An exaggerated preference for simple neural network models of signal evolution?