Summary of Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks, by Adam Karvonen et al.
Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks
by Adam Karvonen, Can Rager, Samuel Marks, Neel Nanda
First submitted to arxiv on: 28 Nov 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 The proposed paper introduces new evaluation metrics for Sparse Autoencoders (SAEs), which are used in neural networks to provide interpretability. The current methods of evaluating SAEs rely on unsupervised proxies, but the authors develop a family of evaluations based on the SHIFT task, which removes spurious cues from a classifier by ablating features deemed irrelevant by a human annotator. They replace this human annotator with an LLM and introduce the Targeted Probe Perturbation (TPP) metric to quantify an SAE’s ability to disentangle similar concepts. The authors apply these metrics to multiple open-source models, demonstrating their effectiveness in differentiating between various SAE training hyperparameters and architectures. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about a new way to test how well a neural network can understand what it’s doing. It uses something called Sparse Autoencoders (SAEs) which help us figure out why the network made a certain decision. Right now, we don’t have good ways to measure how well SAEs are working, so the authors came up with some new metrics to do just that. They used an idea from another paper where humans helped remove unimportant features from a classifier, but instead they used a computer program called an LLM. They also came up with a new metric called Targeted Probe Perturbation (TPP) that helps us see how well SAEs can tell the difference between similar ideas. The authors tested these metrics on lots of different models and showed that they work really well. |
Keywords
» Artificial intelligence » Neural network » Unsupervised
