The Ubiquitous Technology Research Center has made a breakthrough in large-scale heterogeneous data fusion technology for models without training, constructing the FedBone framework, achieving an optimized balance of utility, privacy, and fairness, and implementing the digital ophthalmology federated collaboration platform (FedEYE).
FedBone “links” the large model construction framework, enabling the fusion of heterogeneous task data into a large model with optimized computing power, task optimization, and leading accuracy. It refines the MoFedNet technology framework, allowing pre-trained neural networks to be packaged into graph-structured data for secure transfer across different institutions and domains without training data. Through efficient model fusion, it achieves cross-domain collaboration between different models, balancing privacy and utility. Furthermore, model fusion and group evolution Tare carried out on the premise of model privacy protection, achieving the balance of the “impossible triangle” of utility, privacy, and fairness.
Based on the above core technological achievements, the digital ophthalmology federated collaboration platform FedEYE was constructed. It supports 39 federated learning algorithms and provides diagnostic models for 20 common eye diseases, with an average diagnostic accuracy rate exceeding 94.61%. Related research has been published in the journal Patterns (Cell Press). The international cooperation of “linking” large models has been accelerated. Based on the “Model as a Service” application model, and focusing on the auxiliary diagnosis of neurodegenerative diseases, a successful deployment of a collaborative ecosystem was made in Singapore. Furthermore, the theme of “Fusion Computing Empowering New Productive Forces” was addressed at the 8th Asia-Pacific Forum on Elderly Care and Disability Assistance, actively promoting the construction of an international collaborative ecosystem for life and health in the Asia-Pacific region.
downloadFile