Paper Image

Understanding Hessian alignment for domain generalization

Published on:

22 August 2023

Primary Category:

Machine Learning

Paper Authors:

Sobhan Hemati,

Guojun Zhang,

Amir Estiri,

Xi Chen


Key Details

Shows Hessian alignment reduces transfer measure, improving transferability

Analyzes Hessian alignment as feature matching, relating to CORAL and V-REx

Proposes methods to efficiently match Hessians using Hessian-gradient product

Achieves state-of-the-art on colored MNIST and DomainBed benchmarks

AI generated summary

Understanding Hessian alignment for domain generalization

This paper analyzes the role of aligning the Hessian matrices of neural network classifiers across domains, showing theoretically and empirically that it improves out-of-distribution generalization. The key idea is that matching curvature of the loss landscape makes classifiers similarly near-optimal across domains.

Answers from this paper


No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up