Domain Generalization Capability Enhancement for Binary Neural Networks
Jianming Ye (Peking University), Shunan Mao (Peking University), Shiliang Zhang (Peking University)*
The 33rd British Machine Vision Conference
Abstract
The in-domain performance of Binary Neural Networks (BNNs) has been significantly boosted by recent research efforts. The effect of heavy compression to the generalization capability of BNNs, however, has not been explored. This paper shows that binary compression degrades the generalization capability of BNNs, and addresses this issue by optimizing the distribution of BNN parameters and activations. A novel BNN training scheme is proposed to pursue a flat minimum for binary parameters through optimizing latent real-valued weights. Our method also optimizes the distributions of BNN activations to decrease the quantization errors caused by binarization. Extensive experiments on three domain generalization datasets reveal that, jointly optimizing BNN weights and activations substantially enhances the generalization capability, making our BNN achieve the best performance among its competitors. Our method also exhibits good compatibility to different network architectures, and performs well on general image classification datasets like CIFAR-100 and ImageNet.Video
Citation
@inproceedings{Ye_2022_BMVC,
author = {Jianming Ye and Shunan Mao and Shiliang Zhang},
title = {Domain Generalization Capability Enhancement for Binary Neural Networks},
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/0013.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by