(2021) Relative Flatness and Generalization.
|
Text
flatness-petzka,kamp,etal.pdf Download (979kB) | Preview |
Abstract
Flatness of the loss curve is conjectured to be connected to the generalization ability of machine learning models, in particular neural networks. While it has been empirically observed that flatness measures consistently correlate strongly with generalization, it is still an open theoretical problem why and under which circumstances flatness is connected to generalization, in particular in light of reparameterizations that change certain flatness measures but leave generalization unchanged. We investigate the connection between flatness and generalization by relating it to the interpolation from representative data, deriving notions of representativeness, and feature robustness. The notions allow us to rigorously connect flatness and generalization and to identify conditions under which the connection holds. Moreover, they give rise to a novel, but natural relative flatness measure that correlates strongly with generalization, simplifies to ridge regression for ordinary least squares, and solves the reparameterization issue.
| Item Type: | Conference or Workshop Item (A Paper) (Paper) |
|---|---|
| Divisions: | Jilles Vreeken (Exploratory Data Analysis) |
| Conference: | NeurIPS Conference on Neural Information Processing Systems |
| Depositing User: | Jilles Vreeken |
| Date Deposited: | 28 Mar 2022 11:59 |
| Last Modified: | 28 Mar 2022 11:59 |
| Primary Research Area: | NRA5: Empirical & Behavioral Security |
| URI: | https://publications.cispa.saarland/id/eprint/3595 |
Actions
Actions (login required)
 |
View Item |