| Title |
Lightweight Binary Neural Networks with Reduced Parameters |
| Authors |
김능윤(Neungyun Kim) ; 오선희(Seonhee Oh) ; 김태환(Tae-Hwan Kim) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.12.65 |
| Keywords |
Binary neural networks; Inference; Deep learning; Lightweight model; Embedded systems |
| Abstract |
This paper presents binary neural networks with reduced number of parameters. The proposed approach modifies the conventional techniques, the depthwise-separable and grouped convolution, global average pooling, which have been widely used to reduce the number of the parameters in full-precision neural networks, to accord with the computational structure of the basic blocks composing the binary neural networks. The computational structure of the basic blocks is maintained so that has every of the input and output elements is represented in a single bit. Designed for SVHN and CIFAR-10 classification tasks, the proposed models involve 82.1% and 82.2% fewer parameters with the degradation of the inference accuracy within 1.7% and 1.8%, respectively. |