Research map / Architectures and efficiency / RNNs and normalization
Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs
Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, Andrew G Wilson
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 |
|---|---|---|---|
| Use FGE instead of Snapshot Ensembles | VGG-16 · CIFAR-100, 1B budget | error rate (%): 26.4 ± 0.1 -> 25.7 ± 0.1 | Improved |
| Use FGE instead of Snapshot Ensembles | ResNet-164 · CIFAR-100, 1B budget | error rate (%): 20.9 ± 0.2 -> 20.2 ± 0.1 | Improved |
| Use FGE instead of Snapshot Ensembles | WRN-28-10 · CIFAR-100, 1B budget | error rate (%): 17.9 ± 0.2 -> 17.7 ± 0.2 | Improved |
| Use FGE instead of Snapshot Ensembles | VGG-16 · CIFAR-10, 1B budget | error rate (%): 6.57 ± 0.12 -> 6.48 ± 0.09 | Improved |
| Use FGE instead of Snapshot Ensembles | ResNet-164 · CIFAR-10, 1B budget | error rate (%): 4.66 ± 0.02 -> 4.54 ± 0.05 | Improved |
| Use FGE instead of Snapshot Ensembles | WRN-28-10 · CIFAR-10, 1B budget | error rate (%): 3.73 ± 0.04 -> 3.65 ± 0.1 | Improved |
| Fine-tune a pretrained model with FGE for 5 epochs | ResNet-50 · ImageNet ILSVRC-2012 | top-1 test error rate: 23.87 -> 23.31 | Improved |
| Replace the direct line segment between trained modes with a learned polygonal chain (ablation) | Mode-connecting ResNet-164 · CIFAR-100 | test error (%): 53.69 -> 24.92 | Improved |
| Use a learned Bezier curve instead of a direct segment (ablation) | Mode-connecting ResNet-164 · CIFAR-100 | test error (%): 53.69 -> 24.99 | Improved |
| Temperature-scale predictions from curve points (ablation) | 50-point curve ensemble · CIFAR-100 | test error rate: 21.03% -> 20.7% | Improved |
| Ensemble 50 points along the learned curve rather than only its endpoints (ablation) | ResNet-164 · CIFAR-100 | test error rate: 22.0% -> 21.03% | Improved |
| Use a curve ensemble rather than an independent three-network ensemble | ResNet-164 · CIFAR-100 | test error rate: 21.03% vs 21.01% | No clear effect |
Closest papers by meaning
- Loss Surface Simplexes for Mode Connecting Volumes and Fast Ensembling (ICML 2021 · similarity 0.89)
- Large Scale Structure of Neural Network Loss Landscapes (NeurIPS 2019 · similarity 0.86)
- The Goldilocks Zone: Towards Better Understanding of Neural Network Loss Landscapes (AAAI 2019 · similarity 0.86)
- On Linear Mode Connectivity of Mixture-of-Experts Architectures (NeurIPS 2025 · similarity 0.84)
- Exploring and Exploiting the Asymmetric Valley of Deep Neural Networks (NeurIPS 2024 · similarity 0.84)
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 2025 · 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.
