DENOISING VISION TRANSFORMERS

Summary from the TLDR AI newsletter

Vision Transformers (ViT) have taken over the vision community. They still sometimes display grid-like artifacts in their embeddings. This has made the community hesitant to adopt them for downstream tasks. This work proposes an update to the positional embeddings that eliminates this issue and gives a 25%+ boost for downstream vision tasks.