Research map / Label-efficient and robust learning / Semi-supervised learning
Tied-Augment: Controlling Representation Similarity Improves Data Augmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| add Tied-Augment (L2 feature similarity term between two augmented views) to RandAugment | ResNet-50 on ImageNet, 360 epochs · ImageNet | top-1 accuracy: +2.0 (77.6 to 79.6) | Helped |
| add Tied-Augment to RandAugment | ResNet-50 on ImageNet, 90 epochs · ImageNet | top-1 accuracy: +1.9 (76.3 to 78.2) | Helped |
| add Tied-Augment to Crop-Flip | ResNet-50 on ImageNet · ImageNet | top-1 accuracy: +0.7 at 90 epochs, +0.6 at 360 epochs | Helped |
| add Tied-Augment to mixup | ResNet-50 on ImageNet · ImageNet | top-1 accuracy: +0.9 (77.9 to 78.8) | Helped |
| add Tied-Augment to RandAugment | WRN-28-10 on CIFAR-10/CIFAR-100/CIFAR-4K · CIFAR-10, CIFAR-100, CIFAR-4K | test accuracy: +0.8 (CIFAR-10), +1.7 (CIFAR-100), +3.4 (CIFAR-4K) | Helped |
| add Tied-Augment to Crop-Flip | CIFAR-10/100/4K baselines · CIFAR-10, CIFAR-100, CIFAR-4K | test accuracy: +0.4 to +1.5 across datasets and widths | Helped |
| add Tied-Augment to RandAugment in few-epoch training | ResNet/WRN on CIFAR-10 and CIFAR-100, 1-10 epochs · CIFAR-10/CIFAR-100 few-epoch | test accuracy: up to 6.7% over baseline per paper | Helped |
| add Tied-Augment to RandAugment when finetuning ImageNet-pretrained ResNet-50 | finetuning on Cars, Flowers, Pets, Aircraft, CIFAR-10 · finetuning 2-50 epochs | test accuracy: up to 10.7% per paper | Helped |
| add Tied-Augment to FixMatch | FixMatch semi-supervised CIFAR-10 (4k labels) and CIFAR-100 (10k labels) · CIFAR-10 / CIFAR-100 | accuracy: +0.4 (CIFAR-10), +0.5 (CIFAR-100) | Helped |
| add Tied-Augment to SAM | WRN-28-10 with RandAugment · CIFAR-10 / CIFAR-100 | accuracy: +0.4 (CIFAR-10), +0.3 (CIFAR-100) | Helped |
| use two differently augmented RandAugment views without the similarity loss (vs single RandAugment) (ablation) | ResNet-50 ImageNet / CIFAR-10 / Stanford-Cars 2 epochs · ImageNet, CIFAR-10, Stanford-Cars | accuracy: ImageNet +0.9 (77.6 to 78.5); CIFAR-10 +0.4; Cars -4.5 (58.7 to 54.2) | Helped |
| add L2 feature similarity loss to two-view different-policy training (ablation) | ResNet-50 ImageNet / CIFAR-10 / Stanford-Cars 2 epochs · ImageNet, CIFAR-10, Stanford-Cars | accuracy: ImageNet +1.1 (78.5 to 79.6); CIFAR-10 +0.4; Cars +15.2 (54.2 to 69.4) | Helped |
Closest papers by meaning
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.88)
- Improving Transferability of Representations via Augmentation-Aware Self-Supervision (NeurIPS 2021 · similarity 0.88)
- KeepAugment: A Simple Information-Preserving Data Augmentation Approach (CVPR 2021 · similarity 0.87)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.86)
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · similarity 0.85)
- Understanding Augmentation-based Self-Supervised Representation Learning via RKHS Approximation and Regression (ICLR 2024 · similarity 0.85)
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.