Improved Information Theoretic Generalization Bounds for Distributed and Federated Learning
Authors:
- L. P. BarnesLBL. P. Barnes
- H. V. PoorHPH. V. Poor
Preprint
Abstract
We consider information-theoretic bounds on expected generalization error for statistical learning problems in a networked setting. In this setting, there are $K$ nodes, each with its own independent dataset, and the models from each node have to be aggregated into a final centralized model. We consider both simple averaging of the models as well as more complicated multi-round algorithms. We give upper bounds on the expected generalization error for a variety of problems, such as those with Bregman divergence or Lipschitz continuous losses, that demonstrate an improved dependence of $1/K$ on the number of nodes. These "per node" bounds are in terms of the mutual information between the training dataset and the trained weights at each node, and are therefore useful in describing the generalization properties inherent to having communication or privacy constraints at each node.
Keywords
All Related Versions
- Version 1, 2022-02-04, ArXiv
- Version 2, 2024-01-15, ArXiv
- Published version: Entropy, 24 (9), 1178.
This publication has 5 references indexed in Scilit:
- rTop-k: A Statistical Estimation Approach to Distributed SGDIEEE Journal on Selected Areas in Information Theory, 2020
- Tightening Mutual Information-Based Bounds on Generalization ErrorIEEE Journal on Selected Areas in Information Theory, 2020
- How Much Does Your Data Exploration Overfit? Controlling Bias via Information UsageIEEE Transactions on Information Theory, 2019
- What Can We Learn Privately?SIAM Journal on Computing, 2011
- Clustering with Bregman Divergences2004