Rethinking Loss Reweighting for Imbalance Learning as an Inverse Problem: A Neural Collapse Point of View

Researchers from the machine learning community have proposed a novel approach to address class imbalance in long-tailed classification tasks. Published on arXiv, the study critiques existing loss reweighting strategies for their reliance on heuristics and lack of well-defined targets. Drawing inspiration from the concept of Neural Collapse (NC), specifically the ideal simplex Equiangular Tight Frame (ETF) terminal geometry, the authors identify equal per-class average loss as an optimal objective. They reframe loss reweighting as an inverse problem, introducing a dynamic strategy that infers class weights to align with this ideal equal loss target. Empirical evaluations demonstrate that this method effectively reduces the loss imbalance coefficient and achieves closer alignment with NC geometry. Furthermore, the proposed technique consistently outperforms strong baseline models across various datasets. This work contributes to the field of artificial intelligence by providing a theoretically grounded solution for improving model performance in imbalanced data scenarios, moving beyond traditional heuristic methods.