Reliable uncertainty estimates are essential for deploying machine learning systems in settings with distribution shift and limited data, where models must both recognize uncertainty and oftentimes adapt efficiently from few examples. Gradient-based optimization drives modern machine learning, yet most uncertainty quantification (UQ) methods rely only on information at the final trained parameters, ignoring the structure of the training trajectory. We propose a dynamics-centric framework that models training as a stochastic dynamical system. This enables a decomposition into stable and plastic directions and supports propagation of uncertainty through the learned dynamics. Using this idea, we guide few-shot adaptation as a constrained control problem, where update directions and magnitudes are informed by the learned dynamics and associated uncertainty. We demonstrate the performance of our methodology on out-of-distribution classification and few-shot adaptation tasks, such as the Camelyon-17 dataset.

I’m doing SuperUROP to gain experience with carrying out a year-long independent project. This is a really exciting opportunity to study more of the mathematical foundations of ML and contribute to more trustworthy ML models.