Can sparse autoencoders make sense of gene expression latent variable models?

Viktoria Schuster

Sparse autoencoders (SAEs) have lately been used to uncover interpretable latent features in large language models. By projecting dense embeddings into a much higher-dimensional an…