IC-678Specific attention heads in CLIP ViT-L's last 4 layers encode specific image properties (color, shape, location, counting, texture) that are linearly recoverable via text directions

Yossi Gandelsman, Alexei A Efros, Jacob Steinhardt

SourceInterpreting CLIP's Image Representation via Text-Based Decomposition

Using the TextSpan algorithm, the paper finds a set of text descriptions that span each head's output space in the joint vision-language representation. For many heads, the top descriptions centre on a single property: e.g. layer 22 head 1 is a 'shape' head (semicircular arch, isosceles triangle, oval), layer 22 head 11 is a 'color' head, layer 23 head 10 is a 'counting' head, layer 23 head 12 is a 'texture' head. Projecting each head's 768-dim output to a 60-dim text-interpretable subspace retains 72.77% of the 75.25% base ImageNet accuracy. Retrieval using individual head contributions returns images matching the head's property.

Evidence
observational
Key metric
60 descriptions per head reach 72.77% ImageNet zero-shot accuracy (base: 75.25%). ChatGPT-generated descriptions outperform common words and random vectors at all sizes m.
Caveat
Not all heads have clear single-property roles; some may not correspond to coherent properties, or may require collaborative interpretation.
Model
OpenCLIP ViT-L/14
Concepts
Linear representation
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Methods
TextSpan [primary], Mean Ablation / direct effect mean ablation / Mean token ablation [supporting]
Related work
CLIP [builds-on]
Related findings
IC-677, IC-679, IC-680
Extraction
automatic-extraction