Sensitivity of a Deep JSCC to Changes in the Number of Layers of a CNN

Implementing convolutional neural networks in a deep joint source and channel coding (JSCC) scheme requires many hyperparameters. Although previous research has investigated the efficiency of a deep JSCC compared to traditional transmission schemes, the sensitivity of the model to differing numbers of layers has not been investigated. We have shown that a deep JSCC scheme’s sensitivity to changes in the number of layers is affected by the value of the Signal to Noise Ratio (SNR) for which it is trained. Utilizing code from GitHub for a deep JSCC, we modified the number of layers in the script and trained and evaluated the model to observe the impact. Evaluating models with 3, 5, and 8 layers, we observed varying sensitivity to the different number of layers at values of SNR for which the neural network was optimized. In addition, different regions of test SNR yield different performance comparisons in terms of PSNR. Thus, we observe that the value of the training and test SNR in a deep JSCC scheme can impact how the model reacts to changing numbers of layers.