Title

图像分类在在线持续学习的研究

Abstract

    Online continual learning for image classification studies the problem of learning to classify images from an online stream of data and tasks, where tasks may include **new classes (class incremental)** or **data nonstationarity (domain incremental)**. One of the key challenges of continual learning is to avoid catastrophic forgetting (CF), i.e., forgetting old tasks in the presence of more recent tasks. Over the past few years, a large range of methods and tricks have been introduced to address the continual learning problem, but many have not been fairly and systematically compared under a variety of realistic and practical settings.

任务：new classes（类增量）、data nonstrationarity（域增量） 持续学习的关键：避免灾难性遗忘（在出现较新任务时忘记旧的任务）

   To better understand the relative advantages of various approaches and the settings where they work best, this survey aims to (1) compare state-of-the-art methods such as **Maximally Interfered Retrieval (MIR), iCaRL, and GDumb** (a very strong baseline) and determine which works best at different memory and data settings as well as better understand the key source of CF; (2) determine if the best online class incremental methods are also competitive in the domain incremental setting; and (3) evaluate the performance of 7 simple but effective tricks such as the “review” trick and the nearest class mean (NCM) classifier to assess their relative impact.

• 比较3种方法，确定哪种在不同内存和数据设置下最有效：最大干扰检索（Maximally Interfered Retrieval, MIR），iCaRL，GDumb

当memory shuffle很小时，iCaRL有竞争力 在medium-size数据集上，GDumb优于很多提出的方法 在large-scale数据集上，MIR表现最好

• 确定online class incremental方法在domain方法中是否有用

GDumb表现糟糕 MIR在这种持续学习下具有很强的竞争力

• 评估7个简单但有效的技巧的性能："review" trick，最近类均值分类器NCM

所有技巧都有益 当用"review" trick和NCM时，MIR的性能水平使online continual learning更接近其匹配离线训练的目标

1. Introduction

传统深度学习关注离线训练，每个mini-batch都是从静态数据集中进行iid采样 缺点：为适应数据分布的变化，需要在新的数据集上完全重新训练网络 → 低效，数据由于存储限制或隐私不可用

持续学习关注从non-iid数据中学习，保存和拓展所获得的知识 保存过去的知识，对新知识的快速适应 → 在稳定性和可塑性之间寻找平衡

先前的持续学习：多头评估（为每个任务分配一个单独的输出层） 缺点：需要额外的监督信号（task label）；所有数据存储在内存中进行训练

In this work：

提出了（1）Online Class Incremental, OCI 在线类增量；（2）Online Domain Incremental, ODI 在线域增量 和任务增量设置的区别：（1）采用single-head单头评价 （2）模型在线处理：multiple epoches和repear shuffles对于确保iid条件不可行

Contribution：

1. 比较了OCI最先进的方法，并确定那种方法在不同的内存和数据设置下最有效

当memory shuffle很小时，iCaRL有效 当mediun-size dataset，GDumb是strong baseline 当larger-scale dataset，MIR表现最好 灾难性遗忘的关键原因：由于先前数据和新数据之间的不平衡，导致最近学习偏向最后一个fc中的新类