Continual Learning through Control Minimization

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: de Haan, Sander, Taoudi-Benchekroun, Yassine, Aceituno, Pau Vilimelis, Grewe, Benjamin F.
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866915773998432256
author de Haan, Sander
Taoudi-Benchekroun, Yassine
Aceituno, Pau Vilimelis
Grewe, Benjamin F.
author_facet de Haan, Sander
Taoudi-Benchekroun, Yassine
Aceituno, Pau Vilimelis
Grewe, Benjamin F.
contents Catastrophic forgetting remains a fundamental challenge for neural networks when tasks are trained sequentially. In this work, we reformulate continual learning as a control problem where learning and preservation signals compete within neural activity dynamics. We convert regularization penalties into preservation signals that protect prior-task representations. Learning then proceeds by minimizing the control effort required to integrate new tasks while competing with the preservation of prior tasks. At equilibrium, the neural activities produce weight updates that implicitly encode the full prior-task curvature, a property we term the continual-natural gradient, requiring no explicit curvature storage. Experiments confirm that our learning framework recovers true prior-task curvature and enables task discrimination, outperforming existing methods on standard benchmarks without replay.
format Preprint
id arxiv_https___arxiv_org_abs_2602_04542
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Continual Learning through Control Minimization
de Haan, Sander
Taoudi-Benchekroun, Yassine
Aceituno, Pau Vilimelis
Grewe, Benjamin F.
Machine Learning
Artificial Intelligence
Catastrophic forgetting remains a fundamental challenge for neural networks when tasks are trained sequentially. In this work, we reformulate continual learning as a control problem where learning and preservation signals compete within neural activity dynamics. We convert regularization penalties into preservation signals that protect prior-task representations. Learning then proceeds by minimizing the control effort required to integrate new tasks while competing with the preservation of prior tasks. At equilibrium, the neural activities produce weight updates that implicitly encode the full prior-task curvature, a property we term the continual-natural gradient, requiring no explicit curvature storage. Experiments confirm that our learning framework recovers true prior-task curvature and enables task discrimination, outperforming existing methods on standard benchmarks without replay.
title Continual Learning through Control Minimization
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2602.04542