Summary of Dynamic Layer Tying For Parameter-efficient Transformers, by Tamir David Hay et al.
Dynamic Layer Tying for Parameter-Efficient Transformers
by Tamir David Hay, Lior Wolf
First submitted to arxiv on: 23 Jan 2024
Categories
- Main: Machine Learning (cs.LG)
- 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 This paper proposes a novel approach to reduce the number of trainable parameters in deep transformer networks using Reinforcement Learning (RL). The authors introduce an RL agent that dynamically selects layers during training and ties them together. This agent is asked every few iterations whether to train each layer independently or copy the weights of a previous layer, facilitating weight sharing and reducing the number of trainable parameters. Experimental results show that this model modestly outperforms the baseline transformer model in terms of perplexity while significantly reducing the number of trainable parameters, resulting in up to one order of magnitude less memory consumption during training. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper uses a special kind of computer learning called Reinforcement Learning (RL) to make deep computers work more efficiently. Right now, these computers have many tiny parts that need to be trained separately, which takes a lot of time and space. The authors came up with an idea where the computer learns to decide when to train each part or copy what another part has learned before. This helps share information between parts and reduces how much training is needed overall. They tested this new way and found it works well, making the computers faster and more efficient. |
Keywords
* Artificial intelligence * Perplexity * Reinforcement learning * Transformer
