All terms
Safety & Alignment
Sparse Autoencoder
A tool that breaks a model's internal signals into a large set of features, only a few active at a time.
Definition
A sparse autoencoder (SAE) is a small network that takes a model's internal signals (its activations) and rebuilds them from a large list of learnable features, with a rule that keeps only a few features switched on at once. Researchers use it to pull out features that each stand for one clear, human-readable concept. This gives a tidier picture than studying raw neurons, which tend to blur many concepts together.