Recent work in modulation classification using deep learning has produced promising results in classifying a diverse set of signals with only 128 IQ samples having undergone common radio impairments and frequency selective fading. Using deep learning as an approach differs significantly from established methods in classification where a set of expert features are derived and justified for specific channel conditions and computational resources. Instead, deep learning provides a way to efficiently identify features that are best suited for any set modulations, channel conditions, or available resources that may be of interest. However, deep learning approaches have only recently been used in this area and many questions about its limitations still exist. In this work, an expanded set of signals that include radar signals, multi -carrier signals, and higher order modulations are used with architectures presented in previous work to gain further insight into the flexibility of this approach. With a total of 29 signals to classify, it's shown how previous approaches can be augmented by training the network to classify signals on several hierarchical levels simultaneously with improved classification accuracy and a more flexible network architecture. Branch convolutional neural networks (B-CNN), which have been used to identify the subject of a body of text within a large hierarchy of subjects, are adapted to the problem of modulation classification to improve classification accuracy and facilitate the development of networks that can classify an even more diverse set of signals.