Table 1.
Performances of our model compared to other non-episodic unsupervised feature learning methods on Omniglot and MiniImageNet.
| Omniglot | MiniImageNet | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Methods (M, K) | Clustering | Metric | (5,1) | (5,5) | (20,1) | (20,5) | (5,1) | (5,5) | (5,20) | (5,50) |
| Baseline | N/A | N/A | 57.97 | 79.25 | 34.17 | 59.33 | 25.91 | 32.38 | 37.01 | 38.95 |
| AutoEncoder | N/A | N/A | 53.63 | 77.34 | 32.98 | 55.01 | 26.17 | 33.01 | 37.98 | 39.39 |
| Denoising autoEncoder | N/A | N/A | 59.63 | 79.89 | 34.78 | 60.88 | 27.81 | 34.19 | 39.01 | 40.11 |
| InfoGAN | N/A | N/A | 51.49 | 76.38 | 31.01 | 53.99 | 29.81 | 36.47 | 40.17 | 42.46 |
| BiGAN+KNN | N/A | N/A | 49.55 | 68.06 | 27.37 | 46.70 | 25.56 | 31.10 | 37.31 | 43.60 |
| BiGAN+LC | N/A | N/A | - | - | - | - | 27.08 | 33.91 | 44.00 | 50.41 |
| DeepClustering | Kmeans | Euclidean | 59.07 | 79.81 | 34.05 | 60.12 | 28.91 | 36.01 | 39.29 | 41.98 |
| UFLST | Kmeans | Euclidean | 69.54 | 86.18 | 47.11 | 69.19 | 31.77 | 43.03 | 51.35 | 55.72 |
| UFLST | BSCAN | KRJD | 96.51 | 99.23 | 90.27 | 97.22 | 37.75 | 50.95 | 59.18 | 62.27 |
Baseline performance means training from scratch. Results based on BiGAN are adapted from Hsu et al. (2018). For complete results with confidence intervals, see Appendix 6. The best performances are in bold.