Timezone: »

Robustness via Uncertainty-aware Cycle Consistency
Uddeshya Upadhyay · Yanbei Chen · Zeynep Akata

Tue Dec 07 08:30 AM -- 10:00 AM (PST) @

Unpaired image-to-image translation refers to learning inter-image-domain mapping without corresponding image pairs. Existing methods learn deterministic mappings without explicitly modelling the robustness to outliers or predictive uncertainty, leading to performance degradation when encountering unseen perturbations at test time. To address this, we propose a novel probabilistic method based on Uncertainty-aware Generalized Adaptive Cycle Consistency (UGAC), which models the per-pixel residual by generalized Gaussian distribution, capable of modelling heavy-tailed distributions. We compare our model with a wide variety of state-of-the-art methods on various challenging tasks including unpaired image translation of natural images spanning autonomous driving, maps, facades, and also in the medical imaging domain consisting of MRI. Experimental results demonstrate that our method exhibits stronger robustness towards unseen perturbations in test data. Code is released here: https://github.com/ExplainableML/UncertaintyAwareCycleConsistency.

Author Information

Uddeshya Upadhyay (University of Tuebingen)
Yanbei Chen (University of Tübingen)

I am currently a postdoctoral researcher within the Cluster of Excellence Machine Learning at the University of Tübingen in Germany, working with Prof. Zeynep Akata at the Explainable Machine Learning group. I obtained my Ph.D. from the Queen Mary University of London in the UK, advised by Prof. Shaogang Gong. My research interests lie at the intersection of deep learning and computer vision. I've focused on developing semi-supervised, unsupervised, and cross-domain deep learning algorithms and techniques for visual learning, with the ultimate goal to advance the automatic exploitation of large-scale visual data using minimal human supervision. I am also dedicated to designing multimodal learning algorithms that could connect, correlate, and integrate multiple data modalities (e.g. vision, language, audio) in an explainable way to build intelligent perception systems.

Zeynep Akata (Max Planck Institute for Informatics)

More from the Same Authors