Research map / Architectures and efficiency / CNN design
Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units
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 ReLU with CReLU in lower conv layers (same filter count) | ConvPool-CNN-C · CIFAR-10/100 | test error (Vote / Single): CIFAR-10 Vote 7.55 -> 7.09; CIFAR-100 Single 36.30 -> 31.48 | Helped |
| CReLU with halved filters vs ReLU baseline with same neuron count (ablation) | ConvPool-CNN-C · CIFAR-10 | test error (Vote): 7.55 -> 7.09 with half the parameters (0.7M vs 1.4M) | Helped |
| replace ReLU with absolute value rectification (AVR) (ablation) | ConvPool-CNN-C · CIFAR-10/100 | test error (Average): CIFAR-10 single 8.32 (better than baseline 9.17) but Average 10.26 vs 10.20; CIFAR-100 Average 37.24 vs 38.52 better | Mixed |
| double number of filters in ReLU baseline (ablation) | ConvPool-CNN-C · CIFAR-100 | test error (Average): 38.52 -> 36.73 | Helped |
| apply CReLU after conv1, conv3, conv5 of VGG while halving filters | VGG · CIFAR-10/100 | test error (Vote): CIFAR-10 5.43 -> 5.09 (conv1,3); CIFAR-100 26.85 -> 23.66 (conv1,3,5) | Helped |
| apply CReLU to conv1-4 of All-CNN-B (same filters) | All-CNN-B · ImageNet | top-1 error (center crop): 41.81 -> 39.82 | Helped |
| apply CReLU to conv1,4,7 with halved convs | All-CNN-B · ImageNet | top-1 error (10-patch): 38.03 -> 35.70 | Helped |
| apply CReLU to all conv layers of All-CNN-B with halved convs (ablation) | All-CNN-B · ImageNet | top-1 error (center crop): 41.81 -> 40.93 with 4.7M params | Helped |
| apply AVR instead of ReLU to conv1-4, conv1-7, or conv1-9 of All-CNN-B (ablation) | All-CNN-B · ImageNet | top-1 error (center crop): conv1-4 41.12 (vs 41.81) but conv1-7 42.36 and conv1-9 43.33 (worse) | Mixed |
| extend CReLU from conv1-4 to conv1-9 of All-CNN-B (same filters) (ablation) | All-CNN-B · ImageNet | top-1 error (center crop): conv1-4 39.82 -> conv1-9 40.15 (with more conv layers, performance drops) | Hurt |
Closest papers by meaning
- More Is Less: A More Complicated Network With Less Inference Complexity (CVPR 2017 · similarity 0.88)
- Refining Architectures of Deep Convolutional Neural Networks (CVPR 2016 · similarity 0.87)
- Doubly Convolutional Neural Networks (NeurIPS 2016 · similarity 0.87)
- Modulated Convolutional Networks (CVPR 2018 · similarity 0.87)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.85)
- ReLUs Are Sufficient for Learning Implicit Neural Representations (ICML 2024 · similarity 0.85)
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.