Research map / Architectures and efficiency / CNN design
Generalizing Pooling Functions in CNNs: Mixed, Gated, and Tree
Chen‐Yu Lee, Patrick Gallagher, Zhuowen Tu
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 |
|---|---|---|---|
| replace max pooling with gated max-avg pooling (1 per layer) | DSN-style CNN · CIFAR10 | test error (%): 9.10 -> 7.90 | Improved |
| make the max-avg mix responsive via a gate rather than a fixed learned mix | mixed max-avg pooling · CIFAR10 | test error (%): 8.09 -> 7.90 | Improved |
| increase mixed parameters from 1 per layer to 1 per layer/channel/region (ablation) | mixed max-avg pooling · CIFAR10 | test error (%): 8.09 -> 8.05 | No clear effect |
| use per layer/channel/region mixing on large data (ablation) | mixed max-avg pooling · SVHN | test error (%): 1.76 -> 1.64 | Improved |
| add learning of the mix proportion versus fixed 50/50 (ablation) | max-avg pooling · CIFAR10 | test error (%): 8.11 -> 7.90 | Improved |
| add a second tree level (ablation) | tree pooling · CIFAR10 | test error (%): 8.25 -> 8.22 | No clear effect |
| add a third tree level (ablation) | tree pooling · SVHN | test error (%): 1.80 -> 1.70 | Improved |
| add a third tree level (ablation) | tree pooling · MNIST | test error (%): 0.35 -> 0.37 | Hurt |
| follow tree pooling with gated max-avg in second pooling layer (ablation) | tree pooling · CIFAR10 | test error (%): 8.22 -> 7.62 | Improved |
| add two extra conv layers instead of learned pooling (ablation) | baseline CNN · CIFAR10 | test error (%): 9.10 -> 9.17 | Hurt |
| add one extra conv layer instead of learned pooling (ablation) | baseline CNN · CIFAR10 | test error (%): 9.10 -> 9.08 | No clear effect |
| replace max pooling with stochastic pooling (ablation) | baseline CNN · CIFAR10 | test error (%): 9.10 -> 8.50 | Improved |
Closest papers by meaning
- Rethinking pooling in graph neural networks (NeurIPS 2020 · similarity 0.90)
- Global Feature Guided Local Pooling (ICCV 2019 · similarity 0.90)
- Learning with convolution and pooling operations in kernel methods (NeurIPS 2022 · similarity 0.89)
- Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities (AAAI 2025 · similarity 0.87)
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.87)
- SkipPool: Improved Sparse Hierarchical Graph Pooling with Differentiable Exploration (AAAI 2025 · similarity 0.87)
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.
