核心概念
ConR proposes a contrastive regularizer to address deep imbalanced regression by modeling global and local label similarities in feature space.
摘要
Imbalanced data poses challenges in generalizing models, especially in regression tasks. ConR introduces a contrastive regularizer to prevent features of minority samples from collapsing into majority ones. It effectively translates label relationships to the feature space, boosting performance on deep imbalanced regression benchmarks. The method is efficient, orthogonal to existing approaches, and extends seamlessly to uni- and multi-dimensional label spaces.
統計資料
Our code is publicly available at https://github.com/BorealisAI/ConR.
ConR significantly boosts the performance of all state-of-the-art methods on four large-scale deep imbalanced regression benchmarks.
引述
"ConR addresses the continuous nature of label space with two main strategies in a contrastive manner."
"ConR consolidates essential considerations into a generic, easy-to-integrate, and efficient method."