Predictive Consistency Learning for Long-Tailed Recognition
Nan Kang (Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS)),* Hong Chang (Chinese Academy of Sciences), Bingpeng MA (University of Chinese Academy of Sciences), Shutao Bai (Institute of Computing Technology, Chinese Academy of Sciences), Shiguang Shan (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
The 34th British Machine Vision Conference
Abstract
Real-world data often exhibit long-tailed distributions, on which modern deep networks often make skewed predictions. Post-hoc correction approaches tackle this problem by introducing class-dependent correction biases to adjust the posterior distribution $\hat{p}_s(y|x)$, thereby compensate the discrepancy between training distribution $p_s(y)$ and test distribution $p_t(y)$. Most works along this line focus on the design of correction bias, but little attention has been paid to the estimation of $\hat{p}_s(y|x)$ which is fairly crucial for post-hoc approaches. In this paper, we highlight the inaccurate estimation of $\hat{p}_s(y|x)$ learned through cross-entropy loss minimization, which produces poorly calibrated predictions and limits the effectiveness of post-hoc correction, particularly under large label distribution shifts. To this end, we propose Predictive Consistency Learning (PCL) for long-tailed learning that learns to maintain consistency between current predictions and the aggregation of historical predictions, which iteratively refine $\hat{p}_s(y|x)$ to improve the post-hoc correction. In large-scale dataset, the storage of historical predictions requires high space complexity. To address this issue while maintaining similar performance, we further propose the compressed PCL (ComPCL) that reduces the space complexity of storing historical predictions to linear by label compression and debias operations. Experiments demonstrate that our method achieves significant improvements on several long-tailed recognition benchmarks.Video
Citation
@inproceedings{Kang_2023_BMVC,
author = {Nan Kang and Hong Chang and Bingpeng MA and Shutao Bai and Shiguang Shan and Xilin Chen},
title = {Predictive Consistency Learning for Long-Tailed Recognition},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0264.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by