IC-447CLIP ViT-B/16 produces noisy saliency maps and contains only 42 concept detectors, indicating poor visual interpretability

Shizhan Gong, Haoyu LEI, Qi Dou, Farzan Farnia

SourceBoosting the visual interpretability of CLIP via adversarial fine-tuning

The paper measures the visual interpretability of the original CLIP ViT-B/16 image encoder using multiple feature attribution methods and network dissection. Simple gradient saliency maps are described as 'largely stochastic without highlighting any meaningful patterns,' and Grad-CAM maps suffer from 'background noise interference.' Quantitatively, the pointing game score is 25.77 for simple gradients and 51.00 for Grad-CAM on the ImageNet-Segmentation validation set. Network dissection via CLIP-Dissect reveals only 42 concept detectors in the [CLS] token of the last layer, compared to 159 after AFT. The authors attribute this to the non-smoothness induced by false negatives during contrastive training and the inherent nature of the ViT architecture.

Evidence
correlational
Key metric
PG 25.77 (SG), 51.00 (GC) on ImageNet-Segmentation; 42 concept detectors (Figure 1); PG-energy 31.32 (SG), 43.71 (GC); pixel acc. 65.51 (SG), 61.79 (GC)
Caveat
The paper notes that more advanced attribution methods (M2IB, Grad-ECLIP) already produce better saliency maps on the original CLIP, so the poor interpretability is most evident with basic gradient-based methods.
Model
CLIP / CLIP-ViT (LC)
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], ImageNet-Segmentation [eval], BRODEN [eval], ImageNetette [eval], Caltech-256 [eval]
Methods
Simple Gradient [primary], Grad-CAM [primary], Pointing game [eval], CLIP-Dissect [eval], Network Dissection [eval], M2IB [compared-to], Grad-ECLIP [compared-to], ROAR [validation]
Related findings
IC-448, IC-449, IC-450
Extraction
automatic-extraction