AI System Achieves 93% Accuracy in Identifying Astronomical Transients with Minimal Training

A team of researchers from Oxford University, Google Cloud, and Radboud University has made significant progress in developing an artificial intelligence (AI) system that can help identify genuine astronomical transients from imaging artefacts.

The researchers used the Gemini 1.5 Pro AI model to analyze data from three optical surveys: Pan-STARRS, MeerLICHT, and ATLAS. They provided the model with a written guide and 15 annotated examples for each survey, which guided its interpretation of subsequent images and formatting of responses.

The model achieved an average accuracy of 93% in distinguishing between real and bogus candidates across all three surveys. However, it's essential to note that this accuracy is based on retrospective datasets, not live discovery rates. The experiment also demonstrated the importance of expert curation in selecting representative examples for guiding the AI system.

The team tested the effect of reducing the number of guide triplets from 15 to 12, which resulted in a decrease in accuracy by about 0.5 percentage points. This suggests that the chosen number of examples is not universal and may need to be adjusted depending on the specific task and dataset.

The results of this study are promising for future applications in astronomy, but it's crucial to consider the limitations and challenges associated with deploying AI systems in real-world scenarios. For instance, large models remain too slow and expensive to sit at the front of a survey producing millions of alerts each night.