Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Task2Sim: Towards Effective Pre-Training and Transfer From Synthetic Data
Check 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 |
|---|---|---|---|
| Choose simulation parameters for each task using Task2Sim instead of random parameters | Synthetic pretraining for ResNet-50 · Seen tasks (average, 237 classes / 100k images) | finetuning accuracy: 72.18 -> 76.87 | Helped |
| Choose simulation parameters for each task using Task2Sim instead of domain randomization | Synthetic pretraining for ResNet-50 · Seen tasks (average, 237 classes / 100k images) | finetuning accuracy: 68.51 -> 76.87 | Helped |
| Choose simulation parameters for each task using Task2Sim instead of Imagenet pretraining | ResNet-50 · Seen tasks (average, 237 classes / 100k images) | finetuning accuracy: 77.61 -> 76.87 | Hurt |
| Use Task2Sim-predicted parameters on unseen tasks instead of random parameters | Synthetic pretraining for ResNet-50 · Unseen tasks (average, 237 classes / 100k images) | finetuning accuracy: 83.49 -> 88.77 | Helped |
| Use Task2Sim-predicted parameters on unseen tasks instead of domain randomization | Synthetic pretraining for ResNet-50 · Unseen tasks (average, 237 classes / 100k images) | finetuning accuracy: 78.15 -> 88.77 | Helped |
| Use Task2Sim-predicted parameters instead of Imagenet pretraining | ResNet-50 · Unseen tasks (average, 237 classes / 100k images) | finetuning accuracy: 87.84 -> 88.77 | Helped |
| Add lighting variation (ablation) | Synthetic pretraining configuration · EuroSAT | accuracy: 87.01 -> 88.57 | Helped |
| Add blur variation after lighting variation (ablation) | Synthetic pretraining configuration · SVHN | accuracy: 32.36 -> 35.58 | Helped |
| Add material variation after lighting and blur variations (ablation) | Synthetic pretraining configuration · SVHN | accuracy: 35.58 -> 44.84 | Helped |
| Add material variation after lighting and blur variations (ablation) | Synthetic pretraining configuration · EuroSAT | accuracy: 90.20 -> 84.54 | Hurt |
| Add background variation after the preceding parameter variations (ablation) | Synthetic pretraining configuration · Sketch | accuracy: 30.81 -> 14.60 | Hurt |
Closest papers by meaning
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.87)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.87)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · similarity 0.86)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.86)
- BEHAVIOR Vision Suite: Customizable Dataset Generation via Simulation (CVPR 2024 · similarity 0.86)
- Personalized Representation from Personalized Generation (ICLR 2025 · similarity 0.86)
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.