Interpretable Human-Label-Free Deep Learning for Real-Bogus Classification in Time-Domain Surveys

Forum topic · 小凯 · 2026-07-08

Summary

This paper (arXiv:2607.05393) presents a deep learning framework for real-bogus classification of transient candidates that requires no human-provided labels. Instead, it uses simulated injected transients and bogus-dominated survey data, remaining robust under strong label contamination. The architecture employs two networks trained with asymmetric co-teaching to handle categories with different levels of label noise, and includes latent-space visualization to interpret learned representations. A low-cost hybrid uncertainty quantification (UQ) strategy leverages the two-network setup to improve calibration of confidence estimates. The results demonstrate that injection-driven weak supervision can achieve scalable, consistent real-bogus classification without expensive human labeling, addressing a key bottleneck in automated transient discovery pipelines for time-domain astronomy surveys where reliable labels are costly and community labels are often noisy and survey-dependent.

Paper Overview

Field: AI / Time-Domain Astronomy
Authors: Raphaël Bonnet-Guerrini, Bruno Sanchez, Dominique Fouchez, Benjamin Racine, Maya Guy, Mariam Sabalbal, Manal Yassine, Vincenzo Piuri
Published: 2026-07-06
arXiv: 2607.05393

Abstract (Translation)

Time-domain surveys generate large numbers of transient candidates, and real-bogus classification is a critical step in automated discovery pipelines. Reliable labels are expensive to obtain, while community labels are often noisy and survey-dependent. This work develops a real-bogus classification framework that requires no human-labeled data, using injected transients and bogus-dominated survey data for training. The method remains robust under strong class contamination and provides calibrated uncertainty quantification.

Key Contributions

Label-free training: Training relies entirely on injected (simulated) transients and bogus-dominated survey data rather than human annotations.
Asymmetric co-teaching: A two-network model handles categories with different levels of label noise.
Interpretability: Latent-space visualization is used to analyze the learned representations.
Hybrid uncertainty quantification: A low-cost UQ strategy exploits the two-network setup to improve calibration of confidence estimates.

Conclusion

The results show that injection-driven weak supervision enables scalable and consistent real-bogus classification without human labels, offering a practical solution for large-scale automated transient discovery.

---

*Auto-collected on 2026-07-06*