Contributed Talk: The Influence of Overparameterization and Regularization on Distributed Learning

Speaker: Nicole Muecke, Technical University Berlin
Talk title: The Influence of Overparameterization and Regularization on Distributed Learning

Time: Tuesday, April 20, 9:10am-9:35am (PT)

Abstract:
We study the influence of overparameterization and regularization on distributed learning in linear regression. We give first insights about the double descent phenomenon also in this setting and show how to mitigate this by optimal regularization. In addition, we give a first upper bound on the excess risk in terms on the effective ranks. Our work opens some new research directions for designing efficient algorithms in overparameterized regimes, in particular neural network learning.

Joint work with Enrico Reiss and Markus Klein.

 

Return to workshop schedule