The authors propose that sparse autoencoder features should form an “atomic interpretable unit” basis: a coordinate system whose units support semantically meaningful explanations and interventions.
Sparse autoencoder evaluations can miss ambiguity in their most dominant features
Zhang and Wang propose a unified evaluation framework and report that dominant features can be less interpretable and more redundant despite favorable conventional scores.
Academic
Riteng Zhang · Xiaoqian Wang
Purdue University
Research Digest··2 min read
Sparse autoencoders try to make neural-network representations understandable by separating them into features associated with recognizable concepts.
What it released
CodeWeights
§