摘要
我们提出KGTN-ens框架，该框架在近期提出的知识图谱迁移网络（Knowledge Graph Transfer Network, KGTN）基础上进行扩展，旨在以较低成本融合多种知识图谱嵌入表示。我们在少样本图像分类任务中，采用不同组合的嵌入表示对所提方法进行了评估。此外，我们构建了一种新的知识源——Wikidata嵌入，并将其应用于KGTN与KGTN-ens的性能比较。实验结果表明，在ImageNet-FS数据集上，对于大多数测试设置，我们的方法在top-5准确率方面均优于原始的KGTN模型。
代码仓库
DominikFilipiak/KGTN-ens
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| few-shot-image-classification-on-imagenet-fs | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 62.73 |
| few-shot-image-classification-on-imagenet-fs-1 | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 71.48 |
| few-shot-image-classification-on-imagenet-fs-2 | KGTN-ens (ResNet-50, h+g, mean) | Top-5 Accuracy (%): 78.90 |
| few-shot-image-classification-on-imagenet-fs-3 | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 82.56 |
| few-shot-image-classification-on-imagenet-fs-4 | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 68.58 |
| few-shot-image-classification-on-imagenet-fs-5 | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 75.45 |
| few-shot-image-classification-on-imagenet-fs-6 | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 81.12 |
| few-shot-image-classification-on-imagenet-fs-7 | KGTN-ens (ResNet-50, h+g, max) | Top-5 Accuracy (%): 83.46 |