In this letter, a piecewise linear (PWL) sigmoid function approximation based on the statistical distribution probability of the neurons' values in each layer is proposed to improve the network recognition accuracy with only addition circuit. The sigmoid function is first divided into three fixed regions, and then according to the neurons' values distribution probability, the curve in each region is segmented into sub-regions to reduce the approximation error and improve the recognition accuracy. Experiments performed on Xilinx's FPGA-XC7A200T for MNIST and CIFAR-10 datasets show that the proposed method achieves 97.45% recognition accuracy in DNN, 98.42% in CNN on MNIST and 72.22% on CIFAR-10, up to 0.84%, 0.57% and 2.01% higher than other approximation methods with only addition circuit.
Vantruong NGUYEN
Xidian University
Jueping CAI
Xidian University
Linyu WEI
Xidian University
Jie CHU
Xidian University
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Vantruong NGUYEN, Jueping CAI, Linyu WEI, Jie CHU, "Neural Networks Probability-Based PWL Sigmoid Function Approximation" in IEICE TRANSACTIONS on Information,
vol. E103-D, no. 9, pp. 2023-2026, September 2020, doi: 10.1587/transinf.2020EDL8007.
Abstract: In this letter, a piecewise linear (PWL) sigmoid function approximation based on the statistical distribution probability of the neurons' values in each layer is proposed to improve the network recognition accuracy with only addition circuit. The sigmoid function is first divided into three fixed regions, and then according to the neurons' values distribution probability, the curve in each region is segmented into sub-regions to reduce the approximation error and improve the recognition accuracy. Experiments performed on Xilinx's FPGA-XC7A200T for MNIST and CIFAR-10 datasets show that the proposed method achieves 97.45% recognition accuracy in DNN, 98.42% in CNN on MNIST and 72.22% on CIFAR-10, up to 0.84%, 0.57% and 2.01% higher than other approximation methods with only addition circuit.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.2020EDL8007/_p
Copy
@ARTICLE{e103-d_9_2023,
author={Vantruong NGUYEN, Jueping CAI, Linyu WEI, Jie CHU, },
journal={IEICE TRANSACTIONS on Information},
title={Neural Networks Probability-Based PWL Sigmoid Function Approximation},
year={2020},
volume={E103-D},
number={9},
pages={2023-2026},
abstract={In this letter, a piecewise linear (PWL) sigmoid function approximation based on the statistical distribution probability of the neurons' values in each layer is proposed to improve the network recognition accuracy with only addition circuit. The sigmoid function is first divided into three fixed regions, and then according to the neurons' values distribution probability, the curve in each region is segmented into sub-regions to reduce the approximation error and improve the recognition accuracy. Experiments performed on Xilinx's FPGA-XC7A200T for MNIST and CIFAR-10 datasets show that the proposed method achieves 97.45% recognition accuracy in DNN, 98.42% in CNN on MNIST and 72.22% on CIFAR-10, up to 0.84%, 0.57% and 2.01% higher than other approximation methods with only addition circuit.},
keywords={},
doi={10.1587/transinf.2020EDL8007},
ISSN={1745-1361},
month={September},}
Copy
TY - JOUR
TI - Neural Networks Probability-Based PWL Sigmoid Function Approximation
T2 - IEICE TRANSACTIONS on Information
SP - 2023
EP - 2026
AU - Vantruong NGUYEN
AU - Jueping CAI
AU - Linyu WEI
AU - Jie CHU
PY - 2020
DO - 10.1587/transinf.2020EDL8007
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E103-D
IS - 9
JA - IEICE TRANSACTIONS on Information
Y1 - September 2020
AB - In this letter, a piecewise linear (PWL) sigmoid function approximation based on the statistical distribution probability of the neurons' values in each layer is proposed to improve the network recognition accuracy with only addition circuit. The sigmoid function is first divided into three fixed regions, and then according to the neurons' values distribution probability, the curve in each region is segmented into sub-regions to reduce the approximation error and improve the recognition accuracy. Experiments performed on Xilinx's FPGA-XC7A200T for MNIST and CIFAR-10 datasets show that the proposed method achieves 97.45% recognition accuracy in DNN, 98.42% in CNN on MNIST and 72.22% on CIFAR-10, up to 0.84%, 0.57% and 2.01% higher than other approximation methods with only addition circuit.
ER -