VISION TRANSFORMERS NEED REGISTERS

Summary from the TLDR AI newsletter

One of the coolest and simplest vision papers in recent weeks. Vision Transformers use "useless" pixel values as places to store global information. This makes attention maps uninterpretable. However, if you add a simple [reg] token to the vocabulary, the model will use it and won't store information in pixel values.