Research map / Label-efficient and robust learning / Recognition and long-tail
Synbols: Probing Learning Algorithms with Synthetic Datasets
Alexandre Lacoste, Pau Rodríguez López, Frederic Branchaud-Charron, Parmida Atighehchian, Massimo Caccia, Issam Hadj Laradji, Alexandre Drouin, Matthew Craddock, Laurent Charlin, David Vázquez
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 |
|---|---|---|---|
| add data augmentation (ResNet-12 -> ResNet-12+) (ablation) | ResNet-12 · Synbols Less Variations 100k | accuracy: 39.41 -> 57.42 | Improved |
| add data augmentation (WRN-28-4 -> WRN-28-4+) (ablation) | WRN-28-4 · Synbols Less Variations 100k | accuracy: 23.10 -> 68.42 | Improved |
| replace flattening with global average pooling (ablation) | Conv-4 · Synbols Korean 100k | accuracy: 1.62 -> 33.58 | Improved |
| add camouflage to the background/foreground | WRN-28-4+ · Synbols Default 100k | accuracy: 97.41 -> 95.55 | Hurt |
| increase training set from 100k to 1M | MLP · Synbols Default | accuracy: 14.83 -> 63.05 | Improved |
| stratified rotation split instead of i.i.d. | WRN-28-4+ · Synbols Default | accuracy: 97.41 -> 91.81 | Hurt |
| stratified rotation split instead of i.i.d. | Conv-4 Flat · Synbols Default | accuracy: 68.51 -> 5.54 | Hurt |
| stratified font split instead of i.i.d. | ResNet-12 · Synbols Default | accuracy: 95.43 -> 94.62 | No clear effect |
| stratified x-translation split instead of i.i.d. | Conv-4 Flat · Synbols Default | accuracy: 68.51 -> 44.78 | Hurt |
| stratified x-translation split instead of i.i.d. | ResNet-12+ · Synbols Default | accuracy: 97.16 -> 96.84 | No clear effect |
| test on unseen characters for font prediction | WRN-28-4+ · Synbols Less Variations | accuracy: 68.42 -> 35.41 | Hurt |
| use P-Entropy instead of BALD query function | active learning VGG-16 · 10% Missing | negative log likelihood: 1.24 -> 2.02 | Hurt |
Closest papers by meaning
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.84)
- Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion (ICML 2025 · similarity 0.83)
- CIFAR-10-Warehouse: Broad and More Realistic Testbeds in Model Generalization Analysis (ICLR 2024 · similarity 0.83)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.83)
- CSGO: Content-Style Composition in Text-to-Image Generation (NeurIPS 2025 · similarity 0.83)
- Alchemist: Turning Public Text-to-Image Data into Generative Gold (NeurIPS 2025 · similarity 0.83)
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.
