Summary of Don’t Look Twice: Faster Video Transformers with Run-length Tokenization, by Rohan Choudhury et al.
Don’t Look Twice: Faster Video Transformers with Run-Length Tokenization
by Rohan Choudhury, Guanglei Zhu, Sihan Liu, Koichiro Niinuma, Kris M. Kitani, László Jeni
First submitted to arxiv on: 7 Nov 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 proposed Run-Length Tokenization (RLT) method accelerates video transformers by efficiently removing repeated tokens in videos. This technique, inspired by run-length encoding for data compression, reduces the number of input tokens fed into the model, leading to significant speedups during training and inference. RLT is content-aware and requires no tuning for different datasets, achieving a 30% reduction in wall-clock time to fine-tune video transformers while maintaining baseline performance. Additionally, RLT increases model throughput by 35% with only a minor drop in accuracy, outperforming existing methods that often require significant overhead or dataset-specific tuning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Video transformers are slow because they process many repeated tokens from videos. A new way to speed them up is called Run-Length Tokenization (RLT). RLT works by finding and removing repeated patches of video frames before the model looks at them. This makes training faster, taking 30% less time while still giving good results. It also makes the model run faster during inference, with a boost of 35% in speed without sacrificing accuracy. |
Keywords
» Artificial intelligence » Inference » Tokenization
