centralized-learning

OmniISR proposes a unified framework combining centralized and federated learning via intermediate supervision and regularization at hidden layers, offering theoretical convergence guarantees and reducing the CL–FL gap by 22.60%.