Invited paper for WSA 2021, called 'Low Complexity Channel estimation with Neural Network Solutions'.
%% Cite as:
@INPROCEEDINGS{9739166, author={Luan, Dianxin and Thompson, John}, booktitle={WSA 2021; 25th International ITG Workshop on Smart Antennas}, title={Low Complexity Channel estimation with Neural Network Solutions}, year={2021}, volume={}, number={}, pages={1-6}, doi={}} %%
Low complexity residual convolutional neural network for channel estimation
Conpared with the ReEsNet from the repo Residual_CNN, it has slightly improved performance and the number of parameters is reduced by 82% (when pruning is not applied). I planed to release the code when I sorted out the files.
Run Demonstration_of_H_regression_48_CommuRayleigh.m to test.
Run ResNN_pilot_regression.m for training the neural network.
One thing you need to keep in mind is that, the pruning is applied without retraining because it aims to minimize the computations/latency for low-complexity solutions. Typically, after pruning there should be complex procedures (Learning both weights and connections for efficient neural networks, arXivpreprint) to compensate, but that is not realistic for low-latency and low-complexity online implementation. For sure, that pruning degrades the performance significantly. I just remove the retraining and feedback loop.