Summary of Matryoshka Query Transformer For Large Vision-language Models, by Wenbo Hu et al.
Matryoshka Query Transformer for Large Vision-Language Models
by Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Amita Kamath, Nanyun Peng, Kai-Wei Chang
First submitted to arxiv on: 29 May 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Computation and Language (cs.CL); 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 proposed Matryoshka Query Transformer (MQT) enables Large Vision-Language Models (LVLMs) to adapt to varying computational constraints by encoding images into a dynamic number of visual tokens. Inspired by Matryoshka Representation Learning, MQT uses a query transformer with latent query tokens to compress visual embeddings during inference. This allows for flexible and drastic reductions in the number of visual tokens while maintaining performance comparable to training separate models for each token count. The combined model, MQT-LLAVA, matches LLaVA-1.5 performance across 11 benchmarks using up to 256 tokens, with only a 2.4-point drop in performance when reducing to 16 tokens (8x less TFLOPs). This flexibility enables LVLMs to achieve the best of both worlds. |
Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine you have a super powerful computer that can do lots of calculations quickly. Now imagine you had this same computer, but it was much smaller and could only do half as many calculations at once. You’d still want the computer to be able to get the job done, right? That’s what these researchers did with their “Large Vision-Language Models” (LVLMs). They found a way to make them work better even when they’re not given as much computing power. This is useful because sometimes computers are big and powerful, but other times they might be smaller or less powerful. The researchers came up with an idea called the “Matryoshka Query Transformer” that lets LVLMs adjust how many “visual tokens” (like little building blocks) it uses to process images. This way, the model can still do a good job even when it’s given fewer visual tokens. |
Keywords
» Artificial intelligence » Inference » Representation learning » Token » Transformer