Get the latest tech news
An Intuitive Explanation of Sparse Autoencoders for LLM Interpretability
Sparse Autoencoders (SAEs) have recently become popular for interpretability of machine learning models (although sparse dictionary learning has been around since 1997). Machine learning models and LLMs are becoming more powerful and useful, but they are still black boxes, and we don’t understand how they do the things that they are capable of. It seems like it would be useful if we could understand how they work.
In that case, without additional constraints the task is trivial, and the SAE could use the identity matrix to perfectly reconstruct the input without telling us anything interesting. As an additional constraint, we add a sparsity penalty to the training loss, which incentivizes the SAE to create a sparse intermediate vector. Our training data for these SAEs comes from feeding a diverse range of text through the GPT model and collecting the intermediate activations at each chosen location.
Or read this on Hacker News