Generative AI has begun to transform music creation, with transformer-based systems able to improvise alongside human performers. However, the inner workings of these models remain difficult to interpret, making it challenging for musicians to understand how past notes or motifs influence the output. This project develops on a real-time visualization system that retrieves, downsamples, and transmits attention weights from a transformer music model, allowing its decision-making to be represented in an interpretable way during live performance. In addition to visualization, the work focuses on detecting repetition in token sequences through attention analysis, generating multiple futures for musicians, and quantifying how ‘surprising’ a note to aid interpretability. While music performance is the primary application, the methods aim to contribute more broadly to real-time interpretability and human–AI collaboration with generative transformers.

I am participating in SuperUROP to gain a more intensive undergraduate research experience. I hope to apply machine learning research to music, a field I have been passionate about my entire life. My goal is to develop tools that make generative AI more interpretable for musicians, with the aim of culminating in a live in-concert demonstration. I also hope to deepen my understanding of the research process, from project proposal to publication.