The Equalization Losses: Gradient-Driven Training for Long-tailed Object Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Long-tail distributions are prevalent in real-world data, leading to poor performance on underrepresented categories.
• This performance gap is primarily caused by imbalanced gradients from positive (within-category) and negative (between-category) samples.
• The ratio of accumulated positive to negative gradients indicates category training balance.

Purpose of the Study:

• To investigate the cause of performance degradation in long-tail distributions.
• To develop a novel gradient-driven training mechanism to address imbalanced gradients.
• To introduce a new family of loss functions for improved long-tail learning.

Main Methods:

• A gradient-driven training mechanism was developed to dynamically re-balance positive and negative gradients.
• This mechanism aims to achieve balanced gradient ratios for all categories.
• A new family of loss functions, termed equalization losses, was introduced based on this mechanism.

Main Results:

• The proposed equalization losses consistently outperformed baseline models across diverse visual tasks.
• Experiments were conducted on long-tailed object detection (LVIS), image classification (ImageNet-LT, Places-LT, iNaturalist), and semantic segmentation (ADE20 K).
• The method demonstrated significant improvements in accuracy for tail categories.

Conclusions:

• The proposed gradient-driven mechanism effectively tackles the long-tail problem by addressing imbalanced gradients.
• Equalization losses offer a flexible and effective solution for improving model performance on underrepresented data.
• The approach shows strong generalization capabilities across multiple computer vision tasks.