Isa Camacho retweetledi
Happy to share a fun little project with Ashwin Saraswatula:
"Data Whitening Improves Sparse Autoencoder Learning"
TL;DR: Whitening is cheap, standard preprocessing in sparse coding. It also helps SAEs learn a better solution, which we show on SAEBench.
arxiv.org/abs/2511.13981
@CSHL
[1/11]
English