Research map / Architectures and efficiency / CNN design
Memory-efficient Patch-based Inference for Tiny Deep Learning
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 |
|---|---|---|---|
| run initial stage patch-by-patch (4x4) instead of per-layer | MobileNetV2 · peak SRAM | kB: 1372kB -> 172kB | Helped |
| run initial stage patch-by-patch (4x4) | MobileNetV2 · overall computation | overhead: +10% | Hurt |
| redistribute receptive field (MbV2-RD) (ablation) | MobileNetV2 per-patch · overall computation | overhead: +10% -> +3% | Helped |
| redistribute receptive field (MbV2-RD) (ablation) | MobileNetV2 per-patch · ImageNet | Top-1 accuracy: 72.2% -> 72.1% | No clear effect |
| redistribute receptive field (MbV2-RD) (ablation) | MobileNetV2 per-patch · Pascal VOC | mAP: 75.4% -> 75.7% | No clear effect |
| use overlapping patches with RF redistribution instead of non-overlapping patches (ablation) | MobileNetV2 per-patch · Pascal VOC | mAP: 73.9% -> 75.7% | Helped |
| use overlapping patches with RF redistribution instead of non-overlapping patches (ablation) | MobileNetV2 per-patch · ImageNet | Top-1 accuracy: 71.8% -> 72.1% | Helped |
| use MCUNetV2 patch inference instead of RNNPool (ablation) | MobileNetV2 · ImageNet | Top-1 accuracy: 70.1% -> 72.1% | Helped |
| add patch-based inference to joint NAS (MCUNetV2) vs MCUNet | TinyNAS · ImageNet 256kB SRAM | Top-1 accuracy: 60.3% -> 64.9% | Helped |
| add patch-based inference to joint NAS (MCUNetV2) vs MCUNet | TinyNAS · ImageNet 512kB SRAM | Top-1 accuracy: 68.5% -> 71.8% | Helped |
| joint NAS with patch-based inference enabling higher resolution | MCUNet · Pascal VOC H743 | mAP: 51.4% -> 68.3% | Helped |
| use patch-based inference in search space vs per-layer only | MCUNetV2 NAS · Visual Wake Words | peak SRAM: 4.0x smaller | Helped |
Closest papers by meaning
- msf-CNN: Patch-based Multi-Stage Fusion with Convolutional Neural Networks for TinyML (NeurIPS 2025 · similarity 0.89)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · similarity 0.86)
- DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators (NeurIPS 2024 · similarity 0.86)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.85)
- Auto-reconfiguration for Latency Minimization in CPU-based DNN Serving (ICML 2025 · similarity 0.85)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · 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.