Deterministic Binary Filters for Convolutional Neural Networks
Deterministic Binary Filters for Convolutional Neural Networks
复制标题
DOI：
10.24963/ijcai.2018/380
复制
发表时间：
2018-07
期刊：
影响因子：
--
通讯作者：
V. W. Tseng;S. Bhattacharya;J. Fernández-Marqués;Milad Alizadeh;C. Tong;N. Lane
中科院分区：
文献类型：
--
作者：
V. W. Tseng;S. Bhattacharya;J. Fernández-Marqués;Milad Alizadeh;C. Tong;N. Lane
我们提出了确定性二进制滤波器，一种卷积神经网络的方法，它学习预定义的正交二进制基的加权系数，而不是直接学习卷积滤波器的传统方法。这种方法导致模型架构具有显著更少的参数（4倍至16倍）和更小的模型大小（由于使用二进制而不是浮点精度，因此为32倍）。我们证明了我们的确定性滤波器设计可以集成到众所周知的网络架构（如ResNet和SqueezeNet）中，精度损失仅为2%（在CIFAR-10等数据集下）。在ImageNet下，与sub-megabyte二进制网络相比，它们的模型大小减少了3倍，同时达到了相当的精度水平。
We propose Deterministic Binary Filters, an approach to Convolutional Neural Networks that learns weighting coefficients of predefined orthogonal binary basis instead of the conventional approach of learning directly the convolutional filters. This approach results in model architectures with significantly fewer parameters (4x to 16x) and smaller model sizes (32x due to the use of binary rather than floating point precision). We show our deterministic filter design can be integrated into well-known network architectures (such as ResNet and SqueezeNet) with as little as 2% loss of accuracy (under datasets like CIFAR-10). Under ImageNet, they result in 3x model size reduction compared to sub-megabyte binary networks while reaching comparable accuracy levels.