A Syndrome Based Decoding of Linear Codes Using Deep Learning Method

Using short length codes has hight degree of importance because of I.O.T ubiquitous applications. Of course Deep Learning Methods has acquired state of the art results in object detection, speech processing and some other fields. Meanwhile, Convolutional Neural Network (CNN) models play a fundamental role in Deep Learning methods success. For boosting the accuracy of syndrome based decoding of Low Density Parity Check (LDPC) codes, convolutional neural network was used and for determining the solution of syndrome equations, error pattern recognition method was used. For this purpose a three layered CNN which any layer contains pooling and convolution sub-layer was used. Then output of CNN was applied to Gated Recurrent Unit(GRU). GRU recurrent network with rectified linear unit (ReLU) activation function and node count as three times as codeword length was used. For hyper parameter default value the Tensorflow 2.0 library’s default value was used except some hyper parameters which was altered manually for accuracy improvement reason. comparing between mixed model and bare recurrent model shows that for LDPC code with length of 64, bit error rate drops when mixed model used. In different signal to noise condition, bit error rate drops 0.5 to 0.8 Db toward Maximum Likelihood decoder. Also it is shown that CNNs have potential to improve recurrent networks accuracy when being used together.