Handling Distribution Shift in Visual Learning

Описание: Abstract: While deep learning has achieved remarkable computer vision successes, fundamentally both the theory and practice for these successes have relied on vanilla supervised learning where the training and testing datasets both are sampled from the same distribution. In reality, there is likely to be a significant distribution shift once models are deployed, including noise/weather/illumination/modality changes (covariate shift), new categories (semantic shift), or different label distributions. In this talk, I will present our recent work focusing on the fundamental handling of several of these shifts. For label distribution shifts, we propose a posterior-recalibration of classifiers that can be applied without re-training to handle imbalanced datasets. For covariate and semantic shift, we propose a geometric decoupling of classifiers into feature norms and angles, showing that it can be used to learn more sensitive feature spaces for better calibration and out-of-distribution detection. We demonstrate state-of-art results across multiple benchmark datasets and metrics. In the end, I will present connections to a wider set of problems including continual/lifelong learning, open-set discovery, and semi-supervised learning.