Research map / Architectures and efficiency / CNN design
Convolutional Networks with Dense Connectivity
Gao Ming Huang, Zhuang Liu, Geoff Pleiss, Laurens J. P. van der Maaten, Kilian Q. Weinberger
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| increase depth and growth rate (more parameters) | DenseNet · C10+ | test error (%): 4.51 -> 3.62 -> 3.46 | Improved |
| use DenseNet instead of FractalNet with drop-path, no augmentation | DenseNet (k=24, L=250) · C10 | test error (%): 7.33 -> 5.19 | Improved |
| use DenseNet instead of FractalNet with drop-path, no augmentation | DenseNet (k=24, L=250) · C100 | test error (%): 28.20 -> 19.64 | Improved |
| use 250-layer instead of 100-layer DenseNet | DenseNet · SVHN | test error (%): 1.76 -> 1.74 | No clear effect |
| use DenseNet (0.8M params) instead of 1001-layer pre-activation ResNet (10.2M) | DenseNet k=12 L=100 · C10+ | test error (%): 4.62 -> 4.51 | No clear effect |
| add dense concatenation instead of summation (DenseNet vs ResNet) | DenseNet · CIFAR-10+ small networks | parameter efficiency: 1/3 of the parameters | Improved |
| use smaller growth rate (24 vs 40) at 24% validation error (ablation) | DenseNet · ImageNet | flops: 0.88x10^10 -> 0.50x10^10 | Improved |
| vary compression factor theta 0.3, 0.5, 0.7 (ablation) | DenseNet · ImageNet | error vs flops | No clear effect |
| add full dense connectivity (remove 1x1 conv and compression in transitions) (ablation) | DenseNet · C10+ | parameter and computational efficiency | Hurt |
| restrict connections to most recent M layers (ablation) | DenseNet · ImageNet | validation error | Hurt |
| exponentially increase growth rate per block (ablation) | DenseNet · ImageNet | error vs flops | Mixed |
| use pre-activation instead of post-activation batch normalization (ablation) | DenseNet · ImageNet | efficiency | Improved |
Closest papers by meaning
- Densely Connected Convolutional Networks (CVPR 2017 · similarity 0.99)
- DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTs (ECCV 2024 · similarity 0.94)
- Fully Convolutional Networks for Semantic Segmentation (TPAMI 2016 · similarity 0.89)
- Multi-Scale Dense Networks for Resource Efficient Image Classification (ICLR 2018 · similarity 0.88)
- Densely Connected Multi-Dilated Convolutional Networks for Dense Prediction Tasks (CVPR 2021 · similarity 0.88)
- Feature Pyramid Networks for Object Detection (CVPR 2017 · similarity 0.88)
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.
