AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints

Authors: Zhiwei Dong, Mingzhu Shen, Shihao Bai, Xiuying Wei, Jinyang Guo, Ruihao Gong, Song-Lu Chen, Xianglong Liu, Xu-Cheng Yin

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the effectiveness of the exploitation of the hardware characteristics. Remarkably, Atom Net pioneeringly achieves 3.5% accuracy enhancement and more than 15% latency reduction on 320KB MCU using readily available official inference libraries for Image Net tasks, surpassing the current state-of-the-art method. Experiments on various datasets (Image Net (Deng et al. 2009), VWW (Chowdhery et al. 2019), and Pascal VOC (Everingham et al. 2010)) and tasks (classification and object detection) prove the superiority of our Atom DB and Atom Net.
Researcher Affiliation Collaboration Zhiwei Dong1,3*, Mingzhu Shen3 , Shihao Bai3, Xiuying Wei3, Jinyang Guo2, Ruihao Gong2, 3, Song-Lu Chen1, Xianglong Liu2, Xu-Cheng Yin1 1University of Science and Technology Beijing 2Beihang University 3Sense Time Research
Pseudocode No The paper describes methodologies verbally and with diagrams (Figure 2 "The overall framework of our method"), but no structured pseudocode or algorithm blocks are present.
Open Source Code No The paper does not provide any specific links to source code repositories or explicit statements about releasing their code for the methodology described. It mentions using "readily available official inference libraries" such as "Tiny Engine (Lin et al. 2020)" and "STM32 X-CUBE-AI (STMicroelectronics 2023)" but does not offer its own implementation code.
Open Datasets Yes Experiments on various datasets (Image Net (Deng et al. 2009), VWW (Chowdhery et al. 2019), and Pascal VOC (Everingham et al. 2010)) and tasks (classification and object detection) prove the superiority of our Atom DB and Atom Net.
Dataset Splits No The paper mentions using ImageNet, VWW, and Pascal VOC datasets and describes training configurations, but it does not explicitly specify the training/test/validation splits used for these datasets.
Hardware Specification Yes Deployment. Following (Lin et al. 2020), we quantize our Atom Net to INT8 and use STM32 X-CUBE-AI inference library when deploying them on different hardware: STM32F412 (256KB SRAM/1MB Flash), STM32F746 (320KB SRAM/1MB Flash) and STM32H743 (512KB SRAM/2MB Flash).
Software Dependencies No The paper mentions "STM32 X-CUBE-AI inference library (STMicroelectronics 2023)" but does not provide a specific version number for this library. It also mentions optimizers like "SGD" and "Adam W" without specific software library versions.
Experiment Setup Yes Training configurations. For the Image Net dataset, we use the SGD optimizer for optimization. The initial learning rate, the weight decay, and the batch size are set as 0.4, 3e 4, and 1024, respectively. Cosine learning rate decay is adopted in the training process. For the VWW dataset, we finetune the network trained on the Image Net for 30 epochs as the evaluation model. For the PASCAL VOC dataset, we train the network for 300 epochs with the Adam W optimizer (Loshchilov and Hutter 2017).