Block-Sparse Featurizers
Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds
Block-Sparse Featurizers (BSFs) are Goodfire's open-source interpretability featurizer methods and code for decomposing neural network activations into sparse multidimensional blocks/subspaces rather than single SAE-style directions, applied in the research to vision models including DINOv3, SDXL, and InceptionV1.
Recent stories
1 linked story