Research map / Architectures and efficiency / RNNs and normalization
ReX: An Efficient Approach to Reducing Memory Cost in Image Classification
Xuwei Qian, Renlong Hang, Qingshan Liu
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 early blocks with a ReX layer | ResNet34 (7 exits) · ImageNet-1k | Peak RAM: 3.1MB -> 0.4MB (↓7.75x) | Improved |
| Replace early blocks with a ReX layer | MobileNetV2 (6 exits) · ImageNet-1k | Peak RAM: 2.3MB -> 245KB (↓11.5x) | Improved |
| Replace early blocks with a ReX layer | MobileNetV2 (6 exits) · ImageNet-1k | Top-1 accuracy: 71.7 ± 0.24% -> 72.8 ± 0.11% | Improved |
| Use consistency-based early exit instead of confidence-based exit | ReX-MobileNetV2 (6 exits) · Visual Wake Words | Top-1 accuracy: 89.7% -> 91.0% | Improved |
| Use consistency-based early exit instead of confidence-based exit | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Top-1 accuracy: 71.1% -> 72.8% | Improved |
| Use a 4×4 patch size (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 72.8% -> 71.7% | Hurt |
| Use a 12×12 patch size (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 72.8% -> 72.0% | Hurt |
| Increase stride to 6 (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 72.8% -> 70.5% | Hurt |
| Reduce hidden sizes to h1=8 and h2=8 (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 72.8% -> 71.1% | Hurt |
| Increase hidden sizes to h1=16 and h2=16 (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 72.8% -> 73.1% | Improved |
| Add global vectors V G (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 69.7% -> 72.8% | Improved |
| Add row summary V R (ablation) | ReX-MobileNetV2 (6 exits) · ImageNet-1k | Accuracy: 70.9% -> 72.8% | Improved |
Closest papers by meaning
- Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification (NeurIPS 2020 · similarity 0.87)
- FASTER Recurrent Networks for Efficient Video Classification (AAAI 2020 · similarity 0.86)
- Adaptive Neural Networks for Efficient Inference (ICML 2017 · similarity 0.86)
- Pixel Recurrent Neural Networks (ICML 2016 · similarity 0.86)
- DASH: Warm-Starting Neural Network Training in Stationary Settings without Loss of Plasticity (NeurIPS 2024 · similarity 0.85)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.84)
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.
