eh2022-sae-top-k-sparsity-requirement

IN premisesummaries/2026/08/24/elhage-2022-toy-models-superposition-chunk-1.md

Created 2026-08-25T02:57:58+00:00

A Sparse Autoencoder (SAE) uses top-k sparsity constraint; without it, the autoencoder becomes a dense linear projection and loses interpretability

Summary

The top-k sparsity constraint is what forces each input to activate only a small, fixed set of features rather than blending across all of them. Without that hard limit, the autoencoder degenerates into an ordinary linear transform, and you lose the discrete, human-readable features that make the whole technique useful for understanding what a model is actually computing.