增量学习——《Insights from the Future for Continual Learning》——arxiv202006

Abstract

提出了一个持续学习的新情景，prescient continual learning（测试模型不仅past classes and current classes，还需要考虑future classes）。基于zero-shot learning的启发，提出了Ghost Model，表征空间的生成模型与损失函数的小心微调

Introduction

the future classes(no training samples)，作者认为这个setting需要让模型know the classes and have some prior information about them。类似的考虑future classes的工作《Selfless sequential learning，ICLR2019》; 《Automatically discovering and learning new visual categories with ranking statistics,ICLR2020》提出一个setting,所有类的训练样本一开始就能全部得到，但标签是增量式地获取到的。而prescient continual learning setting与之相反，标签一开始能全部得到，样本是增量地获取到的。举了一个实际应用的例子，fashion product classification。然后介绍方法的insight，integrate the continual and zero-shot learning seamlessly
Zero-shot learning借鉴的两篇文章《Learning to detect unseen object classes by
between-class attribute transfer,CVPR2009》《Zero-shot learning—a comprehensive
evaluation of the good, the bad and the ugly, TPAMI2019》

Method

ghost model是什么？三个部件(feature extractor, feature generator, classifier)，核心就是feature generator的设计，the generator learns the distribution of the features for all classes, aiming to generate plausible samples of features for the future classes.
For the feature generation to work, we must have exploitable prior information about the classes, more precisely, we must be able to map the class labels c into a class attribute space that makes semantic sense.
The exact way to perform that mapping will be data-dependent, but most often, either we will have an explicit set of attributes linked to each class (color, size, material, provenance, etc.), or we will be able to extract a latent semantic vector, using a technique like Word2vec [30, 32].
类别c与一个属性特征关联，对于feature generator，给定类别c的属性特征，输出类别c的样本特征，要与当前模型的特征逼近。

Conclusion

a new setting, trains on a sequence of tasks, each introducing new classes, but has access to prior information about the classes.

Key points:prescient continual learning setting, 需要知道过去的类、当前出现类、未来出现的类具体是什么这些prior information;核心就是generate a representation for the future classes，感觉有点ill-defined;方法设计地有点easy，直接将zero-shot拿过来用；工程实现起来的复杂程度，在复杂数据集比较难work