Reducing Complexity of CSI Feedback based on Deep Learning for Massive MIMO using Tensor-Train Decomposition
Xiangyu Cen, Chan–Tong Lam, Yuanhui Liang, Man Xu, Benjamin K. Ng, Sio‐Kei Im · 2023
To further reduce the complexity of the channel state information (CSI) based on deep learning for massive multiple inputs and multiple outputs (MIMO) system, we propose a method called TT-CsiNet, based on the original CsiNet framework, using Tensor-Train decomposition to decompose the weights of the fully-connected layers, at the UE side. Experimental results show that the memory requirement is reduced by 99% of the original CsiNet for UEs, depending on the compression ratio. Moreover, TT-CsiNet can reduce the number of floating point operations (FLOPs) by about 17.27% to 80.44%, depending on the compression ratio. In addition, TT-CsiNet has comparable performance to the original CsiNet and other variants of the low complexity CsiNet using pruning, quantization, and weight clustering.