Inception

Google

Structures found in this family (2)

Linear Separability, Linear Direction

By model (1)

Inception V3 (ImageNet-trained)

Linear Separability, Linear Direction

Observations (2)

Linear-probe separability increases monotonically with CNN depth

TCAV concept activation vectors are causally significant linear directions

Papers

Understanding Intermediate Layers Using Linear Classifier Probes (2016), Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV) (2018)