TY - GEN
T1 - Swish-Driven GoogleNet for Intelligent Analog Beam Selection in Terahertz Beamspace MIMO
AU - Zarini, Hosein
AU - Mili, Mohammad Robat
AU - Rastiy, Mehdi
AU - Andreev, Sergey
AU - Nardelli, Pedro H.J.
N1 - Funding Information:
ACKNOWLEDGEMENT This work is supported by the Academy of Finland: (a) ee-IoT n.319009, (b) EnergyNet n.321265/n.328869, and (c) FIREMAN n.326270/CHISTERA-17-BDSI-003; and by JAES Foundation via STREAM project.
Publisher Copyright:
© 2022 IEEE.
jufoid=57589
PY - 2022
Y1 - 2022
N2 - In this paper, we propose an intelligent analog beam selection strategy in a terahertz (THz) band beamspace multiple-input multiple-output (MIMO) system. First inspired by transfer learning, we fine-tune the pre-trained off-the-shelf GoogleNet classifier to learn analog beam selection as a multi-class mapping problem. Simulation results show 83% accuracy for the analog beam selection, which subsequently results in 12% spectral efficiency (SE) gain over the existing counterparts. For a more accurate classifier, we replace the conventional rectified linear unit (ReLU) activation function of the GoogleNet with the recently proposed Swish and retrain the fine-tuned GoogleNet to learn analog beam selection. It is numerically indicated that the fine-tuned Swish-driven GoogleNet achieves 86% accuracy, as well as 18% improvement in achievable SE, over the similar schemes. Eventually, a strong ensembled classifier is developed to learn analog beam selection by sequentially training multiple fine-tuned Swish-driven GoogleNet classifiers. According to the simulations, the strong ensembled model is 90% accurate and yields 27% gain in achievable SE in comparison with prior methods.
AB - In this paper, we propose an intelligent analog beam selection strategy in a terahertz (THz) band beamspace multiple-input multiple-output (MIMO) system. First inspired by transfer learning, we fine-tune the pre-trained off-the-shelf GoogleNet classifier to learn analog beam selection as a multi-class mapping problem. Simulation results show 83% accuracy for the analog beam selection, which subsequently results in 12% spectral efficiency (SE) gain over the existing counterparts. For a more accurate classifier, we replace the conventional rectified linear unit (ReLU) activation function of the GoogleNet with the recently proposed Swish and retrain the fine-tuned GoogleNet to learn analog beam selection. It is numerically indicated that the fine-tuned Swish-driven GoogleNet achieves 86% accuracy, as well as 18% improvement in achievable SE, over the similar schemes. Eventually, a strong ensembled classifier is developed to learn analog beam selection by sequentially training multiple fine-tuned Swish-driven GoogleNet classifiers. According to the simulations, the strong ensembled model is 90% accurate and yields 27% gain in achievable SE in comparison with prior methods.
KW - analog beam selection
KW - beamspace
KW - ensembled classifier
KW - GoogleNet
KW - multiple-input multiple-output
KW - Swish
KW - Terahertz (THz) band
U2 - 10.1109/VTC2022-Spring54318.2022.9860549
DO - 10.1109/VTC2022-Spring54318.2022.9860549
M3 - Conference contribution
AN - SCOPUS:85137805389
T3 - IEEE Vehicular Technology Conference
BT - 2022 IEEE 95th Vehicular Technology Conference - Spring, VTC 2022-Spring - Proceedings
PB - IEEE
T2 - IEEE Vehicular Technology Conference
Y2 - 19 June 2022 through 22 June 2022
ER -