Learned spike encoding of the channel response for low-power environment sensing

The use of reflected radio signals to remotely sense the environment at medium-short range, based on the radar principle, has sparked great interest from the research community and the industry alike. This is referred to as Radio-Frequency (RF) sensing, and it is preferable over widely used camera systems in privacy-sensitive applications, such as in pervasive healthcare and assisted living services, vehicle and crowd monitoring, among others. Due to the complexity of the radio propagation channel, Deep Learning (DL) has recently become the tool of choice to carry out advanced RF sensing tasks such as human activity recognition and gait analysis. The current paradigm in DL is to perform the computing processes in the cloud. However, the fast growth of Internet of Things (IoT) and the proliferation of smart devices have pushed the horizon of \emph{edge computing}, which calls for processing the data right at the network edge, where data is collected. Edge computing emerges as the optimal choice for RF sensing applications, since it would avoid continuously transferring of huge amount of RF data that would quickly saturate the network capacity, unlike alternative sensing methods. Nevertheless, moving the computing processes to the network edge would result in an increased energy demand, further exacerbating the already substantial carbon emissions associated with mobile networks. In this thesis, we fill this gap by proposing a \textit{learned} encoding technique for RF signals, obtained using a Convolutional Autoencoder (CAE). Through an encoder network, the RF signal is mapped to a spike train, whose sparsity is promoted through a suitable regularization function. Then, a decoder reconstructs the original signal from the spike representation, enabling training of the CAE through backpropagation of the reconstruction error. To the best of our knowledge, this is the first time neural networks are employed to learn the spike encoding of RF signals. To assess the effectiveness of our encoding technique, we compare it to classical temporal encoding methods, which rely on a fixed threshold to encode the differences in the signal amplitude. Our results on simulated RF signal traces show that the proposed learned encoding outperforms existing methods in terms of signal reconstruction, preservation of the original spectral components, and sparsity of the spike train. The latter two aspects are especially promising for RF sensing on the edge, as they allow accurate sensing with very low power consumption.