Automatic Data Curation for Self-Supervised Learning: A Clustering-Based Approach

This paper presents a method for automating data curation for self-supervised learning by using successive and hierarchical k-means clustering.

The approach aims to create datasets that are large, diverse, and balanced, which can outperform uncurated data and match or surpass manually curated data in performance.

The technique has been tested across various data domains, including web-based images, satellite images, and text.