Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, Victor Lempitsky
Computer vision tasks such as image classification, image retrieval and few-shot learning are currently dominated by Euclidean and spherical embeddings, so that the final decisions about class belongings or the degree of similarity are made using linear hyperplanes, Euclidean distances, or spherical geodesic distances (cosine similarity). In this work, we demonstrate that in many practical scenarios hyperbolic embeddings provide a better alternative.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Image Classification | CUB 200 5-way 5-shot | Accuracy | 72.22 | Hyperbolic ProtoNet |
| Image Classification | CUB 200 5-way 1-shot | Accuracy | 60.52 | Hyperbolic ProtoNet |
| Image Classification | OMNIGLOT - 1-Shot, 5-way | Accuracy | 99 | Hyperbolic ProtoNet |
| Image Classification | OMNIGLOT - 5-Shot, 5-way | Accuracy | 99.4 | Hyperbolic ProtoNet |
| Image Classification | Mini-Imagenet 5-way (5-shot) | Accuracy | 66.27 | Hyperbolic ProtoNet |
| Image Classification | Mini-Imagenet 5-way (1-shot) | Accuracy | 51.57 | Hyperbolic ProtoNet |
| Few-Shot Image Classification | CUB 200 5-way 5-shot | Accuracy | 72.22 | Hyperbolic ProtoNet |
| Few-Shot Image Classification | CUB 200 5-way 1-shot | Accuracy | 60.52 | Hyperbolic ProtoNet |
| Few-Shot Image Classification | OMNIGLOT - 1-Shot, 5-way | Accuracy | 99 | Hyperbolic ProtoNet |
| Few-Shot Image Classification | OMNIGLOT - 5-Shot, 5-way | Accuracy | 99.4 | Hyperbolic ProtoNet |
| Few-Shot Image Classification | Mini-Imagenet 5-way (5-shot) | Accuracy | 66.27 | Hyperbolic ProtoNet |
| Few-Shot Image Classification | Mini-Imagenet 5-way (1-shot) | Accuracy | 51.57 | Hyperbolic ProtoNet |