Research map / Architectures and efficiency / RNNs and normalization
Swapout: Learning an ensemble of deep architectures
Saurabh Singh, Derek Hoiem, David Forsyth
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 |
|---|---|---|---|
| Replace ResNet v1 training with Swapout at the same width | ResNet v1, W × 1 · CIFAR-10 | error (%): 8.54 -> 8.27 | Improved |
| Replace ResNet v2 training with Swapout at the same width | ResNet v2 Ours, W × 1 · CIFAR-10 | error (%): 8.27 -> 7.97 | Improved |
| Train with Swapout instead of ResNet v2 at W × 2 | ResNet v2 Ours, W × 2 · CIFAR-10 | error (%): 6.54 -> 5.68 | Improved |
| Use stochastic inference rather than deterministic inference (ablation) | Swapout v2 (20) W × 2 · CIFAR-10 | error (%): 6.40 -> 5.68 | Improved |
| Increase width from W × 2 to W × 4 (ablation) | Swapout v2 (20) · CIFAR-10 | error (%): 5.68 -> 5.09 | Improved |
| Increase depth from 20 to 32 layers at W × 4 (ablation) | Swapout v2 · CIFAR-10 | error (%): 5.09 -> 4.76 | Improved |
| Use Swapout v2 (32) W × 4 instead of ResNet v2 (1001) | CIFAR-10 image classification · CIFAR-10 | error (%): 4.92 -> 4.76 | Improved |
| Train the 20-layer W × 2 model with Swapout rather than ResNet v2 | ResNet v2 Ours (20) W × 2 · CIFAR-100 | error (%): 28.08 -> 25.86 | Improved |
| Use a wider 32-layer Swapout model rather than ResNet v2 (1001) | CIFAR-100 image classification · CIFAR-100 | error (%): 22.71 -> 22.72 | No clear effect |
| Set θ1=θ2=0.5 across blocks instead of using Linear(1, 0.5) (ablation) | Swapout v2 (20) W × 2 · CIFAR-10 | stochastic error (%): 5.68 -> 6.69 | Hurt |
| Use θ1=0.2 and θ2=0.8 instead of Linear(1, 0.5) (ablation) | Swapout v2 (20) W × 2 · CIFAR-10 | stochastic error (%): 5.68 -> 7.63 | Hurt |
| Use θ1=θ2=Linear(0.5, 1) instead of Linear(1, 0.5) (ablation) | Swapout v2 (20) W × 2 · CIFAR-10 | stochastic error (%): 5.68 -> 6.52 | Hurt |
Closest papers by meaning
- Flipout: Efficient Pseudo-Independent Weight Perturbations on Mini-Batches (ICLR 2018 · similarity 0.87)
- Deep Layers as Stochastic Solvers (ICLR 2019 · similarity 0.86)
- Layer Free-Riding in Forward-Forward Networks: Real, Repairable, but Not Accuracy-Dominant (NeurIPS 2026 · similarity 0.85)
- Understanding and Exploring the Network with Stochastic Architectures (NeurIPS 2020 · similarity 0.85)
- Rethinking Layer-wise Model Merging through Chain of Merges (NeurIPS 2026 · similarity 0.84)
- SWAP-NAS: Sample-Wise Activation Patterns for Ultra-fast NAS (ICLR 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.
