Block-Sparse Featurizers
Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds
Open-source research tool/methods from Goodfire for decomposing neural-network activations into block-sparse multidimensional subspaces/manifolds instead of single SAE directions; demonstrated on vision and image-generation models including DINOv3, InceptionV1, and SDXL.
Recent stories
0 linked stories
No linked stories yet.