Research map / Architectures and efficiency / CNN design
Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets
Check 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 depthwise separable convolutions with BSConv-U | MobileNetV1 · ImageNet | accuracy: 70.8 -> 71.5 | Helped |
| Replace the bottleneck convolution design with BSConv-S | MobileNetV2 · ImageNet | accuracy: 69.7 -> 69.8 | Helped |
| Replace the bottleneck convolution design with BSConv-S | MobileNetV3-large · ImageNet | accuracy: 71.5 -> 71.5 | No clear effect |
| Replace regular convolutions with BSConv-U and increase network depth at similar parameter count | ResNet · CIFAR100 | accuracy: 67.7 -> 70.8 | Helped |
| Replace regular convolutions with BSConv-U and increase network width | WideResNet · CIFAR100 | accuracy: 75.5 -> 77.6 | Helped |
| Replace regular convolutions with BSConv-U | VGG-16 (BN) · Stanford Dogs | accuracy: 60.5 -> 62.4 | Helped |
| Replace regular convolutions with BSConv-U | DenseNet-121 · Stanford Dogs | accuracy: 56.9 -> 59.4 | Helped |
| Replace depthwise separable convolutions with BSConv-U | Xception · Stanford Dogs | accuracy: 59.6 -> 64.3 | Helped |
| Use BSConv-S and orthonormal regularization in the subspace transform | EfficientNet-B0 · Stanford Dogs | accuracy: 54.7 -> 61.2 | Helped |
| Use BSConv-S and orthonormal regularization in the subspace transform | MnasNet · Stanford Dogs | accuracy: 54.8 -> 59.8 | Helped |
| Add orthonormal regularization to the BSConv-S subspace components (ablation) | MobileNetV3-large · Stanford Dogs | accuracy: over 5 percentage points | Helped |
| Use a smaller orthonormal-regularization weight (ablation) | MobileNetV3-large BSConv-S · Stanford Dogs | accuracy | No clear effect |
Closest papers by meaning
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.88)
- MobileODE: An Extra Lightweight Network (NeurIPS 2025 · similarity 0.88)
- KernelDNA: Cross-Layer Kernel Sharing via Decoupled Neural Adapters (NeurIPS 2026 · similarity 0.88)
- FreeNet: Liberating Depth-Wise Separable Operations for Building Faster Mobile Vision Architectures (AAAI 2025 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.86)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · 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.