摘要
本文研究深度直推学习（deep transductive learning），提出了一种名为 TransBoost 的优化方法，用于对任意深度神经网络模型进行微调，以提升其在训练阶段即可获得的任意（未标注）测试集上的性能表现。TransBoost 的设计灵感来源于大间隔原则（large margin principle），具有高效且易于使用的特点。实验结果表明，该方法在多种主流网络架构上均显著提升了 ImageNet 图像分类任务的性能，涵盖 ResNets、MobileNetV3-L、EfficientNetB0、ViT-S 以及 ConvNeXt-T 等模型，实现了当前最优的直推学习性能。此外，我们还验证了 TransBoost 在多种不同图像分类数据集上的广泛有效性。TransBoost 的开源实现已发布于：https://github.com/omerb01/TransBoost。
代码仓库
omerb01/transboost
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| image-classification-on-cifar-10 | TransBoost-ResNet50 | Percentage correct: 97.61 |
| image-classification-on-dtd | TransBoost-ResNet50 | Accuracy: 76.49 |
| image-classification-on-fgvc-aircraft | TransBoost-ResNet50 | Accuracy: 83.80% |
| image-classification-on-flowers-102 | TransBoost-ResNet50 | Accuracy: 97.85% |
| image-classification-on-food-101-1 | TransBoost-ResNet50 | Accuracy (%): 84.30 |
| image-classification-on-imagenet | TransBoost-ResNet50-StrikesBack | Number of params: 25.56M Top 1 Accuracy: 81.15% |
| image-classification-on-imagenet | TransBoost-ResNet152 | Number of params: 60.19M Top 1 Accuracy: 80.64% |
| image-classification-on-imagenet | TransBoost-Swin-T | Number of params: 71.71M Top 1 Accuracy: 82.16% |
| image-classification-on-imagenet | TransBoost-ResNet50 | Top 1 Accuracy: 79.03% |
| image-classification-on-imagenet | TransBoost-ResNet18 | Number of params: 11.69M Top 1 Accuracy: 73.36% |
| image-classification-on-imagenet | TransBoost-ResNet34 | Number of params: 21.8M Top 1 Accuracy: 76.70% |
| image-classification-on-imagenet | TransBoost-ResNet101 | Number of params: 44.55M Top 1 Accuracy: 79.86% |
| image-classification-on-imagenet | TransBoost-EfficientNetB0 | Number of params: 5.29M Top 1 Accuracy: 78.60% |
| image-classification-on-imagenet | TransBoost-MobileNetV3-L | Number of params: 5.48M Top 1 Accuracy: 76.81% |
| image-classification-on-imagenet | TransBoost-ViT-S | Number of params: 22.05M Top 1 Accuracy: 83.67% |
| image-classification-on-imagenet | TransBoost-ConvNext-T | Number of params: 28.59M Top 1 Accuracy: 82.46% |
| image-classification-on-stanford-cars | TransBoost-ResNet50 | Accuracy: 90.80% |
| image-classification-on-sun397 | TransBoost-ResNet50 | Accuracy: 95.94% |