Sub-Terahertz Channel Gain Prediction for Scheduling of Over-The-Air Deep Learning

Enabling artificial intelligence native end-to-end systems in ultra-wideband sub-terahertz spectrum faces several challenges. The particularly complex channel variations and nonlinear behavior of analog components of the transceivers are major obstacles to the over-the-air adaptation of these systems. In this paper, we investigate an edge-based bidirectional long-short-term memory neural network capable of predicting the channel gain variations in Non-Line-of-Sight conditions. We aim to enable end-to-end autoencoders with a predictive model for scheduling the training phase when the power is above the receiver sensitivity, and there are no large fading variations. Otherwise, the training of the end-to-end system will likely fail. With only 16 BiLSTM cells, our model is capable of inferring the channel gain variations with a worst-case root mean squared error lower than 0.0547 (i.e., 1.1% compared to the normalized channel gain range). Also, with lower computational complexity, our model decreased the propagation of the error compared to traditional recurrent neural networks and deep-learning-based forecasting models.