[4Yin2] Interactive session 2

[4Yin2-46] Gradient-Based Communication Network Optimization for Fully Decentralized Learning

〇Naoyuki Terashita¹, Satoshi Hara² (1.Hitachi, Ltd., 2.Osaka University )

Keywords:Distributed Learning, Hyper parameter optimization, Stochastic Gradient Descent, Federated Learning

We propose a gradient-based communication network optimization algorithm for fully decentralized learning.
Our algorithm traces the gradients of network edge weights throughout the training in a fully decentralized manner.
We applied the proposed algorithm to convergence acceleration and evaluated its performance by simulation experiments.