1 iCaRL: Incremental Classifier and Representation Learning (cvpr 2017)

目前来看应该是CIL的（AI阶段）开山作，里边提到了许多CIL相关实验研究的配置和标准。

Abstract：

引出问题，介绍iCaRL，“iCaRL learns strong classifiers and a data representation simultaneously.”。

1 introduction：

· 定义增量学习：Formally, we demand the following three properties of an algorithm to qualify as class-incremental:

i) it should be trainable from a stream of data in which examples of different classes occur at different times,

ii) it should at any time provide a competitive multi-class classifier for the classes observed so far, # 意思就是任何时候你的模型对旧类和新类都要保持很高的准确率；

iii) its computational requirements and memory footprint should remain bounded, or at least grow very slowly, with respect to the number of classes seen so far.

 

· 文中介绍目前满足以上三点的方法有，但是都是过去的基于“fixed data representation”，无法拓展至深度学习可以“learn classifiers and feature representations at the same time”，所以在准确率方面不再具有优势。（详见Sect 3 - related work）

 

· 引出iCaRL：“In this work, we introduce iCaRL (incremental classifier and representation learning), a practical strategy for simultaneously learning classifiers and a feature representation in the class-incremental setting.” 针对上述的三点要求提出了3个模块来解决现有方法存在的shortcomings：

• classification by a nearest-mean-of-exemplars rule,
• prioritized exemplar selection based on herding,
• representation learning using knowledge distillation and prototype rehearsal
 

2 Method

2.1. Class-Incremental Classifier Learning

· Algorithm 1 介绍了分类的方法：针对每个类计算prototype，再衡量一个新数据和这些prototype的距离，离的最近的视为同一类。

· Algorithm 2 介绍了训练的流程：更新参数，限制每个旧类别保留的example的数量，构建当前轮次新类别的examplar set。中间用到了Algorithm 3，5，4模块。

· 文中提到：“All feature vectors are L2-normalized, and the results of any operation on feature vectors, e.g. averages, are also re-normalized”，为了后边方便计算prototype相似度。

· 并且在iCaRL中，“iCaRL uses the network only for representation learning, not for the actual classification step”，意思iCaRL中的网络它只是用于学习更好的特征表示，并不和分类层结合直接用于分类。（特征提取器部分参数量不变，“分类层”对应的参数会随着类别增加增加。

 

2.2. Nearest-Mean-of-Exemplars Classification

 

2.3. Representation Learning
Algorithm 3 介绍了更新参数的流程，这里我有一点疑问：相比金标准，network outputs用于loss就算有什么优势？？？（虽然用金标准肯定是炸了，因为examplar的数量太少了）

这里和finetuning很像，也可以end-to-end正常训练。为了catastrophic forgetting，这里做了两点修改：1）“the training set is augmented”，除了新数据，还有保存的旧类数据（问题同上，这么少怎么起到作用？？？）.2）除了分类loss，还用到了distillation loss，我怎么感觉这里就是简单地over sample了一下old data。。。？？？

2.4. Exemplar Management

· Two routines shown below，分别用于管理当前新来的类和旧类。新类别构建examplars set，旧类要删掉一些由于budget的原因。

· 新类构建examplars set时候，力争用一些emamplars的特征去逼近所有case的特征平均值。最后会提取出a prioritized list。

3 Related work

“iCaRL builds on the insights of multiple earlier attempts to address class-incremental learning.”

主要从两点阐述：“fixed data representations” 和 “also learn the data representation”。

3.1 Learning with a fixed data representation.

· “NCM has been shown to work well and be more robust than standard parametric classifiers in an incremental learning setting [23, 24, 31]”