Summary of Interpbench: Semi-synthetic Transformers For Evaluating Mechanistic Interpretability Techniques, by Rohan Gupta et al.
InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques
by Rohan Gupta, Iván Arcuschin, Thomas Kwa, Adrià Garriga-Alonso
First submitted to arxiv on: 19 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 presents InterpBench, a collection of semi-synthetic transformers with known circuits for evaluating mechanistic interpretability methods. The authors propose Strict IIT (SIIT), a training approach that aligns neural networks’ internal computation with a desired high-level causal model while preventing non-circuit nodes from affecting the output. SIIT is evaluated on sparse transformers produced by Tracr and found to maintain the original circuit, making it more realistic. Additionally, SIIT can train transformers with larger circuits like IOI. The authors also use their benchmark to evaluate existing circuit discovery techniques. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us understand how neural networks work. It’s hard to figure out what a network does when we don’t know the “secret recipe” that makes it tick. To solve this problem, researchers created a special set of fake data and models with known rules for how they work. They called this collection InterpBench. The authors then developed a new way to train neural networks called SIIT. This method makes sure that the network’s internal workings match what we want it to do, while also keeping things simple by not letting unnecessary parts affect the outcome. The researchers tested SIIT on some special models and found that it works well. They even used their benchmark to see how other techniques for discovering how networks work compare. |
