Event

Learning Dynamical Systems: Hybrid Models, Causal Losses, and Uncertainty Quantification

19 July 2024
3:30 pm - 4:00 pm

Location

Room 201

Organizer

Complexity Science Hub
Email
events@csh.ac.at
  • Attendance: in person
  • Language: EN

Event

Learning Dynamical Systems: Hybrid Models, Causal Losses, and Uncertainty Quantification

Talk by Matthew Levine (MIT / Harvard)

The development of data-informed predictive models for dynamical systems is of widespread interest in many disciplines. We present a unifying framework for blending mechanistic and machine-learning approaches to identify dynamical systems from noisily, partially, and irregularly observed data. We will first highlight approximate maximum likelihood approaches that leverage data assimilation to succeed in both synthetic examples (e.g., chaotic Lorenz systems) and real-world biomedical settings (i.e., self-monitoring data collected in a clinical study of Type 1 Diabetes and exercise). We will also introduce novel hybrid/causal losses that incorporate expert knowledge to further improve inference in the real-world diabetes example, which suffers from data sparsity. We will then demonstrate how the above approaches can be cast within a Bayesian framework to quantify uncertainty over learned models; we will also introduce a new Jax package that provides a generic interface for specifying and fitting these models. We will conclude with exciting new questions surrounding identifiability and equivalence classes of dynamical systems in the context of Bayesian inference. Joint work with Andrew Stuart, Youssef Marzouk, Emily Fox, and Iñigo Urteaga.

RSVP

Speaker(s)

0 Pages 0 Press 0 News 0 Events 0 Projects 0 Publications 0 Person 0 Visualisation 0 Art

Signup

CSH Newsletter

Choose your preference
   
Data Protection*