FedHBM: A Local Correction Method for Non-IID Federated Learning

The authors of this paper propose a new method called FEDHBM (Federated Hierarchical Boltzmann Machine) for training deep neural networks on distributed data without compromising data privacy. They address two main challenges in federated learning: (1) dealing with non-IID data across clients, and (2) scaling to large numbers of clients while maintaining the accuracy of the model.
FEDHBM addresses these challenges by introducing a hierarchical Boltzmann machine (HBM) structure that allows for efficient communication between clients and a federated aggregation method that ensures accurate model training. The HBM enables the sharing of information across different layers of the neural network, allowing the model to capture complex relationships in the data.
The authors demonstrate the effectiveness of FEDHBM through experiments on several benchmark datasets, including CIFAR-10 and CIFAR-100. They show that FEDHBM outperforms existing federated learning methods, including FedAvg and Scaffold, in terms of both accuracy and efficiency.
The key insight behind FEDHBM is the use of a hierarchical structure to organize the neural network layers, which enables efficient communication between clients and improves the scalability of the model. By using this approach, FEDHBM can train deep neural networks on large datasets without compromising data privacy or accuracy.
In summary, FEDHBM is a novel method for training deep neural networks in a federated learning setting that addresses the challenges of non-IID data and scalability. By introducing a hierarchical structure to the neural network, FEDHBM enables efficient communication between clients and improves the accuracy of the model.