LEARNING BETTER VISUAL REPRESENTATIONS

Summary from the TLDR AI newsletter

This project proposes an algorithm to filter web datasets used for training CLIP in order to learn better visual representations and achieve state-of-art zero-shot accuracy on vision tasks.