LSTM and ResNets Deep Learning Aided End-to-End Intelligent Communication Systems

Abstract

Deep learning (DL) technology enables communication systems to provide intelligent transmissions to adapt to time changing wireless channel conditions. In this paper, with considerations that information bits are usually grouped into blocks or sequences for transmissions, we propose a DL end-to-end intelligent communication system. In our design, we construct a neural network (NN) constituted by long short-term memory (LSTM) units and residual networks (ResNets) architecture, to process the information-bearing sequences. More explicitly, we propose to add a forward feed path to compose the ResNets, thus the residual learning can be implemented to accelerate the convergence. Moreover, with the LSTM units, information-bearing sequences can be processed sequentially by the NN. Thus for larger number of sequences, better symbol error rate (SER) performances can be achieved since the features of the correlation between messages delivered at different time slots will be extracted to improve the detection performances. Subsequently, simulation results are provided to demonstrate that the proposed LSTM and ResNets based intelligent transmission systems can achieve better performances than benchmark systems over wireless channels undergoing the multiplicative fading and additive noises, while providing satisfactory robustness and convergence performances.