Research map / Architectures and efficiency / Neural architecture search
Learning Strides in Convolutional Neural Networks
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 strides through backpropagation instead of fixing them | ResNet-18 · CIFAR10, mean across stride initializations | accuracy: 87.7 ± 3.4 -> 92.4 ± 0.2 | Helped |
| learn strides through backpropagation instead of fixing them | ResNet-18 · CIFAR100, mean across stride initializations | accuracy: 59.1 ± 6.7 -> 73.5 ± 0.3 | Helped |
| learn strides through backpropagation instead of fixing them | ResNet-18 · ImageNet validation, mean across stride initializations | top-1 accuracy: 65.53 ± 4.49 -> 69.28 ± 0.50 | Helped |
| replace fixed strided convolutions with learnable spectral downsampling | 2D-CNN · Audio classification, single-task mean | accuracy: 85.0 ± 9.3 -> 88.3 ± 8.7 | Helped |
| replace fixed strided convolutions with learnable spectral downsampling | 2D-CNN · Audio classification, multi-task mean | accuracy: 83.5 ± 10.0 -> 85.0 ± 8.9 | Helped |
| share one stride across time and frequency rather than learning a stride per axis (ablation) | DiffStride · Multi-task audio classification (mean) | accuracy: 85.0 ± 8.9 -> 84.3 ± 9.2 | Hurt |
| share one stride across time and frequency rather than learning a stride per axis (ablation) | DiffStride · Birdsong detection, multi-task | accuracy: 78.6 ± 0.5 -> 79.7 ± 1.0 | Helped |
| share one stride across time and frequency rather than learning a stride per axis (ablation) | DiffStride · Music instrument, multi-task | accuracy: 73.0 ± 0.8 -> 70.7 ± 0.5 | Hurt |
| share one stride across time and frequency rather than learning a stride per axis (ablation) | DiffStride · Music pitch, multi-task | accuracy: 89.9 ± 0.3 -> 86.7 ± 0.5 | Hurt |
| share one stride across time and frequency rather than learning a stride per axis (ablation) | DiffStride · Speech commands, multi-task | accuracy: 86.2 ± 0.8 -> 86.8 ± 0.1 | Helped |
| apply the complexity regularizer with λ = 10 (ablation) | DiffStride · ImageNet validation | top-1 accuracy: 58.57% vs 24.54% | Mixed |
| replace strided convolutions with DiffStride | DenseNet-BC · CIFAR10 | accuracy: 91.9 ± 0.3 -> 91.6 ± 0.4 | No clear effect |
Closest papers by meaning
- Making Convolutional Networks Shift-Invariant Again (ICML 2019 · similarity 0.86)
- Deep Pyramidal Residual Networks (CVPR 2017 · similarity 0.84)
- Generalized Deep Image to Image Regression (CVPR 2017 · similarity 0.84)
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes (ICLR 2022 · similarity 0.84)
- Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling (ICLR 2023 · similarity 0.84)
- Towards Learning Convolutions from Scratch (NeurIPS 2020 · 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.