Playing with Vision Embeddings

The article explores the concept of neural network embeddings, specifically the 384-dimensional space of the DINOv3 ViT-S model, and uses techniques such as sparse autoencoders and feature visualization to understand what information is encoded in those numbers.