概述
作为现今主流自监督对比学习使用的对比损失(InfoNCE)的源头,感到学习Noise-contrastive Estimation(NCE)的重要性。该方法的原文为:Noise-contrastive estimation: A new estimation principle for unnormalized statistica…
WHAT SHOULD NOT BE CONTRASTIVE IN CONTRASTIVE LEARNING
一、背景介绍:
现在的自监督对比学习方法,大多通过学习对不同的数据增强保持不变,从而学习到一个好的视觉表征。这里就隐含地假设了一组特定地表示不变性,(比如颜色不变性),但是如果下游任务不满足这种假设条…
参考文献: [1] Training Products of Experts by Minimizing Contrastive Divergence [2] Training Restricted Boltzmann Machines using Approximations to the Likelihood Gradient [3]Training Products of Experts by Minimizing Contrastive Divergence阅读笔…
文章目录 IntroductionMethodPredictive LearningContrastive Learning with Two Views距离函数的定义contrastive Learning with more than two viewsImplementation the Contrastive Lossexperiments代码分析Introduction codeing theory 的主要思想是:以自编码器或者生成器…
Contrastive Loss 参考:https://zhuanlan.zhihu.com/p/35027284
对比损失函数,来自Yann LeCun的论文《Dimensionality Reduction by Learning an Invariant Mapping》,出发点:增大类间差异并且减小类内差异。
公式如下ÿ…