In June 25, CVPR 2021 launched the 2nd CLVision workshop (IEEE CVPR2021 Workshop on Continual Learning in Computer
Vision, website: https://sites.google.com/view/clvision2021/overview), which discusses
the recent progress, current limitations, and future directions of continual/incremental
learning in computer vision. The workshop received a total of 46 submissions
from MIT, Imperial College London, DeepMind etc., which revolve around tasks,
methods and libraries of continual/incremental learning. Chen He in our lab won
the Best Paper Award as the first author out of the 46 submissions.
Information of the paper：
Chen He, Ruiping Wang, Xilin Chen, “A Tale
of Two CILs: The Connections Between Class Incremental Learning and Class
Imbalanced Learning, and Beyond,” IEEE CVPR 2021 Workshop on Continual Learning
in Computer Vision (CLVision), pp. 3559–3569, June 19-25, 2021.
This paper systematically shows the
connections between Class Incremental Learning and Class Imbalanced Learning.
Specifically, it demonstrates that many techniques in Class Incremental
Learning share similar ideas in Class Imbalanced Learning. By introducing a
simple post-scaling technique that originates in Class Imbalanced Learning, the
performance is on par or even higher than SOTAs in Class Incremental Learning. By
leveraging visualization tools, the paper finds that post-scaling and another
effective “weight aligning” technique translates and rotates the decision
boundary to alleviate the biasing problem. Based on the theoretical analyses
and experimental results mentioned above, the paper reflects upon the recent
progress of Class Incremental Learning and raises the question that Class
Incremental Learning seems to degenerate into Class Imbalanced Learning. It
further provides the authors’ preliminary thoughts on the future directions of
this field.
Chen He participated in the workshop and
performed a 10-minute oral presentation about the paper to over 100 researchers.
Fig.
1. Certificate of the Best Paper Award
Fig.
2. Oral presentation by Chen He
Download: