Rico Zhu

I am an undergraduate student at Duke University, double majoring in Computer Science and Math, currently working under Prof. Cynthia Rudin. My research is broadly focused on interpretable generative models, previously in the domain of music composition, and currently in the domain of physics discovery.

I am fortunate to also work with Prof. Simon Mak as part of the JETSCAPE high-energy physics (HEP) collaboration, Prof. Yue Jiang as part of the Interpretable ML Lab, and Prof. Rong Ge. My given name is Yifan Zhu, but I have gone by Rico since about when I was 10, and is my preferred name.

Email  /  CV  /  Google Scholar  /  Github

Research Interests

My interests broadly span applications of interpretable ML for science discovery, with an emphasis in the high-energy physics (HEP) domain. I am currently interested in using methods from geometric deep learning to design more interpretable, and more realistic generative models for scientific emulation.

Publications

Current music harmonization models fail to compose with affect, and those which do are uninterpretable. We propose a novel affective embedding and sentiment representation in tandem with an efficient, interpretable generative model.

Music theory is well understood under the framework of Schenkerian Analysis, a hierarchical approach to harmonic analysis. We propose a novel graph-based formulation of this task.

A novel generative framework for music composition that is architecturally compatible with music theory, making the generation process easily interpretable for music experts.

Demos

How can we generate music affectively? In this demo set in the city of New Orleans, we present a music generation model which uses mixtures to compose with emotion.

Research Experience

Developing gauge invariant generative model for emulating particle collision events. Designed a general, modular software framework for performing closure tests to verify experiment results.

Interpretable generative models for music composition, with a focus on designing models with built-in interpretability under the Schenkerian music theory framework.

With the ATLAS Collaboration, worked on using GNNs to model collider geometry for jet reconstruction as an alternative to the costly Particle Flow algorithm.

Teaching Experience

Souce code and website design from Jon Barron.