Large Scale Incremental Learning

About

Modern machine learning suffers from catastrophic forgetting when learning new classes incrementally. The performance dramatically degrades due to the missing data of old classes. Incremental learning methods have been proposed to retain the knowledge acquired from the old classes, by using knowledge distilling and keeping a few exemplars from the old classes. However, these methods struggle to scale up to a large number of classes. We believe this is because of the combination of two factors: (a) the data imbalance between the old and new classes, and (b) the increasing number of visually similar classes. Distinguishing between an increasing number of visually similar classes is particularly challenging, when the training data is unbalanced. We propose a simple and effective method to address this data imbalance issue. We found that the last fully connected layer has a strong bias towards the new classes, and this bias can be corrected by a linear model. With two bias parameters, our method performs remarkably well on two large datasets: ImageNet (1000 classes) and MS-Celeb-1M (10000 classes), outperforming the state-of-the-art algorithms by 11.1% and 13.2% respectively.

Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, Yun Fu · 2019

Related benchmarks

Task	Dataset	Result
Class-incremental learning	CIFAR-100	Averaged Incremental Accuracy72.83	281
Class-incremental learning	CIFAR-100	Average Accuracy—	150
Class-incremental learning	ImageNet-R	Last Accuracy—	147
Class-incremental learning	CIFAR100 (test)	Avg Acc69.96	116
Class-incremental learning	ImageNet-100	Last Acc—	108
Class-incremental learning	CIFAR-100 10 (test)	Average Top-1 Accuracy57.8	105
Class-incremental learning	CIFAR-100 (10-split)	Accuracy68.8	87
Incremental Learning	TinyImageNet	Avg Incremental Accuracy40.56	83
Image Classification	MNIST 5 tasks (test)	Accuracy94.8	75
Class-incremental learning	CIFAR-100 Split (test)	Avg Acc81.42	75