Research map / Architectures and efficiency / CNN design
Deep Continuous 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 |
|---|---|---|---|
| Replace pixel 3x3 kernels with SRF filters in ODE blocks (DCN-ODE) | neural ODE CNN · CIFAR-10 small-data (52 images/class) | accuracy (%): 44.5 vs 41.7 for ODE-Net | Helped |
| Replace pixel 3x3 kernels with SRF filters in ODE blocks (DCN-ODE) | neural ODE CNN · CIFAR-10 small-data (1024 images/class) | accuracy (%): 79.7 vs 77.4 for ODE-Net | Helped |
| Replace pixel 3x3 kernels with SRF filters in ODE blocks (DCN-ODE) | neural ODE CNN · CIFAR-10 full training validation | accuracy (%): 89.5 vs 89.6 for ODE-Net | No clear effect |
| Replace pixel 3x3 kernels with SRF filters in ODE blocks (DCN-ODE) for reconstruction | encoder-decoder with ODE blocks · CIFAR-10 reconstruction validation | reconstruction loss (%): 17.1 vs 20.2 for ODE-Net (lower is better) | Helped |
| Use SRF filters in ResNet blocks only (no ODE) (ablation) | ResNet blocks · CIFAR-10 full training validation | accuracy (%): 88.3 vs 89.0 for ResNet-blocks | Hurt |
| Use SRF filters everywhere including downsampling layers (ResNet-SRF-full) (ablation) | ResNet blocks · CIFAR-10 full training validation | accuracy (%): 89.3 vs 89.0 for ResNet-blocks | Helped |
| Use SRF filters everywhere including downsampling layers (DCN-full) instead of DCN-ODE (ablation) | DCN · CIFAR-10 full training validation | accuracy (%): 89.2 vs 89.5 for DCN-ODE | Hurt |
| Learn scale sigma independently per input and output channel pair (DCN σji) (ablation) | DCN-ODE · CIFAR-10 full training validation | accuracy (%): 89.7 vs 89.5 for DCN-ODE | Helped |
| Meta-parametrize sigma as a function of depth (linear or quadratic) or both sigma and alpha (ablation) | DCN · CIFAR-10 validation | accuracy (%): 89.88 to 89.97 across variants; paper cites ~89.5 for static DCN-ODE | No clear effect |
| Scale input feature maps by contrast c and scale integration time by c at test time (ablation) | DCN and ODE-Net · CIFAR-10 validation, contrast robustness | validation accuracy and NFE: total NFE reduced from 102 to 60 at c=0.06 with <0.5% accuracy loss | Helped |
| Mask n x n center pixels at test time (pattern completion) | DCN vs ResNet-blocks and ODE-Net · CIFAR-10 validation with masks | validation accuracy: DCNs marginally more robust to zero-masking than baselines | Helped |
Closest papers by meaning
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.88)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.88)
- The Master Key Filters Hypothesis: Deep Filters Are General (AAAI 2025 · similarity 0.87)
- Continuity-Preserving Convolutional Autoencoders for Learning Continuous Latent Dynamical Models from Images (ICLR 2025 · similarity 0.86)
- Layer Compression of Deep Networks with Straight Flows (AAAI 2024 · similarity 0.86)
- The Quest for Universal Master Key Filters in DS-CNNs (NeurIPS 2025 · 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.