Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Soupervision: Cooking Model Soups without Labels
Anthony Fuller, James R Green, Evan Shelhamer
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 |
|---|---|---|---|
| mix ingredients inter-trained with different SSL algorithms instead of supervised-only ingredients | uniform-mix soup · ImageNet-C | top-1 accuracy: 30.91 -> 32.34 | Improved |
| mix ingredients inter-trained with different SSL algorithms instead of supervised-only ingredients | uniform-mix soup · LAION-C | top-1 accuracy: 22.05 -> 22.92 | Improved |
| mix ingredients inter-trained with different SSL algorithms | uniform-mix soup · ImageNet-Val | top-1 accuracy: 79.12 -> 79.09 | No clear effect |
| add SSL inter-training on corrupted test distribution (disjoint samples) on ImageNet-C | Self-Soup · ImageNet-C | top-1 accuracy: 32.34 -> 35.72 | Improved |
| add SSL inter-training on LAION-C distribution (disjoint samples) | Self-Soup · LAION-C | top-1 accuracy: 22.92 -> 29.48 | Improved |
| inter-train on the test samples themselves rather than disjoint samples | Self-Soup on IN-C · ImageNet-C | top-1 accuracy: 35.72 -> 36.02 | Improved |
| inter-train on the test samples themselves rather than disjoint samples | Self-Soup on LA-C · LAION-C | top-1 accuracy: 29.48 -> 30.32 | Improved |
| inter-train on both IN-C and LAION-C rather than a single corruption set | Self-Soup · ImageNet-C | top-1 accuracy: 35.72 -> 34.43 | Hurt |
| replace continued SSL (single algorithm) soup with Self-Soup (varying algorithms) | soup on LAION-C · LAION-C | top-1 accuracy: 27.43 -> 29.48 | Improved |
| add test-time adaptation (SAR) to Self-Soup on ImageNet-C | Self-Soup on ImageNet-C · ImageNet-C | top-1 accuracy: 35.72 -> 37.50 | Improved |
| use Self-Soup instead of supervised-only soup on mini-VTAB-C | uniform mix · mini-VTAB-C mean | top-1 accuracy: 34.6 -> 35.1 | Improved |
| use Self-Soup instead of supervised-only soup on noise corruptions | Self-Soup · VTAB noise corruptions | top-1 accuracy: -1% versus the supervised-only soup | Hurt |
Closest papers by meaning
- LT-Soups: Bridging Head and Tail Classes via Subsampled Model Soups (NeurIPS 2025 · similarity 0.85)
- Seasoning Model Soups for Robustness to Adversarial and Natural Distribution Shifts (CVPR 2023 · similarity 0.85)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.83)
- Self-Supervised Dataset Distillation for Transfer Learning (ICLR 2024 · similarity 0.83)
- Look Ahead or Look Around? A Theoretical Comparison Between Autoregressive and Masked Pretraining (ICML 2024 · similarity 0.83)
- Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation (NeurIPS 2025 · similarity 0.82)
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.
