Summary of Compression Via Pre-trained Transformers: a Study on Byte-level Multimodal Data, by David Heurtel-depeiges et al.
Compression via Pre-trained Transformers: A Study on Byte-Level Multimodal Data
by David Heurtel-Depeiges, Anian Ruoss, Joel Veness, Tim Genewein
First submitted to arxiv on: 7 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Information Theory (cs.IT)
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 study investigates whether pre-trained vanilla transformers can achieve competitive compression ratios when reduced in size. Researchers train families of models on massive datasets and compress out-of-distribution data from various modalities (text, image, audio). The findings show that small models (millions of parameters) can outperform standard compression algorithms and domain-specific compressors, achieving the lowest compression ratio of 0.49 on OOD audio data. Ablation studies reveal that even small models can be trained to perform well on multiple modalities, but transfer to unseen modalities is generally weak. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study looks at how well pre-trained transformers can squeeze down and still compress data well. They train lots of models and test them on different types of data (text, pictures, audio). It turns out that smaller models (with millions of parameters) can do a better job than usual compression tools or ones designed for specific types of files. The smallest model they tested did 0.49 times as much data in the same space. They also found that even small models can be trained to work well on multiple types of data, but if they’re not familiar with the type of data, they don’t do very well. |
