Table 2:
MIG metrics on the held-out data
| Dataset | d (u) | Proposed | VAE | β-VAE | Factor-VAE | TCVAE | DSD |
|---|---|---|---|---|---|---|---|
| MNIST | 2 | 0.68 | 0.01 | 0.03 | 0.33 | 0.04 | 0.01 |
| Fashion-MNIST | 2 | 0.52 | 0.11 | 0.28 | 0.36 | 0.19 | 0.01 |
| Yale Faces | 1 | 0.42 | 0.02 | 0.07 | 0.06 | 0.29 | N/A |
| 3D chairs | 2 | 0.37 | 0.02 | 0.15 | 0.11 | 0.08 | N/A |
| 3D cars | 2 | 0.41 | 0.02 | 0.22 | 0.15 | 0.16 | N/A |
DSD is designed for analyzing binary similarities, and cannot analyze real-valued similarities.