Ahstract-A new loss function is proposed for training complex-valued neural networks that reconstruct radio signals. Given a complex time series, this method incorporates loss from spectrograms with multiple aspect ratios, cross-correlation loss, and loss from amplitude envelopes in the time & frequency domains. When training a neural network, an optimizer will observe batch loss and backpropagate this value through the network to determine how to update the model parameters. The proposed loss function is robust to typical radio impairments and co-channel interference that would explode a naive mean-square error approach. This robust loss enables higher quality steps along the loss surface which in turn enables training of models specifically designed for impaired radio input. Loss versus channel impairment is shown in comparison to mean-square error for an ensemble of common channel effects. A PyTorch implementation is available at https://github.com/the-aerospace-corporation/glaucus.
Complex-Valued Radio Signal Loss for Neural Networks
2023-03-04
604061 byte
Conference paper
Electronic Resource
English
Glaucus: A Complex-Valued Radio Signal Autoencoder
IEEE | 2023
|Improvements of the traffic signal control by complex-valued Hopfield networks
Tema Archive | 2007
|Complex-Valued Signal Capsule Network for Automatic Modulation Classification
Springer Verlag | 2022
|