Research map / Detection and segmentation / CNN features
ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks
Shuxuan Guo, Jose M. Alvarez, Mathieu Salzmann
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 |
|---|---|---|---|
| expand convolutional kernels (CK) instead of only FC layers (ablation) | SmallNet 7x7 · CIFAR-10 | Top-1 accuracy: 78.64 -> 80.27 | Improved |
| expand conv layers with CL (ablation) | SmallNet 7x7 · CIFAR-10 | Top-1 accuracy: 78.63 -> 78.47 | No clear effect |
| add FC expansion to CK (ablation) | ExpandNet-CK · CIFAR-100 | Top-1 accuracy: 48.55 -> 48.62 | Improved |
| add knowledge distillation to ExpandNet-CK+FC | ExpandNet-CK+FC · CIFAR-100 | Top-1 accuracy: 48.62 -> 49.13 | Improved |
| add nonlinear-counterpart initialization (ablation) | ExpandNet-CK+FC · CIFAR-100 | Top-1 accuracy: 48.62 -> 49.82 | Improved |
| add nonlinear-counterpart initialization (ablation) | YOLO-LITE ExpandNet-CL · PASCAL VOC2007 test | mAP: 30.97 -> 35.14 | Improved |
| expand all 3x3 convs with CL | MobileNet · ImageNet val | Top-1 accuracy: 66.48 -> 69.40 | Improved |
| expand non-residual 3x3 convs with CL | MobileNetV2 · ImageNet val | Top-1 accuracy: 63.75 -> 65.62 | Improved |
| expand with CL | ShuffleNetV2 0.5x · ImageNet val | Top-1 accuracy: 56.89 -> 57.38 | Improved |
| train baseline longer to match ExpandNet training time (ablation) | MobileNet · CIFAR-10 | Top-1 accuracy: 89.61 -> 88.87 | Hurt |
| initialize compact net with contracted ExpandNet instead of training ExpandNet (ablation) | MobileNetV2 · ImageNet val | Top-1 accuracy: 63.75 -> 63.07 | Hurt |
| lower expansion rate r from 4 to 0.25 (ablation) | ExpandNet-CK · CIFAR-10 | Top-1 accuracy: 80.27 -> 72.32 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MobileNetV2 — MobileNetV2: Inverted Residuals and Linear Bottlenecks (CVPR 2018) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Deep Expander Networks: Efficient Deep Networks from Graph Theory (ECCV 2018 · similarity 0.89)
- Learning Efficient Convolutional Networks Through Network Slimming (ICCV 2017 · similarity 0.87)
- Network Expansion for Practical Training Acceleration (CVPR 2023 · similarity 0.87)
- Go Wider Instead of Deeper (AAAI 2022 · similarity 0.87)
- Learning Compact Neural Networks with Regularization (ICML 2018 · similarity 0.86)
- Expand-and-Cluster: Parameter Recovery of Neural Networks (ICML 2024 · 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.
