Summary of Llava-prumerge: Adaptive Token Reduction For Efficient Large Multimodal Models, by Yuzhang Shang et al.
LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models
by Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, Yan Yan
First submitted to arxiv on: 22 Mar 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 Large Multimodal Models (LMMs) have achieved impressive visual reasoning capabilities by combining a visual encoder and a large language model. Recent LMMs process complex visual inputs like high-resolution images and videos, increasing the number of visual tokens significantly. However, this design choice leads to quadratic computational costs with the number of input tokens. To address this issue, we propose PruMerge, an adaptive visual token reduction strategy that identifies significant spatial redundancy among visual tokens without compromising performance. By exploiting sparsity in the visual encoder’s attention scores, we dynamically select crucial tokens and cluster them based on their similarity. This approach can compress visual tokens by 14 times on average while achieving comparable results across various visual question-answering and reasoning tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research paper is about making computer models better at understanding pictures and videos. These models are very good at answering questions, but they need a lot of information to do it. The problem is that using too much information makes the model slower and uses more energy. To fix this, we developed a new way to reduce the amount of information used by the model without making it any worse. We did this by looking for patterns in how the model pays attention to different parts of the picture or video. This allowed us to keep only the most important parts and still get good results. |
Keywords
» Artificial intelligence » Attention » Encoder » Large language model » Question answering » Token
