Human–Machine Hybrid Federated Learning: Concept and Applications

Federated Learning (FL) has become a state-of-the-art technique for addressing data isolation and privacy problems. However, the traditional FL framework has limitations on lack of labeled data, adaptation to evolving environments and tasks, and insufficient generalization of the global model due to non-independent and identically distributed (non-IID) data. These issues suggest incorporating human knowledge and interaction into the FL workflow can be beneficial. Human-machine hybrid intelligence is an area that human abilities are considered to prompt the usability and robustness of the system by providing human domain knowledge. Combining FL and human-machine hybrid intelligence can fully utilize their benefits and complement each other perfectly. This article presents our vision of the next generation of FL, Human-Machine Hybrid Federated Learning, namely HIFL. And this work first defines the concept of HIFL and proposes three patterns of collaboration for HIFL: (1) local hybrid intelligence (LocalHIFL), (2) separate hybrid intelligence (SeparateHIFL), and (3) cross hybrid intelligence (CrossHIFL). In each pattern, we survey methodologies and techniques that are utilized to address specific problems that occurred after adding human-machine collaboration in FL process. Besides, we exhibit some potential application scenarios, and provide several open challenges and opportunities contained in HIFL. This survey intends to provide a high-level summarization for improving FL by combining human-machine hybrid intelligence, and to motivate interested readers to consider approaches for designing effective FL approaches and ways of intelligence fusion according to their requirements.