Dominik Filipiak, Anna Fensel, Agata Filipowska
We propose KGTN-ens, a framework extending the recent Knowledge Graph Transfer Network (KGTN) in order to incorporate multiple knowledge graph embeddings at a small cost. We evaluate it with different combinations of embeddings in a few-shot image classification task. We also construct a new knowledge source - Wikidata embeddings - and evaluate it with KGTN and KGTN-ens. Our approach outperforms KGTN in terms of the top-5 accuracy on the ImageNet-FS dataset for the majority of tested settings.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Image Classification | ImageNet-FS (10-shot, all) | Top-5 Accuracy (%) | 83.46 | KGTN-ens (ResNet-50, h+g, max) |
| Image Classification | ImageNet-FS (5-shot, novel) | Top-5 Accuracy (%) | 78.9 | KGTN-ens (ResNet-50, h+g, mean) |
| Image Classification | ImageNet-FS (2-shot, all) | Top-5 Accuracy (%) | 75.45 | KGTN-ens (ResNet-50, h+g, max) |
| Image Classification | ImageNet-FS (5-shot, all) | Top-5 Accuracy (%) | 81.12 | KGTN-ens (ResNet-50, h+g, max) |
| Image Classification | ImageNet-FS (1-shot, novel) | Top-5 Accuracy (%) | 62.73 | KGTN-ens (ResNet-50, h+g, max) |
| Image Classification | ImageNet-FS (1-shot, all) | Top-5 Accuracy (%) | 68.58 | KGTN-ens (ResNet-50, h+g, max) |
| Image Classification | ImageNet-FS (2-shot, novel) | Top-5 Accuracy (%) | 71.48 | KGTN-ens (ResNet-50, h+g, max) |
| Image Classification | ImageNet-FS (10-shot, novel) | Top-5 Accuracy (%) | 82.56 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (10-shot, all) | Top-5 Accuracy (%) | 83.46 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (5-shot, novel) | Top-5 Accuracy (%) | 78.9 | KGTN-ens (ResNet-50, h+g, mean) |
| Few-Shot Image Classification | ImageNet-FS (2-shot, all) | Top-5 Accuracy (%) | 75.45 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (5-shot, all) | Top-5 Accuracy (%) | 81.12 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (1-shot, novel) | Top-5 Accuracy (%) | 62.73 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (1-shot, all) | Top-5 Accuracy (%) | 68.58 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (2-shot, novel) | Top-5 Accuracy (%) | 71.48 | KGTN-ens (ResNet-50, h+g, max) |
| Few-Shot Image Classification | ImageNet-FS (10-shot, novel) | Top-5 Accuracy (%) | 82.56 | KGTN-ens (ResNet-50, h+g, max) |