Advances in Machine Learning and Data Augmentation

The field of machine learning is rapidly advancing, with a focus on developing new techniques for improving model performance and generalization. One key area of research is data augmentation, which involves modifying training data to increase its diversity and reduce overfitting. Recent work has shown that data augmentation can be optimized using Bayesian model selection, leading to improved calibration and robust performance. Another area of focus is contrastive learning, which involves training models to distinguish between similar and dissimilar examples. Researchers have proposed new methods for contrastive learning, including weakly-supervised approaches that can handle imprecise class labels. Additionally, there has been progress in developing new regularization techniques, such as relevance-driven input dropout, which can improve model generalization by selectively occluding important features. Overall, these advances have the potential to improve the performance and reliability of machine learning models in a wide range of applications. Noteworthy papers include: Locality-Sensitive Hashing for Efficient Hard Negative Sampling in Contrastive Learning, which proposes a novel hashing scheme for efficient hard negative sampling. Boosting Open Set Recognition Performance through Modulated Representation Learning, which introduces a novel negative cosine scheduling scheme to improve open set recognition performance.

Sources

The Quasi-Polynomial Low-Degree Conjecture is False

Locality-Sensitive Hashing for Efficient Hard Negative Sampling in Contrastive Learning

Boosting Open Set Recognition Performance through Modulated Representation Learning

Self-Organizing Visual Prototypes for Non-Parametric Representation Learning

Relevance-driven Input Dropout: an Explanation-guided Regularization Technique

Optimizing Data Augmentation through Bayesian Model Selection

Weakly-Supervised Contrastive Learning for Imprecise Class Labels

The informativeness of the gradient revisited

An Augmentation-Aware Theory for Self-Supervised Contrastive Learning

Progressive Data Dropout: An Embarrassingly Simple Approach to Faster Training

Multiclass Loss Geometry Matters for Generalization of Gradient Descent in Separable Classification

Refining Labeling Functions with Limited Labeled Data

Built with on top of