Research map / Architectures and efficiency / CNN design
CondConv: Conditionally Parameterized Convolutions for Efficient Inference
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 standard convolutions with 8-expert CondConv | MobileNetV1 (1.0x) · ImageNet validation | Top-1 accuracy (%): 71.9 -> 73.7 | Helped |
| Replace standard convolutions with 8-expert CondConv | MobileNetV2 (1.0x) · ImageNet validation | Top-1 accuracy (%): 71.6 -> 74.6 | Helped |
| Replace standard convolutions with 8-expert CondConv | MnasNet-A1 · ImageNet validation | Top-1 accuracy (%): 74.9 -> 76.2 | Helped |
| Replace standard convolutions with 8-expert CondConv | ResNet-50 · ImageNet validation | Top-1 accuracy (%): 77.7 -> 78.6 | Helped |
| Replace standard convolutions with 8-expert CondConv | EfficientNet-B0 · ImageNet validation | Top-1 accuracy (%): 77.2 -> 78.3 | Helped |
| Replace standard convolutions with 8-expert CondConv | MobileNetV1 SSD300 · COCO minival | mAP: 14.4 -> 18.0; 18.2 -> 21.0; 20.3 -> 22.4 | Helped |
| Reuse routing weights from block 7 in all subsequent layers instead of routing at multiple depths (ablation) | CondConv-MobileNetV1 (0.25x) · ImageNet validation | Top-1 accuracy (%): 62.0 -> 56.5 | Hurt |
| Share routing weights between every other layer (ablation) | CondConv-MobileNetV1 (0.25x) · ImageNet validation | Top-1 accuracy (%): 62.0 -> 62.5 | Helped |
| Add a small hidden layer to the routing function (ablation) | CondConv-MobileNetV1 (0.25x) · ImageNet validation | Top-1 accuracy (%): 62.0 -> 57.7 | Hurt |
| Add a medium hidden layer to the routing function (ablation) | CondConv-MobileNetV1 (0.25x) · ImageNet validation | Top-1 accuracy (%): 62.0 -> 62.2 | No clear effect |
| Add a large hidden layer to the routing function (ablation) | CondConv-MobileNetV1 (0.25x) · ImageNet validation | Top-1 accuracy (%): 62.0 -> 54.1 | Hurt |
| Make routing hierarchical by concatenating previous-layer routing weights (ablation) | CondConv-MobileNetV1 (0.25x) · ImageNet validation | Top-1 accuracy (%): 62.0 -> 60.3 | Hurt |
Closest papers by meaning
- Omni-Dimensional Dynamic Convolution (ICLR 2022 · similarity 0.89)
- Convolutional Normalization: Improving Deep Convolutional Network Robustness and Training (NeurIPS 2021 · similarity 0.87)
- EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (ICML 2019 · similarity 0.87)
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes (ICLR 2022 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.85)
- Verification of Neural Networks Against Convolutional Perturbations via Parameterised Kernels (AAAI 2025 · similarity 0.85)
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.