Robust Feature Learning and Global Variance-Driven Classifier Alignment for Long-Tail Class Incremental Learning
2 November 2023
Computer Vision and Pattern Recognition
Proposes a two-stage framework for long-tail class incremental learning
Learns robust features in stage 1 using mixup regularization
Aligns classifiers in stage 2 based on global variance measure
Avoids need for balanced data or added model layers
Shows consistent gains over prior state-of-the-art on CIFAR and ImageNet
Learning to classify images incrementally from imbalanced data
This paper introduces a two-stage approach to enable image classification models to continuously learn new classes from imbalanced long-tail distributions of data. The approach mitigates catastrophic forgetting and bias towards majority classes by learning robust features in stage 1, then aligning classifiers using global variance and prototypes in stage 2.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper