Research map / Optimization / Optimizers
A Continuous Mapping For Augmentation Design
Keyu Tian, Chen Lin, Ser-Nam Lim, Wanli Ouyang, Puneet K. Dokania, Philip Torr
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| keep only color adjustment augmentations (remove warping and filtering) (ablation) | MCMC-Aug · CIFAR-10 WRN-40-2 | top-1 error: 2.96 -> 3.32 | Hurt |
| keep only image warping augmentations (ablation) | MCMC-Aug · CIFAR-10 WRN-40-2 | top-1 error: 2.96 -> 3.45 | Hurt |
| keep only image filtering augmentations (ablation) | MCMC-Aug · CIFAR-10 WRN-40-2 | top-1 error: 2.96 -> 3.61 | Hurt |
| stop SGLD sample collection at 75% of epochs (ablation) | MCMC-Aug · CIFAR-10 WRN-40-2 | top-1 error: 2.96 -> 3.06 | Hurt |
| stop SGLD sample collection at 50% of epochs (ablation) | MCMC-Aug · CIFAR-10 WRN-40-2 | top-1 error: 2.96 -> 3.15 | Hurt |
| stop SGLD sample collection at 25% of epochs (ablation) | MCMC-Aug · CIFAR-10 WRN-40-2 | top-1 error: 2.96 -> 3.20 | Hurt |
| continuous random Gaussian policy (std=0.5) instead of discrete AutoAugment random policy | augmentation space · CIFAR-10 WRN-40-2 | top-1 error: 4.08 -> 3.25 | Improved |
| continuous random policy (std=0.5) instead of searched AutoAugment | augmentation space · CIFAR-10 WRN-40-2 | top-1 error: 3.7 -> 3.25 | Improved |
| larger std 1.0 instead of 0.5 for random continuous policy | example random policy · CIFAR-10 WRN-40-2 | top-1 error: 3.25 -> 3.34 | Hurt |
| MCMC-Aug instead of no augmentation search (baseline) | WRN-40-2 · CIFAR-10 | top-1 error: 5.3 -> 2.96 | Improved |
| MCMC-Aug instead of AWS on WRN-28-10 | AWS · CIFAR-100 | top-1 error: 15.28 -> 15.64 | Hurt |
| MCMC-Aug instead of AWS on Shake-Shake | AWS · CIFAR-10 | top-1 error: 1.65 -> 1.53 | Improved |
Closest papers by meaning
- Faster AutoAugment: Learning Augmentation Strategies Using Backpropagation (ECCV 2020 · similarity 0.86)
- FreeAugment: Data Augmentation Search Across All Degrees of Freedom (ECCV 2024 · similarity 0.86)
- A Data-Augmentation Is Worth A Thousand Samples: Analytical Moments And Sampling-Free Training (NeurIPS 2022 · similarity 0.86)
- AIM: An Auto-Augmenter for Images and Meshes (CVPR 2022 · similarity 0.85)
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · similarity 0.84)
- Distribution-Aware Data Expansion with Diffusion Models (NeurIPS 2024 · similarity 0.84)
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.
