Manifold Capacity Regularization for State Representation Learning and Self-Supervised Learning
Nordic Machine Intelligence · 19.05.2026 · 6 (1) · 1-18
DOI: 10.5617/nmi.13235 ↗Kurz gefasst
Fachartikel in der Zeitschrift Nordic Machine Intelligence, online erschienen am 19.05.2026. Verfasst von Li Meng und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The expanding research on manifold-based Self-Supervised Learning (SSL) is based on the manifold hypothesis, which suggests that the inherent complexity of high-dimensional data can be unraveled through lower-dimensional manifold embeddings. Taking advantage of this, manifold-based State Representation Learning (SRL) has emerged as a powerful tool and yielded impressive results for state representations in Reinforcement Learning. Meanwhile, it remains unexplored whether the manifold capacity theory can lead to a general regularization method to improve the performance of existing SRL and SSL methods. Bridging this gap, we present a simple yet effective manifold encoding framework named "C+" that improves representational diversity at the batch level. In experimentation with the Atari Annotated RAM Interface, our method DIM-C+ improves DIM-UA significantly with the same number of target encoding dimensions. The mean F1 score averaged over categories is 78% compared to 75% of DIM-UA. There are also compelling gains across datasets when implementing SimCLR and Barlow Twins with C+. This substantiates our claim that the C+ framework represents an effective regularization alternative, allowing for more consistent high-dimensional data representations.
Zitieren
APA
Meng, L., Goodwin, M., Yazidi, A., & Engelstad, P. (2026). Manifold Capacity Regularization for State Representation Learning and Self-Supervised Learning. Nordic Machine Intelligence, 6(1), 1-18. https://doi.org/10.5617/nmi.13235
Vancouver
Meng L, Goodwin M, Yazidi A, Engelstad P. Manifold Capacity Regularization for State Representation Learning and Self-Supervised Learning. Nordic Machine Intelligence. 2026;6(1):1-18. doi:10.5617/nmi.13235
BibTeX
@article{Meng2026,
author = {Meng, Li and Goodwin, Morten and Yazidi, Anis and Engelstad, Paal},
title = {Manifold Capacity Regularization for State Representation Learning and Self-Supervised Learning},
journal = {Nordic Machine Intelligence},
year = {2026},
volume = {6},
number = {1},
pages = {1--18},
doi = {10.5617/nmi.13235}
}
EUVIMED