Command Palette
Search for a command to run...
摘要
我们提出了一种面向下游少样本分类任务的无监督嵌入自适应方法。基于深度神经网络在记忆之前优先学习泛化能力的发现，我们提出了早期阶段特征重建（Early-Stage Feature Reconstruction, ESFR）——一种新颖的自适应策略，该策略结合了特征重建机制与基于维度驱动的早停机制，能够有效识别出具有泛化能力的特征。在所有标准设置下，将ESFR融入基线方法均能持续提升其性能，包括最近提出的归纳式方法。当与归纳式方法联合使用时，ESFR在mini-ImageNet、tiered-ImageNet和CUB数据集上均取得了当前最优的性能表现，尤其在1-shot设置下，准确率相较此前最优方法提升了1.2%~2.0%。
代码仓库
movinghoon/ESFR
官方
tf
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| few-shot-image-classification-on-cub-200-5 | BD-CSPN + ESFR (ResNet-18) | Accuracy: 88.65 |
| few-shot-image-classification-on-cub-200-5-1 | BD-CSPN + ESFR (ResNet-18) | Accuracy: 82.68 |
| few-shot-image-classification-on-mini-2 | BD-CSPN + ESFR (WRN) | Accuracy: 76.84 |
| few-shot-image-classification-on-mini-2 | BD-CSPN + ESFR (ResNet-18) | Accuracy: 73.98 |
| few-shot-image-classification-on-mini-3 | BD-CSPN + ESFR (WRN) | Accuracy: 84.36 |
| few-shot-image-classification-on-mini-3 | BD-CSPN + ESFR (ResNet-18) | Accuracy: 82.32 |
| few-shot-image-classification-on-tiered | BD-CSPN + ESFR (ResNet-18) | Accuracy: 80.13 |
| few-shot-image-classification-on-tiered | BD-CSPN + ESFR (WRN) | Accuracy: 81.77 |
| few-shot-image-classification-on-tiered-1 | BD-CSPN + ESFR (WRN) | Accuracy: 87.61 |
| few-shot-image-classification-on-tiered-1 | BD-CSPN + ESFR (ResNet-18) | Accuracy: 86.34 |