BONSAI (Bilevel Optimization for Simulation-based Inference) is an ANR JCJC project coordinated by Michael Arbel. The ANR project record lists a March 2024 start and a 48-month duration, corresponding to February 2028.

Project description

Despite impressive progress in machine learning, it remains challenging to systematically and reliably use learning methods in new disciplines to accelerate scientific discovery. A central obstacle is the lack of a consistent way to incorporate expert knowledge, such as simulators, into modern learning frameworks. BONSAI addresses this challenge from theory to applications by combining simulations with deep learning through bilevel optimization for inference.

BONSAI develops methods for inference problems in which simulators encode valuable scientific knowledge but cannot be directly combined with modern learning systems. The project brings together theory, optimization, and applications to study how bilevel optimization can make that connection practical and reliable.

Research questions

The research agenda connects three questions:

  • How can a learned approximation to an intractable simulator quantity be optimized jointly with the scientific inference task, rather than treated as a fixed surrogate?
  • What statistical guarantees and efficient gradient methods are possible when the inner problem uses an expressive learned model and only finitely many simulations are available?
  • How can inference remain reliable when a simulator only approximates the real system, including the calibration of uncertainty from experimental data?

Members and partners

Scientific leader

Permanent researchers

PhD students

Former PhD participants

External present and past collaborations