Summary of Krony-pt: Gpt2 Compressed with Kronecker Products, by M. Ayoub Ben Ayad et al.
Krony-PT: GPT2 compressed with Kronecker Products
by M. Ayoub Ben Ayad, Jelena Mitrovic, Michael Granitzer
First submitted to arxiv on: 16 Dec 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: 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 We present Krony-PT, a novel compression technique for GPT2-based transformer models. By applying Kronecker Products to MLP layers and systematically compressing feed-forward layer matrices, we reduce the model’s size while preserving its language modeling capabilities. Our method, which initializes new factors using modified Van Loan decomposition and pruning-based techniques, yields smaller models with competitive performance on standard datasets. The 81M variant outperforms distilgpt2 on next-token prediction tasks, demonstrating the effectiveness of Krony-PT in compressing GPT2-based models for efficient language modeling. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary We created a new way to make large language models smaller while still keeping their ability to predict words. We used a special math technique called Kronecker Products to shrink the model’s size and keep its language skills intact. Our method works by breaking down big pieces of the model into smaller ones, which makes it easier to train and use on devices with limited resources. Our smallest model is 80M parameters, and our largest is 96M. We tested this new approach and found that our 81M model can predict words just as well as larger models, making it a useful tool for people who want to use language models without using too much computer power. |
Keywords
» Artificial intelligence » Pruning » Token » Transformer
