Research map / Architectures and efficiency / CNN design
FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes
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 |
|---|---|---|---|
| learn kernel sizes with FlexConv rather than use fixed-size convolutions (ablation) | FlexNet-16 · CIFAR-10 | accuracy: 78.0 ± 0.3 -> 92.2 ± 0.1 | Helped |
| use MAGNet kernels instead of N-Jet kernels (ablation) | FlexNet-16 · CIFAR-10 | accuracy: 91.7 ± 0.1 -> 92.2 ± 0.1 | Helped |
| replace SIREN with MAGNet parameterization (ablation) | FlexNet-16 · CIFAR-10 | accuracy: 89.0 ± 0.3 -> 92.2 ± 0.1 | Helped |
| use anisotropic MAGNet rather than isotropic Gabor parameterization (ablation) | FlexNet-16 · CIFAR-10 | accuracy: 91.9 ± 0.2 -> 92.2 ± 0.1 | Helped |
| add aliasing regularization on MAGNet frequencies (ablation) | FlexNet-16 · CIFAR-10 at 32 px after upsampling from 16 px | accuracy difference: -34.4 ± 14.3 -> -3.8 ± 2.0 | Helped |
| include the FlexConv mask frequency in aliasing regularization rather than regularize MAGNet frequency only (ablation) | FlexNet-16 · CIFAR-10 at 32 px after upsampling from 16 px | accuracy difference: -3.8 ± 2.0 vs -3.3 ± 0.3 | No clear effect |
| increase FlexTCN depth from 2 to 6 blocks (ablation) | FlexTCN · sCIFAR10 | test accuracy: 78.99 -> 80.82 | Helped |
| increase FlexTCN depth from 2 to 6 blocks (ablation) | FlexTCN · SC_raw | test accuracy: 88.03 -> 91.73 | Helped |
| use the improved layer-dependent MAGNet initialization (ablation) | MAGNet · Kodak | PSNR: 25.791 -> 25.893 | Helped |
| train at 16×16 instead of 32×32 resolution (ablation) | FlexNet-7 · CIFAR-10 | training time per batch: 222ms (± 9ms) -> 179ms (± 7ms) | Helped |
| initialize sequential mask centers at the first element to prioritize far-past information (ablation) | FlexTCN-2 · npCIFAR10 | accuracy: 68.33% acc. | Helped |
Closest papers by meaning
- EffConv: Efficient Learning of Kernel Sizes for Convolution Layers of CNNs (AAAI 2023 · similarity 0.90)
- The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods (ICLR 2021 · similarity 0.89)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.88)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.87)
- Reparameterized Multi-Resolution Convolutions for Long Sequence Modelling (NeurIPS 2024 · similarity 0.86)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 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.