Research map / Architectures and efficiency / CNN design
Efficient Learning of CNNs using Patch Based Features
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 |
|---|---|---|---|
| use full-information patch embedding (Phi_full) instead of hard assignment (Phi_hard) | patch-based linear classifier with data patches · CIFAR-10 | test accuracy: 81.23% vs 78.80% | Helped |
| use clustered data patches as dictionary instead of Gaussian random patches (ablation) | Phi_full embedding · CIFAR-10 | test accuracy: 81.23% vs 76.04% | Helped |
| use clustered data patches as dictionary instead of Gaussian random patches (ablation) | Phi_hard embedding · CIFAR-10 | test accuracy: 78.80% vs 71.36% | Helped |
| Phi_full patch-based linear model (with pooling and bottleneck) vs vanilla one-hidden-layer CNN | one-hidden-layer model · CIFAR-10 | test accuracy: 81.23% vs 80.08% | Helped |
| constrain linear classifier to shared per-patch linear function (ablation) | Phi_full linear classifier · CIFAR-10 | test accuracy: 76.61% vs 76.38% with 75x fewer parameters | No clear effect |
| add average pooling after the per-patch outputs (ablation) | Phi_full constrained linear model · CIFAR-10 | test accuracy: 80.42% vs 78.98% | Helped |
| add 1x1 conv bottleneck (ablation) | Phi_full constrained linear model · CIFAR-10 | test accuracy: 80.38% vs 78.98% | Helped |
| add both average pooling and 1x1 bottleneck (ablation) | Phi_full constrained linear model · CIFAR-10 | test accuracy: 81.23% vs 78.98% with 39x fewer parameters | Helped |
| layer-wise stacking of Phi_full to depth 4 (ablation) | Phi_full · CIFAR-10 | test accuracy: 82.88% at depth 4 vs 81.33% at depth 1 | Helped |
| layer-wise stacking of Phi_hard to depth 4 (ablation) | Phi_hard · CIFAR-10 | test accuracy: 50.69% at depth 4 vs 78.30% at depth 1 | Hurt |
| sample-based patch dictionary via k-means with k=0.25N neighbors, layer-wise deep model vs end-to-end CNN | deep patch model · CIFAR-10 | test accuracy: 82.88% vs 89.62% for depth-4 CNN | Hurt |
Closest papers by meaning
- Learning One Convolutional Layer with Overlapping Patches (ICML 2018 · similarity 0.90)
- Greedy Layerwise Learning Can Scale To ImageNet (ICML 2019 · similarity 0.88)
- PatchUp: A Feature-Space Block-Level Regularization Technique for Convolutional Neural Networks (AAAI 2022 · similarity 0.88)
- The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods (ICLR 2021 · similarity 0.87)
- Patch-Level Representation Learning for Self-Supervised Vision Transformers (CVPR 2022 · similarity 0.87)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 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.