Research map / Architectures and efficiency / CNN design
Learning to Learn Parameterized Classification Networks for Scalable Input Images
Duo Li, Anbang Yao, Qifeng Chen
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 |
|---|---|---|---|
| add scale distillation from large to small resolutions (ablation) | SAN (MobileNetV2) · ImageNet val, train=test res 96 | top-1 accuracy: 61.86 -> 65.07 | Improved |
| use privatized BN layers per resolution instead of shared BN (ablation) | SAN w/o SD (MobileNetV2) · ImageNet val, train=test res 96 | top-1 accuracy: 1.38 -> 61.86 | Improved |
| generate convolution weights with a meta learner instead of privatized convolution per resolution (ablation) | SAN (ResNet-18) · ImageNet val, train=test res 96 | top-1 accuracy: 64.53 -> 65.30 | Improved |
| use a two-layer non-linear meta learner (8 hidden units) instead of linear (ablation) | SAN (ResNet-18) · ImageNet val, train 224 test 64 | top-1 accuracy: 1.50 | Hurt |
| use a two-layer non-linear meta learner (8 hidden units) instead of linear (ablation) | SAN (ResNet-18) · ImageNet val, medium resolutions | top-1 accuracy: 0.01 | No clear effect |
| double hidden units of non-linear meta learner to 16 (ablation) | SAN (ResNet-18) · ImageNet val, train 192 test 64 | top-1 accuracy: 2.74 | Hurt |
| apply BN calibration for ideal inference at interpolated resolutions (ablation) | SAN (ResNet-18) · ImageNet val, interpolated resolutions | top-1 accuracy | Improved |
| use proxy inference instead of ideal inference with calibration (ablation) | SAN (ResNet-18) · ImageNet val, test res 208/176/144/112 | top-1 accuracy | No clear effect |
| replace independently trained model with SAN | ResNet-50 · ImageNet val, train 224 test 96 | top-1 accuracy: 70.61 -> 72.83 | Improved |
| replace independently trained model with SAN | ResNet-18 · ImageNet val, train 224 test 64 | top-1 accuracy: 20.40 -> 34.98 | Improved |
| replace FixRes with SAN at large test resolution | ResNet-50 (train 224) · ImageNet val, test res 384 | top-1 accuracy: 79.0 -> 77.0 | Hurt |
| replace FixRes with SAN at small test resolution | ResNet-50 (train 224) · ImageNet val, test res 128 | top-1 accuracy: 67.7 -> 72.5 | Improved |
Closest papers by meaning
- Supplementary Meta-Learning: Towards a Dynamic Model for Deep Neural Networks (ICCV 2017 · similarity 0.90)
- Deep Neural Network Parameter Selection via Dataset Similarity Under Meta-Learning Framework (TPAMI 2025 · similarity 0.88)
- Empowering Convolutional Neural Nets with MetaSin Activation (NeurIPS 2023 · similarity 0.88)
- Multi-Scale Dense Networks for Resource Efficient Image Classification (ICLR 2018 · similarity 0.88)
- Meta Convolutional Neural Networks for Single Domain Generalization (CVPR 2022 · similarity 0.87)
- Dynamic Kernel Selection for Improved Generalization and Memory Efficiency in Meta-Learning (CVPR 2022 · 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.
