Label-Aware Distribution Calibration for Long-Tailed Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Natural Language Processing

Background:

• Real-world data often exhibits long-tailed distributions, leading to imbalanced datasets.
• Training neural networks on imbalanced data results in poor performance on underrepresented (tail) classes due to sparse training instances and biased distribution estimation.
• Existing methods like data resampling and synthesis have limitations in addressing the core challenge of tail class sparseness.

Purpose of the Study:

• To address the challenge of imbalanced data in machine learning by proposing a novel distribution calibration method.
• To leverage transferable knowledge from head classes to enrich tail classes and improve model generalization.
• To introduce label-aware distribution calibration (LADC) for more accurate distribution estimation and classifier rebalancing.

Main Methods:

• Proposed a novel distribution calibration (DC) approach named label-aware DC (LADC).
• LADC transfers statistical information from relevant head classes to infer the distribution of tail classes.
• Utilized sampling from the calibrated distribution to rebalance the classifier.

Main Results:

• LADC significantly outperforms existing methods on both image and text long-tailed datasets.
• Experimental results demonstrate LADC's effectiveness in improving performance on tail classes.
• Visualizations confirm that LADC provides a more accurate estimation of data distribution.

Conclusions:

• LADC offers a novel and effective solution for handling long-tailed distributions in machine learning.
• The proposed method successfully transfers knowledge from head to tail classes, enhancing classifier performance.
• LADC represents a significant advancement in training robust neural networks on imbalanced real-world data.