IC-151Each CLIP neuron's second-order effect is approximately a single linear direction in the joint text-image space, significant for fewer than 2% of images

Yossi Gandelsman, Alexei A Efros, Jacob Steinhardt

SourceInterpreting the Second-Order Effects of Neurons in CLIP

The paper shows that the second-order effect of any individual neuron can be well approximated by one direction vector in CLIP's shared text-image representation space, scaled by an image-dependent coefficient. Replacing each neuron's full second-order contribution with this rank-1 approximation causes negligible accuracy loss on ImageNet. The effect is also highly selective: for each neuron, the contribution is large in norm for only about 2% of validation images; ablating the effect on the remaining 98% of images has no measurable impact. Sparse coding of the direction into text representations recovers the neuron's function with 128 descriptions.

Evidence
interventional
Key metric
Second-order effect significant for < 2% of images. First PC explains 48.2% of variance (Table 1). Reconstruction from PC #1 yields negligible accuracy drop from baseline (Figure 3). 128 text descriptions per neuron approach original classification accuracy (Figure 4).
Caveat
The rank-1 approximation is computed from ~5000 ImageNet training images; the paper notes the same behavior holds for ViT-L/14 and ImageNet-R but does not quantify the approximation error beyond the accuracy drop.
Model
CLIP / CLIP-ViT (LC)
Concepts
Linear representation
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], ImageNet-R / ImageNet-Rendition [eval]
Methods
Mean Ablation / direct effect mean ablation / Mean token ablation [primary], Orthogonal Matching Pursuit [primary]
Related work
Gandelsman et al. 2024 (TextSpan) [builds-on]
Related findings
IC-150, IC-152
Extraction
automatic-extraction