Generalized Parametric Contrastive Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Supervised contrastive loss can exhibit bias towards high-frequency classes, complicating imbalanced learning.
• Existing methods struggle with effectively handling imbalanced data distributions in deep learning models.

Purpose of the Study:

• To introduce Generalized Parametric Contrastive Learning (GPaCo), a novel method for robust learning on both imbalanced and balanced datasets.
• To theoretically and empirically demonstrate GPaCo's ability to mitigate class imbalance issues and improve model performance.

Main Methods:

• Developed a Generalized Parametric Contrastive Learning (GPaCo) framework incorporating parametric class-wise learnable centers.
• Introduced a rebalancing mechanism from an optimization perspective to address class imbalance.
• Analyzed the GPaCo loss function under balanced settings to understand its adaptive properties.

Main Results:

• GPaCo achieves state-of-the-art results on long-tailed recognition benchmarks.
• Models trained with GPaCo exhibit superior generalization and robustness on ImageNet compared to MAE models.
• Significant improvements were observed when applying GPaCo to semantic segmentation tasks across four popular benchmarks.

Conclusions:

• GPaCo offers an effective solution for imbalanced learning by rebalancing class importance through learnable centers.
• The proposed method enhances feature representation, benefiting hard example learning and improving overall model performance.
• GPaCo demonstrates broad applicability and effectiveness across diverse computer vision tasks and architectures.