Bayesian Online Test Time Adaptation

D. Corrales Alonso, D. Ríos Insua

Standard supervised machine learning algorithms typically assume identically independent distributed samples in both training and deployment. Furthermore, these algorithms do not typically leverage learning from unlabeled data, effectively wasting task-relevant information. In the standard test-time adaptation literature, machine learning models are first trained using a labeled source dataset and then adapted in deployment using only the pre-trained model and the incoming unlabeled data. This work addresses the issue of adaptation in scenarios of test-time distributional shifts by designing a fully probabilistic framework that generalises most common approaches in the literature, and, importantly, provides uncertainty estimates on predictions. We exemplify the applicability of this methodology with a set of use cases, growing in complexity in modelling and data dynamics, thus proving the usefulness and scalability of the framework.

Keywords: Test-time-adaptation, online-learning, Bayesian, covariate-shift

GT Inferencia Bayesiana: Sesión de Jóvenes Bayesianos en honor a Mª Eugenia Castellanos
September 5, 2026  10:00 AM
Aula 20

P. García Arce, R. Naveiro, D. Ríos Insua

M. Chacón Falcón, D. Rios Insua

C. Mulet, G. García-Donato