Uncertainty Estimation

Quantile-based and distributional methods for calibrated, multivariate uncertainty.

A point estimate says little about what the data actually support. I develop methods for multivariate conditional distributions, in particular vector quantile regression built on optimal transport, that scale to large datasets and extend to manifolds. These give calibrated uncertainty that can inform downstream decisions, including what to measure next.

Selected work

* Equally-contributing first author, † Equally-contributing senior author

2024

  1. pegoraro2023vector.png
    AISTATS2024
    Vector Quantile Regression on Manifolds
    M. Pegoraro, S. Vedula, A. Rosenberg, I. Tallini, E. Rodola, A. Bronstein
    Proc. International Conference on Artificial Intelligence and Statistics (AISTATS)

2023

  1. vedula2023continuous.png
    ICML Workshop2023
    Continuous Vector Quantile Regression
    S. Vedula*, I. Tallini*, A. Rosenberg, M. Pegoraro, E. Rodola, Y. Romano, A. Bronstein
    ICML Workshop on New Frontiers in Learning, Control, and Dynamical Systems
  2. rosenberg2023fast.png
    ICLR2023
    Fast Nonlinear Vector Quantile Regression
    A. Rosenberg*, S. Vedula*, Y. Romano, A. Bronstein
    Proc. International Conference on Learning Representations (ICLR)
    HTML

← All research areas