hrdkbhatnagar

Toy Models of Fea­ture Ab­sorp­tion in SAEs

[Paper] A is for Ab­sorp­tion: Study­ing Fea­ture Split­ting and Ab­sorp­tion in Sparse Autoencoders