低比特宽度CNN优化
本文探讨了使用低比特宽度卷积神经网络(CNN)在ImageNet数据集上的表现，包括AlexNet、ResNet系列等模型的二值化及多值量化技术。重点介绍了Bi-RealNet、Binary Ensemble等网络结构在保持精度的同时减少计算资源消耗的方法，以及Semi-Binary Decomposition(SBD)训练策略。此外，还讨论了Group-Net如何通过结构化的二值化实现更准确的图像分类和语义分割。
ImageNet using AlexNet
Towards Effective Low-bitwidth Convolutional Neural Networks
SBD: Training Binary Weight Networks via Semi-Binary Decomposition
ImageNet using ResNet-18 (W-weights, A-activation)
Bi-Real net 56.4%
TBN 55.6
Binary Ensemble 61.0%
![Full-Precision DNN [27, 43] 32 32 69.3%](https://i-blog.csdnimg.cn/blog_migrate/c3108b75925cf9fd7018825966c0f9fe.png)
ImageNet with Res-18 and Res-34
Binary Ensemble Neural Network: More Bits per Network or More Networks per Bit
ImageNet with Res-50
Towards Effective Low-bitwidth Convolutional Neural Networks
SBD: Training Binary Weight Networks via Semi-Binary Decomposition
Group-Net: Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation
Group-Net: Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation
Binary Ensemble Neural Network: More Bits per Network or More Networks per Bit?
CVPR2019
PQ+TS+Guided:
Towards Effective Low-bitwidth Convolutional Neural Networks CVPR2018
SBD: Training Binary Weight Networks via Semi-Binary Decomposition
ECCV2018
Bi-Real Net: Enhancing the Performance of 1-bit CNNs With Improved Representational Capability and Advanced Training Algorithm
ECCV2018
XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
ECCV2016
Binarized Neural Networks
NIPS2016
11
扫一扫
730
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
