Sparse Training for Federated Learning with Regularized Error Correction

The paper presents a novel algorithm, Federated Learning with Accumulated Regularized Embeddings (FLARE), which addresses challenges in decentralized machine learning, specifically in federated learning (FL). The algorithm uses sparse training and error correction methods to reduce computational and communication costs, without harming convergence. The authors claim that FLARE can achieve remarkable sparsity levels and significantly improved accuracy. The algorithm is validated through experiments on diverse and complex models.

Publication date: 22 Dec 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2312.13795