Research map / Architectures and efficiency / CNN design
RNNPool: Efficient Non-linear Pooling for RAM Constrained Inference
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 early CNN blocks with RNNPool | MobileNetV2 · ImageNet-10 | accuracy: 94.20 -> 94.40 | Helped |
| Replace early CNN blocks with RNNPool | EfficientNet-B0 · ImageNet-10 | accuracy: 96.00 -> 96.40 | Helped |
| Replace early CNN blocks with RNNPool | DenseNet121 · ImageNet-10 | Peak RAM: 1.53 MB -> 0.77 MB | Helped |
| Replace early CNN blocks with RNNPool | MobileNetV2 · ImageNet-1K | accuracy: 71.81 -> 70.14 | Hurt |
| Replace early blocks with RNNPool | MobileNetV2 · ImageNet-1K | Peak RAM: 2.29MB -> 0.24MB | Helped |
| Replace the final average pooling layer with RNNPool (ablation) | MobileNetV2 · ImageNet-10 | accuracy: 94.20 -> 95.00 | Helped |
| Replace standard 4×4 downsampling with RNNPool | MobileNetV2 · ImageNet-10 | accuracy: 90.80 -> 94.40 | Helped |
| Replace max pooling downsampling with RNNPool | MobileNetV2 · ImageNet-10 | accuracy: 92.80 -> 94.40 | Helped |
| Replace strided convolution downsampling with RNNPool | MobileNetV2 · ImageNet-10 | accuracy: 93.00 -> 94.40 | Helped |
| Replace ReNet downsampling with RNNPool | MobileNetV2 · ImageNet-10 | accuracy: 92.20 -> 94.40 | Helped |
| Add RNNPool at the final pooling layer (ablation) | MobileNetV2-RNNPool · ImageNet-10 | accuracy: 94.40 -> 95.60 | Helped |
| Increase RNNPool face-model hidden dimension from 4 to 16 (ablation) | RNNPool-Face · WIDER FACE validation | MAP, Medium: 0.75 -> 0.89 | Helped |
Closest papers by meaning
- S3Pool: Pooling With Stochastic Spatial Sampling (CVPR 2017 · similarity 0.87)
- Recombinator Networks: Learning Coarse-To-Fine Feature Aggregation (CVPR 2016 · similarity 0.87)
- Pooling Revisited: Your Receptive Field Is Suboptimal (CVPR 2022 · similarity 0.86)
- R-FCN++: Towards Accurate Region-Based Fully Convolutional Networks for Object Detection (AAAI 2018 · similarity 0.86)
- TI-Pooling: Transformation-Invariant Pooling for Feature Learning in Convolutional Neural Networks (CVPR 2016 · similarity 0.85)
- Memory-efficient Patch-based Inference for Tiny Deep Learning (NeurIPS 2021 · 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.