The Memory Perturbation Equation: Understanding Model's Sensitivity to Data

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Nickl, Peter, Xu, Lu, Tailor, Dharmesh, Möllenhoff, Thomas, Khan, Mohammad Emtiyaz
Formato: Preprint
Publicado: 2023
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866929210623262720
author Nickl, Peter
Xu, Lu
Tailor, Dharmesh
Möllenhoff, Thomas
Khan, Mohammad Emtiyaz
author_facet Nickl, Peter
Xu, Lu
Tailor, Dharmesh
Möllenhoff, Thomas
Khan, Mohammad Emtiyaz
contents Understanding model's sensitivity to its training data is crucial but can also be challenging and costly, especially during training. To simplify such issues, we present the Memory-Perturbation Equation (MPE) which relates model's sensitivity to perturbation in its training data. Derived using Bayesian principles, the MPE unifies existing sensitivity measures, generalizes them to a wide-variety of models and algorithms, and unravels useful properties regarding sensitivities. Our empirical results show that sensitivity estimates obtained during training can be used to faithfully predict generalization on unseen test data. The proposed equation is expected to be useful for future research on robust and adaptive learning.
format Preprint
id arxiv_https___arxiv_org_abs_2310_19273
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle The Memory Perturbation Equation: Understanding Model's Sensitivity to Data
Nickl, Peter
Xu, Lu
Tailor, Dharmesh
Möllenhoff, Thomas
Khan, Mohammad Emtiyaz
Machine Learning
Artificial Intelligence
Understanding model's sensitivity to its training data is crucial but can also be challenging and costly, especially during training. To simplify such issues, we present the Memory-Perturbation Equation (MPE) which relates model's sensitivity to perturbation in its training data. Derived using Bayesian principles, the MPE unifies existing sensitivity measures, generalizes them to a wide-variety of models and algorithms, and unravels useful properties regarding sensitivities. Our empirical results show that sensitivity estimates obtained during training can be used to faithfully predict generalization on unseen test data. The proposed equation is expected to be useful for future research on robust and adaptive learning.
title The Memory Perturbation Equation: Understanding Model's Sensitivity to Data
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2310.19273