Research map / Architectures and efficiency / Neural architecture search
MCUNet: Tiny Deep Learning on IoT Devices
Ji Lin, Wei-Ming Chen, Yujun Lin, john cohn, Chuang Gan, Song Han
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 |
|---|---|---|---|
| jointly optimize TinyNAS architecture search and TinyEngine inference | MCUNet · ImageNet, STM32H743 | top-1 accuracy: 70.7% vs 69.8% | Improved |
| use TinyEngine rather than CMSIS-NN with scaled MobileNetV2 (ablation) | MobileNetV2 · ImageNet, 320kB SRAM and 1MB Flash | top-1 accuracy: 35.2% -> 47.4% | Improved |
| use TinyEngine rather than CMSIS-NN with scaled ProxylessNAS (ablation) | ProxylessNAS · ImageNet, 320kB SRAM and 1MB Flash | top-1 accuracy: 49.5% -> 56.4% | Improved |
| add TinyNAS to TinyEngine-based model deployment (ablation) | MCUNet · ImageNet, 320kB SRAM and 1MB Flash | top-1 accuracy: 47.4% -> 61.8%; 56.4% -> 61.8% | Improved |
| optimize the search space using satisfying-network FLOPs (ablation) | TinyNAS · ImageNet-100 | top-1 accuracy: 2.3% better accuracy | Improved |
| use a huge variable-resolution and variable-width search space instead of an optimized space (ablation) | TinyNAS · ImageNet-100 | top-1 accuracy | Hurt |
| use evolutionary search rather than random search (ablation) | TinyNAS with TinyEngine · search accuracy | best accuracy: 1% | Improved |
| replace layer-wise scheduling with model-adaptive im2col scheduling (ablation) | TinyEngine · inference | inference efficiency: 13% | Improved |
| specialize computation kernels with tiling, unrolling, and operation fusion (ablation) | TinyEngine · inference | inference efficiency: 22% | Improved |
| use in-place depth-wise convolution (ablation) | TinyEngine · inference | memory usage: 1.6× | Improved |
| replace runtime interpretation with code generation (ablation) | TinyEngine · inference versus CMSIS-NN | memory usage: 2.1× | Improved |
| replace runtime interpretation with code generation (ablation) | TinyEngine · inference versus CMSIS-NN | inference efficiency: 22% | Improved |
Later papers that compared with this method
7 later accepted papers report MCUNetV2 as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.
- msf-CNN: Patch-based Multi-Stage Fusion with Convolutional Neural Networks for TinyML (NeurIPS 2025 · better on 3 of 3 shared results)
- StreamNet: Memory-Efficient Streaming Tiny Deep Learning Inference on the Microcontroller (NeurIPS 2023 · better on 10 of 10 shared results)
- MCUFormer: Deploying Vision Tranformers on Microcontrollers with Limited Memory (NeurIPS 2023 · better on 6 of 6 shared results)
- Entropy-Driven Mixed-Precision Quantization for Deep Network Design (NeurIPS 2022 · better on 3 of 5 shared results)
- EtinyNet: Extremely Tiny Network for TinyML (AAAI 2022 · better on 3 of 3 shared results)
- UDC: Unified DNAS for Compressible TinyML Models for Neural Processing Units (NeurIPS 2022 · better on 2 of 2 shared results)
- Memory-efficient Patch-based Inference for Tiny Deep Learning (NeurIPS 2021 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Microcontrollers (NeurIPS 2019 · similarity 0.90)
- Constrained deep neural network architecture search for IoT devices accounting for hardware calibration (NeurIPS 2019 · similarity 0.89)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.88)
- TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices (NeurIPS 2024 · similarity 0.87)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · similarity 0.87)
- DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators (NeurIPS 2024 · 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.
