KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Xia, Zixuan, Davtyan, Aram, Favaro, Paolo
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866918168687017984
author Xia, Zixuan
Davtyan, Aram
Favaro, Paolo
author_facet Xia, Zixuan
Davtyan, Aram
Favaro, Paolo
contents We propose KOALA++, a scalable Kalman-based optimization algorithm that explicitly models structured gradient uncertainty in neural network training. Unlike second-order methods, which rely on expensive second order gradient calculation, our method directly estimates the parameter covariance matrix by recursively updating compact gradient covariance products. This design improves upon the original KOALA framework that assumed diagonal covariance by implicitly capturing richer uncertainty structure without storing the full covariance matrix and avoiding large matrix inversions. Across diverse tasks, including image classification and language modeling, KOALA++ achieves accuracy on par or better than state-of-the-art first- and second-order optimizers while maintaining the efficiency of first-order methods.
format Preprint
id arxiv_https___arxiv_org_abs_2506_04432
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products
Xia, Zixuan
Davtyan, Aram
Favaro, Paolo
Machine Learning
Optimization and Control
We propose KOALA++, a scalable Kalman-based optimization algorithm that explicitly models structured gradient uncertainty in neural network training. Unlike second-order methods, which rely on expensive second order gradient calculation, our method directly estimates the parameter covariance matrix by recursively updating compact gradient covariance products. This design improves upon the original KOALA framework that assumed diagonal covariance by implicitly capturing richer uncertainty structure without storing the full covariance matrix and avoiding large matrix inversions. Across diverse tasks, including image classification and language modeling, KOALA++ achieves accuracy on par or better than state-of-the-art first- and second-order optimizers while maintaining the efficiency of first-order methods.
title KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2506.04432