Google Research· ResearchStory of the weekResearch paper

Teaching Gemini to spot exploding stars with just a few examples

Google Research, in collaboration with the University of Oxford, published a paper in Nature Astronomy demonstrating that Google's Gemini model, using few-shot learning with only 15 annotated examples per survey, can classify cosmic transients (e.g., supernovae) with 93% accuracy across three datasets (Pan-STARRS, MeerLICHT, ATLAS), while also providing textual explanations and interest scores. The model matches specialized CNNs but offers transparency, and a panel of astronomers confirmed the coherence of its outputs.

From the source

In a publication in Nature Astronomy, we show how Google's Gemini model can be transformed into an expert astronomy assistant that classifies cosmic events with high accuracy and explains its reasoning in plain language, achieving 93% accuracy across three datasets by learning from just 15 annotated examples per survey.

research.google