Title: Exploiting Nontrivial Connectivity for Automatic Speech Recognition

Abstract: Nontrivial connectivity has allowed the training of very deep networks by addressing the problem of vanishing gradients and offering a more efficient method of reusing parameters. In this paper we make a comparison between residual networks, densely-connected networks and highway networks on an image classification task. Next, we show that these methodologies can easily be deployed into automatic speech recognition and provide significant improvements to existing models.