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

Imagine you're trying to identify real astronomical events from a flood of data, but getting reliable labels is tough and expensive. Traditionally, researchers rely on human labels, which can be inconsistent and vary from one survey to another. This leads to problems like misclassification, where real events get mixed up with false ones, making it hard to trust the results. This issue is known as label noise, and it can severely impact the accuracy of any classification system you build.

What this paper does is propose a clever way to tackle these challenges without needing those costly human labels. Instead of relying on them, the authors use a combination of simulated data and existing noisy survey data to train a dual-network model. This model is designed to handle different levels of label noise effectively, which means it can still perform well even when the data is messy. They also introduce a method for quantifying uncertainty in their predictions, which helps users understand how confident they can be in the model's classifications.

The results are promising: the method shows strong performance in identifying real versus bogus transients and maintains stability even when faced with significant label noise. This is a big step forward compared to previous methods that struggled with similar issues. For anyone building systems in astronomy or related fields, this approach could save time and resources by allowing for scalable classification without the need for extensive human labeling.

The paper introduces a novel framework for Real-Bogus classification that leverages weak supervision and uncertainty quantification.

The approach is validated on benchmark datasets and includes a robust evaluation of uncertainty quantification methods.