Even small correlation and diversity shifts pose dataset-bias issues

Autor(es)

Bissoto, Alceu ; Barata, Catarina ; Valle, Eduardo ; Avila, Sandra

• Tipo de documento: Artigo Científico
• Fonte: PATTERN RECOGNITION LETTERS; v. 179, p. 7-pg., 2024-02-07.

Resumo

Distribution shifts hinder the deployment of deep learning in real-world problems. Distribution shifts appear when train and test data come from different sources, which commonly happens in practice. Despite shifts occurring concurrently in many forms (e.g., correlation and diversity shifts) and intensities, the literature focuses only on severe and isolated shifts. In this work, we propose a comprehensive examination of distribution shifts across different intensity levels, investigating the nuanced impacts of both mild and severe shifts on the learning process and assessing the interplay between correlation and diversity shifts. We train models in three different scenarios considering synthetic and real correlation and diversity shifts, spamming across eight different levels of correlation shift, and evaluate them in both in-distribution and diversity-shifted test sets. Our experiments reveal three major findings: (1) Even small correlation shifts pose dataset-bias issues, presenting a risk of accumulating and combining unaccountable weak biases; (2) Models learn robust features in high- and low-shift scenarios but prefer spurious ones during test regardless; (3) Diversity shift can attenuate the reliance on spurious correlations. Our work has implications for distribution shift research and practice, providing new insights into how models learn and rely on spurious correlations under different types and intensities of shifts. (AU)

• Processo FAPESP: 13/08293-7 - CECC - Centro de Engenharia e Ciências Computacionais
• Beneficiário: Munir Salomao Skaf
• Modalidade de apoio: Auxílio à Pesquisa - Centros de Pesquisa, Inovação e Difusão - CEPIDs
• Processo FAPESP: 19/19619-7 - Geração ilimitada de imagens de lesões de pele usando redes generativas adversariais
• Beneficiário: Alceu Emanuel Bissoto
• Modalidade de apoio: Bolsas no Brasil - Doutorado
• Processo FAPESP: 22/09606-8 - Compreensão do papel de atalhos e mudanças de distribuição para generalização de redes neurais profundas
• Beneficiário: Alceu Emanuel Bissoto
• Modalidade de apoio: Bolsas no Exterior - Estágio de Pesquisa - Doutorado
• Processo FAPESP: 20/09838-0 - BI0S - Brazilian Institute of Data Science
• Beneficiário: João Marcos Travassos Romano
• Modalidade de apoio: Auxílio à Pesquisa - Programa Centros de Pesquisa em Engenharia