AI BriefingLG AI Research: Long-Tail Distribution Learning
Key point
This article covers the concept of Long-Tail Distribution (LTD) and various learning methodologies to address data imbalance problems.
Details
Large-scale datasets collected in real-world industrial settings often exhibit a Long-Tail Distribution (LTD), where some classes have many samples while others have very few. This data imbalance causes the model to be trained with a bias toward the majority classes (Head class), sharply degrading performance on the minority classes (Tail class).
The main methodologies for addressing this are as follows.
- Re-Sampling: A data-level approach that includes Under-Sampling, which reduces samples of the head class, and Over-Sampling, which increases samples of the tail class.
- Cost-Sensitive Learning: Trains the model by adjusting the loss function according to class imbalance.
- Transfer Learning: Uses pre-trained models to supplement performance on the tail class.
- Representation Learning: Enhances representational power so that features between classes can be better distinguished.
- Decoupled Training: Separates the training process to proceed with optimization.
This article takes an in-depth look at these traditional approaches as well as the latest deep learning-based solutions, such as PASCL (Partial and Asymmetric Supervised Contrastive Learning), presented at ICML 2022.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.