In real-world scenarios, data often exhibits a long-tailed distribution, leading to bias towards head classes due to imbalanced gradients. The Gradient-Aware Logit Adjustment (GALA) loss is introduced to adjust logits based on accumulated gradients for optimization balance. A post hoc prediction re-balancing strategy further mitigates bias towards head classes. Extensive experiments on benchmark datasets show superior performance over existing methods. The GALA loss effectively balances gradient ratios and negative distributions, reducing classifier biases. The prediction re-balance strategy normalizes predictions across classes, addressing biases from biased classifiers or CNNs. Achieving top-1 accuracy improvements on CIFAR100-LT, Places-LT, and iNaturalist datasets validates the effectiveness of the proposed approach.
To Another Language
from source content
arxiv.org
Önemli Bilgiler Şuradan Elde Edildi
by Fan Zhang,We... : arxiv.org 03-15-2024
https://arxiv.org/pdf/2403.09036.pdfDaha Derin Sorular