Research map / Label-efficient and robust learning / Domain adaptation and generalization
Robust and Generalizable Visual Representation Learning via Random Convolutions
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 |
|---|---|---|---|
| Use multi-scale random convolutions to augment training images | CNN digit classifier · MNIST-M | accuracy: 67.94 -> 87.76 | Helped |
| Use random-convolution augmentation | CNN digit classifier · MNIST-C | accuracy: 88.20 (2.10) -> 91.62 (0.77) | Helped |
| Use multi-scale random-convolution mixing | AlexNet Deep-All · PACS Sketch | accuracy: 55.28 (4.71) -> 69.74 (0.66) | Helped |
| Use random-convolution image augmentation with consistency loss | AlexNet · ImageNet-Sketch | Top1 accuracy: 10.28 -> 18.09 | Helped |
| Pretrain on ImageNet with RandConv and fine-tune on PACS with RandConv | AlexNet · PACS Sketch | accuracy: 71.74 (0.43) -> 76.11 (0.43) | Helped |
| Pretrain on ImageNet with RandConv | AlexNet Deep-All fine-tuned on PACS · PACS Photo | accuracy: 86.77 (0.42) -> 84.48 (0.52) | Hurt |
| Combine RandConv mixing with MultiAug | CNN digit classifier · MNIST-C | accuracy: 91.62 (0.77) -> 91.40 (0.93) | Hurt |
| Increase maximum filter size from 1 to 7 (ablation) | RC_mix · MNIST digits DG-Avg | accuracy: 69.26 (1.35) -> 71.19 (1.21) | Helped |
| Increase maximum filter size from 1 to 7 (ablation) | RC_img, p=0.5 · MNIST digits DG-Avg | accuracy: 70.55 (0.15) -> 71.34 (0.61) | Helped |
| Set original-image fraction p to 0.5 rather than p=0 (ablation) | RC_img with size-1 filters · MNIST digits DG-Avg | accuracy: 69.32 (1.19) -> 70.63 (0.42) | Helped |
| Add consistency loss with λ=10 (ablation) | RC_mix 1–7 · MNIST digits DG-Avg | accuracy: 71.19 (1.21) -> 72.88 (0.58) | Helped |
| Add consistency loss with λ=10 (ablation) | RC_img 1–7, p=0.5 · MNIST digits DG-Avg | accuracy: 71.34 (0.61) -> 72.02 (1.15) | Helped |
Closest papers by meaning
- Progressive Random Convolutions for Single Domain Generalization (CVPR 2023 · similarity 0.91)
- Learning Robust Global Representations by Penalizing Local Predictive Power (NeurIPS 2019 · similarity 0.89)
- A Powerful Generative Model Using Random Weights for the Deep Image Representation (NeurIPS 2016 · similarity 0.88)
- Informative Dropout for Robust Representation Learning: A Shape-bias Perspective (ICML 2020 · similarity 0.87)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.86)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 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.