Summary of A Manifold Representation Of the Key in Vision Transformers, by Li Meng et al.
A Manifold Representation of the Key in Vision Transformers
by Li Meng, Morten Goodwin, Anis Yazidi, Paal Engelstad
First submitted to arxiv on: 1 Feb 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper explores a novel approach in Vision Transformers by disentangling the key from the query and value, and adopting a manifold representation for the key. The existing multi-head self-attention mechanism stacks multiple attention blocks, where the query, key, and value are intertwined and generated within those blocks via a single shared linear transformation. By decoupling the key from the query and value, the authors demonstrate that this approach can enhance the model’s performance on various tasks such as image classification, object detection, and instance segmentation. Specifically, ViT-B exhibits a 0.87% increase in top-1 accuracy and Swin-T sees a boost of 0.52% in top-1 accuracy on the ImageNet-1K dataset. The authors’ approach also yields positive results on the COCO dataset. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps Vision Transformers work better by making changes to how they pay attention to things. Normally, these models use a special trick called multi-head self-attention that makes them good at understanding images. But the researchers found that if they break apart this trick into three parts – query, key, and value – and give the key its own special way of working, it can make the model even better. They tested this idea on lots of different tasks and found that it works really well. For example, one type of model called ViT-B got 0.87% better at guessing what was in a picture, while another model called Swin-T got 0.52% better. |
Keywords
* Artificial intelligence * Attention * Image classification * Instance segmentation * Object detection * Self attention * Vit
