| Title |
Efficient Training Acceleration System for Binary Neural Networks |
| Authors |
류창호(Changho Ryu) ; 이형석(Hyeongseok Lee) ; 김태환(Tae-Hwan Kim) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.1.3 |
| Keywords |
Binary neural networks; Training; Backpropagation; Deep learning; FPGA |
| Abstract |
This paper presents an efficient implementation of a training acceleration system for binary neural networks. As the convolution operations take up dominant time in training, a dedicated hardware component to perform the convolution operations is designed and integrated to the system. An efficient training algorithm is employed, where the parameters can be updated selectively to reduce the computations without degrading the performance. In addition, a novel method to manage the memory is proposed to maximize the memory reuse by reordering the training processes. In the proposed system, the overall training time is reduced by 89.6% in comparison to that of the software implementation. |