Distribution Shift
Distribution shift, the discrepancy between training and deployment data distributions, is a critical challenge in machine learning, hindering model generalization and reliability. Current research focuses on developing methods to detect, adapt to, and mitigate the impact of various shift types (e.g., covariate, concept, label, and performative shifts), employing techniques like data augmentation, model retraining with regularization, and adaptive normalization. These advancements are crucial for improving the robustness and trustworthiness of machine learning models across diverse real-world applications, particularly in safety-critical domains like healthcare and autonomous driving, where unexpected performance degradation can have significant consequences.
Papers
GraphMETRO: Mitigating Complex Graph Distribution Shifts via Mixture of Aligned Experts
Shirley Wu, Kaidi Cao, Bruno Ribeiro, James Zou, Jure Leskovec
Adversarial Learning for Feature Shift Detection and Correction
Miriam Barrabes, Daniel Mas Montserrat, Margarita Geleta, Xavier Giro-i-Nieto, Alexander G. Ioannidis
Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation
Haojie Zhang, Yongyi Su, Xun Xu, Kui Jia
Complementary Benefits of Contrastive Learning and Self-Training Under Distribution Shift
Saurabh Garg, Amrith Setlur, Zachary Chase Lipton, Sivaraman Balakrishnan, Virginia Smith, Aditi Raghunathan