Summary of Repeat After Me: Transformers Are Better Than State Space Models at Copying, by Samy Jelassi et al.
Repeat After Me: Transformers are Better than State Space Models at Copying
by Samy Jelassi, David Brandfonbrener, Sham M. Kakade, Eran Malach
First submitted to arxiv on: 1 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); 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 Transformers have dominated sequence modeling, but researchers are exploring alternative architectures called Generalized State Space Models (GSSMs). While GSSMs show promise in terms of inference-time efficiency, they fall short compared to transformers on tasks requiring copying from the input context. A theoretical analysis reveals that a two-layer transformer can copy strings of exponential length, whereas GSSMs are limited by their fixed-size latent state. Empirical results confirm transformers outperform GSSMs in efficiency and generalization on synthetic tasks. Pretrained large language models also favor transformers over GSSMs for copying and retrieving information from context. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Transformers are super smart at processing sequences, but scientists want to know if they can do better with a different approach called Generalized State Space Models (GSSMs). They’re looking into how well these new models work on tasks that need them to copy information from the start. Surprisingly, transformers can copy really long strings while GSSMs are stuck with their limited size. In real-life tests, transformers did better at copying and remembering info from context. |
Keywords
* Artificial intelligence * Generalization * Inference * Transformer
