Home // ICDT 2021, The Sixteenth International Conference on Digital Telecommunications // View article

An Evaluation of Neural Network Performance Using Complex-Valued Input Data

Authors:
Kushal Thapa
Stan McClellan

Keywords: complex-valued data; machine learning; neural network; real spectrogram; imaginary spectrogram

Abstract:
Complex-valued data is ubiquitous in many scientific fields. However, machine learning for complex-valued input is still in the developmental stage. Alternatively, complex data can be transformed to real data in a few different ways to fit the traditional machine learning framework. In this research, we compare the performance of two such ways - combining real and imaginary components or stacking them - on a simple neural network. To compare these two methods, we create magnitude (combined) and rectangular (stacked) spectrograms from artificial time-series data. Then, we feed the raw 1D time-series dataset, 2D magnitude spectrogram dataset, and 3D rectangular spectrogram dataset to a neural network for training and validation. As a measure of performance, we track the accuracy of each dataset model. From our experimentation, we found out that the rectangular dataset outperforms the magnitude spectrogram in most cases.

Pages: 27 to 31

Copyright: Copyright (c) IARIA, 2021

Publication date: April 18, 2021

Published in: conference

ISSN: 2308-3964

ISBN: 978-1-61208-835-8

Location: Porto, Portugal

Dates: from April 18, 2021 to April 22, 2021