection layer parameters. Selective parameter averaging enables certain model components to remain local, enhancing adaptability. Nodes share predefined attention block parameters, which the server aggregates or customizes for specific models. This method improves model initialization, personalization, and adaptability in federated learning for non-image data tasks.
|
技術摘要：
專利行銷平台
