NeurIPS 2020 talk: Rethinking the Value of Labels for Improving Class-Imbalanced Learning
Code (data & pretrained models): https://github.com/YyzHarry/imbalanced-semi-self
Project page: https://www.mit.edu/~yuzhe/imbalanced-semi-self.html
We show theoretically and empirically that, both semi-supervised learning (using unlabeled data) and self-supervised pre-training (first pre-train the model with self-supervision) can substantially improve the performance on imbalanced (long-tailed) datasets, regardless of the imbalanceness on labeled/unlabeled data and the base training techniques.
