Interpretable Human-Label-Free Deep Learning for Real-Bogus Classification with Uncertainty Quantification

Researchers have developed a deep learning framework for Real-Bogus classification that operates without human-labeled data, instead utilizing injected transients and bogus-dominated survey data. This approach enables the model to learn from noisy and survey-dependent community labels, reducing the need for reliable but costly labels. The framework incorporates uncertainty quantification, allowing for more accurate predictions and robust performance. By leveraging this method, astronomers can improve the efficiency of automated discovery pipelines, particularly in time-domain surveys where numerous transient candidates are generated. The ability to classify real and bogus signals without relying on human labels has significant implications for astronomical research, enabling scientists to focus on higher-level analysis and decision-making¹. This advancement matters to practitioners as it streamlines the discovery process, freeing up resources for more complex and high-value tasks.