TABLE 4.
Cosine similarities of individual CAVs trained on concept data. Results are given averaged over all datasets, as well as the averages over model architectures trained on single datasets. Cosine similarity of 1 indicates full alignment, while 0 indicates orthogonality.
| Dataset | SCDB | ISIC | EyePACS | Overall |
|---|---|---|---|---|
| Baseline | 0.44 | 0.46 | 0.40 | 0.44 |
| Overfit | 0.37 | 0.44 | 0.38 | 0.40 |
| DP | 0.35 | 0.48 | 0.42 | 0.42 |