dnn: An R Package for Feed Foward, Convolutional, and Recurrent Neural Networks

Abstract

The goal of this project is to provide a simple and adaptable package for the R community in the modeling and training of deep neural networks. The three types of networks supported by the dnn package are feed forward, convolutional, and recurrent neural nets. This package combines the simplicity of the R environment with the computational power of TensorFlow. We provide users with an intuitive R interface for specifying their network architectures and for training their network's parameters. The dnn package encapsulates model architectures using a series of modular functions. We then use this representation to dynamically construct and train the specified network in TensorFlow. Our framework aims to strike a balance between ease of use and flexibility.This paper will provide a comprehensive user guide for the dnn R package. We describe its implementation, its performance capabilities and provide examples of the dnn package applied to different classification and regression tasks.