Research map / Architectures and efficiency / CNN design
Modelling Long Range Dependencies in $N$D: From Task-Specific to a General Purpose CNN
David M Knigge, David W. Romero, Albert Gu, Efstratios Gavves, Erik J Bekkers, Jakub Mikolaj Tomczak, Mark Hoogendoorn, Jan-jakob Sonke
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use CCNN₆,₃₈₀ instead of treating the image as a flattened sequence | CCNN on LRA Pathfinder · LRA Pathfinder (2DPathfinder vs Pathfinder) | accuracy: 91.51 -> 96.00 | Improved |
| Use CCNN₆,₃₈₀ instead of treating the image as a flattened sequence | CCNN on LRA Image · LRA Image (2DImage vs Image) | accuracy: 88.90 -> 91.12 | Improved |
| Replace the FlexCNNBlock with the proposed CCNNBlock (ablation) | CCNN₄,₁₄₀ · SpeechCommands-MFCC | accuracy: 94.38 -> 95.01 | Improved |
| Replace the FlexCNNBlock with the proposed CCNNBlock (ablation) | CCNN₄,₁₄₀ · sCIFAR10 | accuracy: 85.54 -> 90.30 | Improved |
| Replace the FlexCNNBlock with the proposed CCNNBlock (ablation) | CCNN₄,₁₄₀ · CIFAR10 | accuracy: 86.34 -> 92.78 | Improved |
| Replace the FlexCNNBlock with the proposed CCNNBlock (ablation) | CCNN₆,₃₈₀ · SpeechCommands-MFCC | accuracy: 94.92 -> 95.01 | Improved |
| Replace the FlexCNNBlock with the proposed CCNNBlock (ablation) | CCNN₆,₃₈₀ · sCIFAR10 | accuracy: 84.68 -> 93.08 | Improved |
| Replace the FlexCNNBlock with the proposed CCNNBlock (ablation) | CCNN₆,₃₈₀ · CIFAR10 | accuracy: 92.24 -> 95.20 | Improved |
| Use the S4Block instead of the proposed CCNNBlock (ablation) | CCNN₄,₁₄₀ · SpeechCommands-MFCC | accuracy: 95.01 -> 73.60 | Hurt |
| Use the S4Block instead of the proposed CCNNBlock (ablation) | CCNN₄,₁₄₀ · sCIFAR10 | accuracy: 90.30 -> 53.25 | Hurt |
| Use the S4Block instead of the proposed CCNNBlock (ablation) | CCNN₄,₁₄₀ · CIFAR10 | accuracy: 92.78 -> 60.06 | Hurt |
| Train CCNN₄,₁₄₀ on voxelized ModelNet10, then evaluate on point clouds | CCNN₄,₁₄₀ · ModelNet10 test set | accuracy: 90.64 -> 90.99 | Improved |
Closest papers by meaning
- CKConv: Continuous Kernel Convolution For Sequential Data (ICLR 2022 · similarity 0.86)
- Reparameterized Multi-Resolution Convolutions for Long Sequence Modelling (NeurIPS 2024 · similarity 0.85)
- The Unreasonable Effectiveness of Fully-Connected Layers for Low-Data Regimes (NeurIPS 2022 · similarity 0.85)
- Deep Continuous Networks (ICML 2021 · similarity 0.85)
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes (ICLR 2022 · similarity 0.85)
- Viewing Transformers Through the Lens of Long Convolutions Layers (ICML 2024 · 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.
