For their high expressive power and accuracy, machine learning (ML) models are now found in countless application domains. These include autonomous and cyber-physical systems found in high-risk and safety-critical domains, such as healthcare and automotive. These systems nowadays integrate multiple ML components for e.g., sensing, end-to-end control, predictive monitoring, and anomaly detection. Hence, data-driven analysis has become necessary in this context, one where rigorous model-driven techniques like model checking have been the go-to solution for years.
In this project you will develop data-driven analysis techniques for autonomous systems based on conformal prediction (CP) [1,2], an increasingly popular approach to providing guarantees on the generalization error of ML models: it can be applied on top of any supervised learning model and it provides so-called prediction regions (instead of single-point predictions) guaranteed to contain the (unknown) ground truth with a given probability. Crucially, these coverage guarantees are finite-sample (as opposed to asymptotic) and do not rely on any parametric or distributional assumptions.
Our group has a track record of developing CP-based methods for predictive monitoring of autonomous and cyber-physical systems [3-6]. With this project, you will contribute to this endeavor by working on challenge problems including off-policy prediction [7,8], data-driven optimization, causal inference [9,10], robust inference under distribution shifts [11,12], and uncertain distributions [13,14].
The proposed techniques will be evaluated in standard relevant benchmarks and different real-world scenarios coming from the REXASI-PRO EU project [15], which focuses on the safe navigation of autonomous wheelchairs in crowded environments for people with reduced mobility.
The overall goal of this project is to provide correctness guarantees for ML models used in high-stake and safety-critical autonomous and cyber-physical systems, by expanding conformal prediction techniques to tackle novel and challenging problem domains.