Selected publications:

  • Architecture diagram for Latent Domain Modeling Improves Robustness to Geographic Shifts

    Latent Domain Modeling Improves Robustness to Geographic Shifts

    Ruth Crasto, Esther Rolf

    CVPRF 2026

    We propose a framework for improving the performance of image understanding models trained on data with geographically varying subpopulations (e.g. different countries). The key idea is to model continuous, latent domain assignment using geographic location encoders and to condition the main task predictor on the jointly-trained latents.

  • Computation graph for ICML'23 workshop paper

    Statistics estimation in neural network training: a recursive identification approach

    Ruth Crasto, Xuchan Bao, Roger Grosse

    ICML 2023 Workshop on New Frontiers in Learning, Control, and Dynamical Systems

    We propose a method for adapting the decay term in an EMA (e.g. Adam optimizer) that leads to improved training convergence. Our method is inspired by ideas from the field of recursive identification.

A full list of publications can be found on Google Scholar.