人工智能培训

搜索

人工智能教程:深度不平衡学习在人脸识别和属性预测中的应用(Deep Imbalanced Learning for Face Recognition and A

[复制链接]
meidou. 发表于 2018-6-4 08:34:24 | 显示全部楼层 |阅读模式
meidou. 2018-6-4 08:34:24 970 0 显示全部楼层
人工智能教程:深度不平衡学习在人脸识别和属性预测中的应用(Deep Imbalanced Learning for Face Recognition and Attribute Prediction)用于脸部分析的数据通常表现出高度偏斜的类别分布,即大多数数据属于少数多数类别,而少数类别仅包含少量情况。为了缓解这个问题,当代的深度学习方法通​​常遵循经典的策略,例如课堂重新采样或成本敏感的培训。在本文中,我们进行了广泛而系统的实验来验证这些经典方案在类别不平衡数据上的表征学习的有效性。我们进一步证明,通过执行深层网络来维持班级内部和班级之间的集群边际,可以学习到更多有区别的深层表示。这种紧密约束有效地减少了本地数据邻域中固有的类别不平衡,从而在本地雕刻更平衡的类别边界。说明在超球体流形上的群集分布之间部署角边缘是很容易的。这种学习到的基于群集的LargeMargin局部嵌入(CLMLE)与简单的k-最近的群体算法相结合时,与现有方法相比,在显示不平衡类别分布的人脸识别和人脸属性预测任务方面的准确度显着提高。
Data for face analysis often exhibit highly-skewed class distribution, i.e.,most data belong to a few majority classes, while the minority classes onlycontain a scarce amount of instances.To mitigate this issue, contemporary deeplearning methods typically follow classic strategies such as class re-samplingor cost-sensitive training.In this paper, we conduct extensive and systematicexperiments to validate the effectiveness of these classic schemes forrepresentation learning on class-imbalanced data.We further demonstrate thatmore discriminative deep representation can be learned by enforcing a deepnetwork to maintain inter-cluster margins both within and between classes.Thistight constraint effectively reduces the class imbalance inherent in the localdata neighborhood, thus carving much more balanced class boundaries locally.Weshow that it is easy to deploy angular margins between the clusterdistributions on a hypersphere manifold.Such learned Cluster-based LargeMargin Local Embedding (CLMLE), when combined with a simple k-nearest clusteralgorithm, shows significant improvements in accuracy over existing methods onboth face recognition and face attribute prediction tasks that exhibitimbalanced class distribution.人工智能教程:深度不平衡学习在人脸识别和属性预测中的应用(Deep Imbalanced Learning for Face Recognition and Attribute Prediction) TQ4jfiO35mqnCX5V.jpg
URL地址:https://arxiv.org/abs/1806.00194     ----pdf下载地址:http://arxiv.org/pdf/1806.00194    ----人工智能教程:深度不平衡学习在人脸识别和属性预测中的应用(Deep Imbalanced Learning for Face Recognition and Attribute Prediction)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

meidou.当前离线
新手上路

查看:970 | 回复:0

快速回复 返回顶部 返回列表