Why do continual physical learners need sleep?

Sleep is widespread across the animal kingdom, including in organisms without a centralized nervous system. Does learning create a need for sleep, even without biology? To explore this question, we study a disordered spring network trained to perform a sequence of mechanical allosteric tasks through local changes in bond stiffness. We find that continual learning progressively reduces the network's ability to acquire new tasks as individual bonds become increasingly specialized. We identify an exact algebraic relationship between the polarization of bond dominance and the network's remaining learning capacity, providing a quantitative measure of this loss of adaptability. We then demonstrate that random internal perturbations allow individual bonds to estimate their mechanical dominance and locally adjust their stiffnesses to restore learning capacity. However, the forces associated with active learning bias these local measurements and interfere with restoration. Separating learning and repair into wake-like and sleep-like phases removes this interference and improves continual-learning performance. Our results suggest that the need for offline restorative phases can emerge from the mechanics of physical learning, even in systems without biological components.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-17
DOI
https://doi.org/10.48550/arxiv.2611.22070
Primary Topic
Neural Networks and Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Why do continual physical learners need sleep?

Nidhi Pashine, J. M. Schwarz, Nada Elmeligy
arXiv (Cornell University)
Neural Networks and Applications
preprint

Why do continual physical learners need sleep?

Nidhi Pashine, J. M. Schwarz, Nada Elmeligy
preprint en

Abstract

Sleep is widespread across the animal kingdom, including in organisms without a centralized nervous system. Does learning create a need for sleep, even without biology? To explore this question, we study a disordered spring network trained to perform a sequence of mechanical allosteric tasks through local changes in bond stiffness. We find that continual learning progressively reduces the network's ability to acquire new tasks as individual bonds become increasingly specialized. We identify an exact algebraic relationship between the polarization of bond dominance and the network's remaining learning capacity, providing a quantitative measure of this loss of adaptability. We then demonstrate that random internal perturbations allow individual bonds to estimate their mechanical dominance and locally adjust their stiffnesses to restore learning capacity. However, the forces associated with active learning bias these local measurements and interfere with restoration. Separating learning and repair into wake-like and sleep-like phases removes this interference and improves continual-learning performance. Our results suggest that the need for offline restorative phases can emerge from the mechanics of physical learning, even in systems without biological components.

arXiv (Cornell University)
Neural Networks and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon