Quantization SAE

Creator
Creator
Seonglae ChoSeonglae Cho
Created
Created
2024 Oct 24 0:3
Editor
Edited
Edited
2024 Nov 19 22:35
Refs
Refs
 
 
 
 
 
 

Codebook feature
VQ-VAE

Residual Vector Quantization
bottleneck 양자화된 벡터가 더 유의미한 정보만을 포함하도록

RQAE (Residual Quantization Autoencoder)
Residual Vector Quantization
(Stacked SAEs)

RQAE learns features hierarchically, by iteratively applying vector quantization on different subspaces of the residual stream of an LLM.
RQAE has much higher capacity than SAEs
RQAE has much higher capacity than SAEs
 
 

Recommendations