EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Efficient Low-rank Bayesian Neural Networks for Incremental Pulsar Candidate Identification

Yi Liu, Suxun Zhu, Jing Jin, Hongyang Zhao, Taofei Jiang, Jie Liu

The Astrophysical Journal Supplement Series · 01.08.2026 · 285 (2) · 56

DOI: 10.3847/1538-4365/ae809c ↗

Kurz gefasst

Fachartikel in der Zeitschrift The Astrophysical Journal Supplement Series, online erschienen am 28.07.2026. Verfasst von Yi Liu und 5 weiteren Autor:innen. Die Arbeit nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: MOST ∣ National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the deployment of next-generation radio telescopes such as the Five Hundred Meter Aperture Spherical Radio Telescope (FAST) and the Square Kilometre Array (SKA), pulsar candidate identification faces two critical hurdles: data distribution discrepancies across facilities and catastrophic forgetting in traditional neural networks. While Bayesian neural networks (BNNs) offer a theoretical solution via probabilistic modeling, their prohibitive computational cost hinders deployment in data-intensive astronomical surveys. This paper bridges the gap between theoretical BNN capability and engineering practicality by proposing an improved incremental learning framework with multimodal fusion. Crucially, to transition BNNs from “theoretically feasible” to “operationally viable” for the SKA era, we introduce the Low-Rank plus Diagonal Gaussian Approximation (LRDGA) method. This technique optimizes variational inference, reducing computational complexity from O ( d 2 ) to O ( dr ) and transforming the resource-heavy posterior updates into an efficient process. The model treats diagnostic plots as independent modalities, employing Bayesian convolutional neural networks with a hybrid focal loss/Kullback–Leibler divergence objective to handle imbalance. Experiments on FAST and HTRU datasets demonstrate that in domain incremental scenarios, the method achieves a precision of 97.54%, recall of 96.06%, and F1-score of 97.05%. Notably, LRDGA improves training efficiency by 60% and reduces memory consumption by 72%, effectively resolving the computational bottleneck of BNNs and providing a scalable, automated solution essential for the exascale data challenges of future pulsar searches.

Förderung

  • MOST ∣ National Natural Science Foundation of China

Zitieren

APA

Liu, Y., Zhu, S., Jin, J., Zhao, H., Jiang, T., & Liu, J. (2026). Efficient Low-rank Bayesian Neural Networks for Incremental Pulsar Candidate Identification. The Astrophysical Journal Supplement Series, 285(2), 56. https://doi.org/10.3847/1538-4365/ae809c

Vancouver

Liu Y, Zhu S, Jin J, Zhao H, Jiang T, Liu J. Efficient Low-rank Bayesian Neural Networks for Incremental Pulsar Candidate Identification. The Astrophysical Journal Supplement Series. 2026;285(2):56. doi:10.3847/1538-4365/ae809c

BibTeX
@article{Liu2026,
  author = {Liu, Yi and Zhu, Suxun and Jin, Jing and Zhao, Hongyang and Jiang, Taofei and Liu, Jie},
  title = {Efficient Low-rank Bayesian Neural Networks for Incremental Pulsar Candidate Identification},
  journal = {The Astrophysical Journal Supplement Series},
  year = {2026},
  volume = {285},
  number = {2},
  pages = {56},
  doi = {10.3847/1538-4365/ae809c}
}
Kostenlos starten →