Summary of Boosting Multimodal Large Language Models with Visual Tokens Withdrawal For Rapid Inference, by Zhihang Lin et al.
Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference
by Zhihang Lin, Mingbao Lin, Luxi Lin, Rongrong Ji
First submitted to arxiv on: 9 May 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
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 introduces Visual Tokens Withdrawal (VTW), a module to accelerate inference in Multimodal Large Language Models (MLLMs). MLLMs require significant computations due to their extensive parameters and additional visual information representation. VTW is inspired by two phenomena: attention sink, where initial tokens receive most attention, and information migration, which transfers visual info to text tokens in early layers. The approach strategically withdraws vision tokens at a certain layer, allowing only text tokens to engage with subsequent layers. To determine the ideal layer for VTW, the authors analyze a limited set of tiny datasets using the Kullback-Leibler divergence criterion. This results in a 40% reduction in computational overhead across diverse multimodal tasks while maintaining performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper introduces Visual Tokens Withdrawal (VTW), a new way to make Multimodal Large Language Models (MLLMs) faster and more efficient. Right now, MLLMs need a lot of computer power because they have so many parameters and they need to process visual information too. VTW helps by taking out the extra visual information that isn’t needed in the later parts of the model. This makes the model run about 40% faster for different tasks while still getting good results. |
Keywords
» Artificial intelligence » Attention » Inference
