Flynet: Max it, Excite it, Quantize it
Luis Guerra (Monash University),* Tom Drummond (University of Melbourne)
The 33rd British Machine Vision Conference
Abstract
We present a new efficient Convolutional Neural Network (CNN) architecture targeted for tiny machine learning (TinyML) on vision tasks. Through a series of architectural improvements to state-of-the-art mobile networks, we are able to significantly reduce the number of parameters while maintaining a reasonable number of multiply-accumulate operations. We switch the load from expensive pointwise convolutions into lighter multihead-depthwise convolutions with non-linear Max-out aggregation. By incorporating our contributions to a MobileNetV3 backbone, we achieved comparable accuracy with up to 0.5x reduction in parameters in the ImageNet dataset, and achieved 61\% top-1 accuracy, matching MicroNet-M2, ShufflenetV2-0.5x and EfficientNet-B, with 1MB. We additionally reported results in the tasks of object detection in the COCO dataset. Finally we performed ablation studies to demonstrate the effectiveness of our improvements. Code will be made available online.Video
Citation
@inproceedings{Guerra_2022_BMVC,
author = {Luis Guerra and Tom Drummond},
title = {Flynet: Max it, Excite it, Quantize it},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0407.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by