Understanding the role of shortcuts and distribution shifts in deep learning generalization

• Grant number: 22/09606-8
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: October 27, 2022
• End date: April 26, 2023
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Sandra Eliza Fontes de Avila
• Grantee: Alceu Emanuel Bissoto
• Supervisor: Ana Catarina Fidalgo Barata
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Institution abroad: Universidade de Lisboa, Portugal
• Associated to the scholarship: 19/19619-7 - Generating unlimited skin lesion images with generative adversarial networks, BP.DR

Abstract

Deep learning achieved amazing success in computer vision, and natural language processing. Spurious correlations (biases) in data are able to poison models, causing lack of generalization to out-of-distribution data, which is an obstacle for the application of deep learning to real-world problems as medical imaging. Despite the existence of robust learning algorithms, their performance highly depends on hyperparameter selection, and on the characteristics of the data and its biases. Also, solutions are still dependent on the annotation of spurious sources, and the phenomenons surrounding out-of-distribution generalization is still not near to be understood. Also, despite years of research on domain generalization, evaluation of robust solutions only show minor improvements over consolidated baselines. In this scenario, we study the interaction between two important phenomenons: shortcut learning and domain shift. Shortcuts appear as spurious correlations on training data, and domain shift appears as out-of-distribution test sets. We think that by understanding how base and robust models are affected by these phenomenons, we can design better generalization evaluation methods, debiasing techniques, and better understand the learning mechanism in deep neural networks. (AU)