Research map / Architectures and efficiency / Neural architecture search
MaskConnect: Connectivity Learning by Gradient Descent
Karim Ahmed, Lorenzo Torresani
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 |
|---|---|---|---|
| add learned connectivity instead of fixed previous-block connectivity | ResNet-38 · CIFAR-100 | accuracy: 68.54 -> 70.40 | Improved |
| add learned connectivity instead of fixed previous-block connectivity | ResNet-110 · CIFAR-100 | accuracy: 71.21 -> 73.15 | Improved |
| learn connectivity rather than use random fixed connectivity with same fan-in (ablation) | ResNet-38 · CIFAR-100 | accuracy: 62.67 -> 70.40 | Improved |
| learn connectivity rather than random fixed connectivity (ablation) | ResNet-74 · CIFAR-100 | accuracy: 66.93 -> 72.81 | Improved |
| learn connectivity with K=4 instead of fixed full ResNeXt connectivity | ResNeXt {29,8,8} · CIFAR-100 | accuracy: 73.52 -> 75.89 | Improved |
| learn connectivity with K=4 instead of fixed full ResNeXt connectivity | ResNeXt {29,64,8} · CIFAR-100 | accuracy: 82.23 -> 84.05 | Improved |
| learn connectivity rather than random fixed with K=4 (ablation) | ResNeXt {29,8,8} · CIFAR-100 | accuracy: 72.85 -> 75.89 | Improved |
| use fan-in K=1 learned connectivity (prunes unused blocks) instead of full connectivity | ResNeXt {29,64,8} · CIFAR-100 | accuracy and test params: 82.23 -> 82.31 with 34.4M -> 20.5M params | Improved |
| add learned connectivity (K=10) to ResNet-50 | ResNet-50 · ImageNet | top-1 accuracy: 76.15 -> 78.09 | Improved |
| add learned connectivity (K=C/2) to ResNeXt | ResNeXt {50,4,32} · ImageNet | top-1 accuracy: 77.8 -> 79.1 | Improved |
| add learned connectivity (K=C/2) to ResNeXt | ResNeXt {101,4,64} · ImageNet | top-1 accuracy: 79.6 -> 79.8 | Improved |
| add learned connectivity K=4 instead of fixed full | ResNeXt {20,4,8} · Mini-ImageNet | top-1 accuracy: 62.12 -> 66.09 | Improved |
Closest papers by meaning
- Learning Connectivity of Neural Networks from a Topological Perspective (ECCV 2020 · similarity 0.89)
- SparseMask: Differentiable Connectivity Learning for Dense Image Prediction (ICCV 2019 · similarity 0.88)
- Discovering Neural Wirings (NeurIPS 2019 · similarity 0.87)
- Differentiable Dynamic Wirings for Neural Networks (ICCV 2021 · similarity 0.87)
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 2025 · similarity 0.84)
- TrAct: Making First-layer Pre-Activations Trainable (NeurIPS 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.
