notesum.ai
Published at December 9Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers
cs.CV
cs.AI
cs.LG
Released Date: December 9, 2024
Authors: Johanna Vielhaben1, Dilyara Bareeva1, Jim Berend1, Wojciech Samek2, Nils Strodthoff3
Aff.: 1Fraunhofer Heinrich-Hertz-Institute Berlin; 2Fraunhofer Heinrich-Hertz-Institute Berlin, Technische Universitat Berlin, BIFOLD – Berlin Institute for the Foundations of Learning and Data; 3Carl von Ossietzky Universitat Oldenburg

| FS | CLIP | DINO | MAE | ||
|---|---|---|---|---|---|
| SEQ | PCA | 0.45 | 0.90 | 1. | 0.60 |
| MCD | 0.81 | 0.90 | 0.81 | 0.70 | |
| KMeans | 0.90 | 0.90 | 0.81 | 1. | |
| NLMCD | 0.90 | 1. | 0.90 | 1. | |
| CLS | PCA | 0.75 | 0.72 | 0.50 | 0.33 |
| MCD | 0.33 | 0.16 | 0.16 | 0.16 | |
| KMeans | 0.83 | 0.91 | 0.50 | 0.66 | |
| NLMCD | 1. | 1. | 0.83 | 0.75 |