# Sparse Autoencoder **Domain:** Machine Learning / Interpretability **Doc Type:** Technical Concept Node **Maturity:** Established Method, Active Frontier Use **Related:** [[Monosemantic Feature]], [[Superposition]], [[Latent Space]], [[Dictionary Learning]] --- ## Definition A **sparse autoencoder** learns to reconstruct input activations through a larger set of latent features while encouraging only a small subset to activate for any one input. In model interpretability, it can serve as a learned dictionary that decomposes dense activation patterns into more legible feature directions. ## Boundary Recovered features are analytical representations, not guaranteed atomic concepts or proof that a model reasons through the same categories a human observer assigns to them.