Core Technology
Foundational methods for reasoning, probabilistic inference, and model-building in complex environments.
People
Matthew Levine
Research Scientist, Cambridge, MA
Matthew Levine develops methods at the intersection of machine learning, dynamical systems, and uncertainty quantification, with a focus on improving prediction and inference in biological and physical systems. His work blends mechanistic modeling, probabilistic inference, and modern machine learning to build principled tools that hold up in real, noisy settings.
About
I am a research scientist at Basis, where I work on principled machine learning methods for complex dynamical systems. My research focuses on improving prediction and inference in biological and physical systems by combining mechanistic modeling, probabilistic reasoning, and data-driven learning.
I am especially interested in approaches that preserve scientific structure while remaining flexible enough to learn from partial, noisy, real-world data. Across projects, I aim to develop theory that is mathematically grounded, software that is practically useful, and applications that matter in domains such as biomedicine, epidemiology, and scientific discovery.
Before joining Basis, I was a postdoctoral fellow at the Broad Institute’s Eric and Wendy Schmidt Center and completed my PhD in Computing and Mathematical Sciences at Caltech, where I was advised by Andrew Stuart. Earlier, I worked in biomedical informatics at Columbia University on predictive modeling and decision support for health applications.
Projects
Current and recent Basis projects.
Collaborative Intelligent Systems
Tools for understanding and reasoning about complex, multi-agent, real-world dynamical systems using multimodal animal behavior data.
Collaborations
Ways to work with Matt at Basis.
Collaborative research projects
I would love to collaborate both with methodologists interested in advancing the state of the art in scientific machine learning, dynamical systems, uncertainty quantification, and data-driven inference, and with scientists who want to bring these methods to bear on important domain problems in areas such as biomedicine, epidemiology, and the atmospheric sciences.
Co-advising postdocs and trainees
I am open to co-advising postdoctoral researchers and other trainees whose work sits between rigorous methodology, real-world scientific applications, and applied machine learning.
In the News
Talks, announcements, and recent highlights.
ICML 2024 oral presentation for Hybrid2 Neural ODE Causal Modeling
Selected for oral presentation at ICML 2024.
Invited Stanford EE talk on principled inference of dynamics in the AI era
Talk on machine learning, uncertainty quantification, and biomedicine.
Software
Tools for dynamical systems and scientific machine learning.
Dynestyx
Dynestyx is an extension of NumPyro for first-class support of dynamical systems, with a unified interface to structured inference methods for state-space models.
CD-Dynamax
JAX software for learning dynamical systems from irregularly sampled, partially observed, noisy time-series data.
Recent Publications
Recent papers and preprints.
Continuum attention for neural operators
Edoardo Calvello, Nikola B. Kovachki, M. Levine, Andrew M. Stuart
Contextual computation by competitive protein dimerization networks
Jacob Parres-Gold, M. Levine, Benjamin Emert, Andrew Stuart, Michael B. Elowitz
Learning about structural errors in models of complex dynamical systems
Jin-Long Wu, M. Levine, Tapio Schneider, Andrew Stuart
Selected Publications
Earlier selected work.
A framework for machine learning of model error in dynamical systems
M. Levine, Andrew M. Stuart
Ensemble Kalman methods with constraints
David J. Albers, Paul-Adrien Blancquart, M. Levine, Elnaz Esmaeilzadeh Seylabi, Andrew Stuart