GoogLeNet (Inception v1)

Google

Structures found in this family (2)

By model (1)

GoogLeNet (ImageNet-trained)

Observations (2)

Block-sparse featurizers recover concepts as two- to four-dimensional subspaces in DINOv3, one curve manifold in InceptionV1, and steerable blocks in SDXL

TCAV concept activation vectors are causally significant linear directions

Papers

Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds (2026), Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV) (2018)