VISION TRANSFORMERS WITH SELF-EMERGING TOKEN LABELING FOR ROBUSTNESS

Summary from the TLDR AI newsletter

In a significant advancement for Vision Transformers (ViTs), researchers have improved the robustness of the Fully Attentional Network (FAN) models using a new self-emerging token labeling (STL) framework. This approach involves training a FAN token labeler to create meaningful patch token labels, followed by training a FAN student model.