Research map / Architectures and efficiency / Pruning and efficient tuning
TransBoost: Improving the Best ImageNet Performance using Deep Transduction
Omer Belhasin, Guy Bar-Shalom, Ran El-Yaniv
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace separation loss with reciprocal (pull-together) loss (ablation) | TransBoost ResNet50 · ImageNet-1k val | top-1 accuracy (%): 79.03 -> 74.64 | Hurt |
| add reciprocal pull-together loss to separation loss (ablation) | TransBoost ResNet50 · ImageNet-1k val | top-1 accuracy (%): 79.03 -> 79.00 | No clear effect |
| add knowledge distillation stage (SimCLRv2 vs SimCLR) | SimCLR ResNet50 transduction · ImageNet-1k val | top-1 accuracy (%): 75.95 -> 74.68 | Hurt |
| increase test set size with full training set | TransBoost ResNet50 · ImageNet subsets | top-1 gain (%): +1.80 -> +2.88 | Improved |
| increase test set size with 5% training set | TransBoost ResNet50 · ImageNet 5% train | top-1 gain (%): +3.42 -> +1.97 | Hurt |
| apply TransBoost to a larger ResNet | TransBoost · ImageNet-1k val | top-1 gain (%): +3.60 -> +2.31 | Hurt |
| use ViT-S instead of Swin-T | TransBoost · ImageNet-1k val | top-1 gain (%): +0.78 -> +2.28 | Improved |
| apply TransBoost on top of stronger recipe ResNet50-StrikesBack | TransBoost · ImageNet-1k val | top-1 accuracy (%): 80.38 -> 81.15 | Improved |
| apply TransBoost to ConvNext-T | TransBoost · ImageNet-1k val | top-1 accuracy (%): 82.05 -> 82.46 | Improved |
| apply TIM-GD transductive few-shot method | ResNet50 · ImageNet-1k val | top-1 accuracy (%): 76.15 -> 71.51 | Hurt |
| apply Ent-min transductive few-shot method | ResNet50 · ImageNet-1k val | top-1 accuracy (%): 76.15 -> 77.55 | Improved |
| apply TransBoost on CIFAR-10 | ResNet50 · CIFAR-10 | top-1 accuracy (%): 97.60 -> 97.61 | No clear effect |
Closest papers by meaning
- Transductive Semi-Supervised Deep Learning using Min-Max Features (ECCV 2018 · similarity 0.84)
- Boosting Vision-Language Models with Transduction (NeurIPS 2024 · similarity 0.83)
- HydraNets: Specialized Dynamic Architectures for Efficient Inference (CVPR 2018 · similarity 0.82)
- LEARNING TO PROPAGATE LABELS: TRANSDUCTIVE PROPAGATION NETWORK FOR FEW-SHOT LEARNING (ICLR 2019 · similarity 0.82)
- Transductive Zero-Shot and Few-Shot CLIP (CVPR 2024 · similarity 0.81)
- Efficient ConvBN Blocks for Transfer Learning and Beyond (ICLR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
