Comparison Between CNN, ViT and CCT for Channel Frequency Response Interpretation and Application to G.Fast

Abstract

Convolutional Neural Networks (CNN) and more recently Visual Transformers (ViT) have been heavily used in specific areas like computer vision. Through this work, we explore and compare the CNNs and ViT models applied to a telecommunication signal, more specifically to interpret a G.fast channel frequency response. As both CNNs and ViT bring advantages, we have deepened the research by using a combination of both convolutions and transformers using Compact Convolutional Transformers (CCT) models. This study demonstrates that using transformer based models on a 1-D signal processing use case, we have significantly gained in accuracy compared to traditional convolution based models.