边距(机器学习)
嵌入
计算机科学
人工智能
模式识别(心理学)
判别式
一般化
特征向量
特征(语言学)
代表(政治)
水准点(测量)
机器学习
数学
哲学
大地测量学
政治
数学分析
语言学
政治学
法学
地理
作者
Yulin Su,Boan Chen,Zi‐Ming Feng,Junchi Yan
标识
DOI:10.1007/978-3-031-44198-1_4
摘要
Visual recognition methods assume models will be evaluated on the same class distribution as training data, but real-world data is often heavily class-imbalanced. To address this, the essential idea is to provide discriminative fitting abilities for classes with different sample sizes, i.e., the model achieves better generalization on less frequent classes, while maintaining high classification ability on the recurring classes. In this work, we propose to unify representation learning and classification learning with robust margin adjustment, which enforces a suitable margin in logit space and regularizes the distribution of embeddings. This procedure reduces representation bias in the feature space and reduces classification bias in the logit space at the same time. We further augment the under-represented tail classes on the feature level via re-balanced sampling from the robust prototype, calibrated with the knowledge from well-represented head classes and adaptive embedding uncertainty estimation. We conduct extensive experiments on a common long-tailed benchmark CIFAR100-LT. Experimental results demonstrate the advantage of the proposed AMDRG for the long-tailed recognition problem.
科研通智能强力驱动
Strongly Powered by AbleSci AI