Toward Identifiable Sparse Autoencoders
Two new research papers explore challenges and solutions for sparse autoencoders (SAEs), a tool used to interpret neural network representations. One paper introduces "identifiable SAEs" (iSAEs) that offer improved stability and lower reconstruction error by addressing architectural and training issues. The other paper identifies "activation outliers" as the cause of "feature death" in SAEs, where learned features fail to activate, and proposes mean-centering as a solution to prevent this issue across various model types. AI
IMPACT These papers offer methods to improve the interpretability and stability of neural network representations, potentially aiding in debugging and understanding complex models.