The content discusses the challenges of long-tailed distributions in real-world data and introduces ProCo, a probabilistic contrastive learning algorithm. ProCo estimates feature distributions using von Mises-Fisher distributions and samples contrastive pairs effectively. The method is validated through experiments on various datasets, showcasing improved performance compared to existing methods.
Recent advancements in deep learning have led to significant progress in computer vision tasks. However, real-world data often exhibits long-tailed patterns with imbalanced class distributions. This imbalance poses challenges for training deep models as they may struggle to generalize to infrequent categories due to limited training data.
Supervised contrastive learning (SCL) has shown promise in addressing long-tail distribution issues by integrating label information into the formulation of positive and negative pairs for the contrastive loss function. However, SCL requires large batch sizes for generating sufficient contrastive pairs, leading to computational and memory overheads.
To overcome these challenges, the author proposes ProCo, a novel probabilistic contrastive learning algorithm that estimates feature distributions using von Mises-Fisher distributions. By sampling contrastive pairs efficiently, ProCo eliminates the need for large batch sizes while improving performance on imbalanced datasets.
Empirical evaluations on supervised/semi-supervised visual recognition tasks demonstrate that ProCo consistently outperforms existing methods across various datasets. The method also shows enhanced performance on balanced datasets and can be applied to semi-supervised learning scenarios effectively.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Chaoqun Du,Y... kl. arxiv.org 03-12-2024
https://arxiv.org/pdf/2403.06726.pdfDybere Forespørgsler