Session 3J: RNNintendo: Music Composition with Neural Networks

A fledgling technological with extraordinary potential, artificial neural networks have risen to prominence in the field of machine learning. Loosely inspired by their biological counterparts, neural networks function as computational systems composed of interconnected neurons. By exploiting certain architectural and computational properties of the system, neural networks are capable of a dizzying array of tasks, encompassing everything from computer vision to handwriting recognition to even language processing. This independent study focused on the application of a particular class of neural networks termed “recurrent neural networks” (RNN) to music composition. RNN’s are unique in that they rely on data from previous computations as well as fresh input. More specifically, we seek to use “Long Short-Term Memory” (LSTM) units, which differ from traditional RNN’s in that they leverage the use of “memory cells,” specific aspects of the architecture designed to preserve the state of the data across many transformations. As a result, LSTM’s are particularly well suited for applications like sentence comprehension and speech analysis, where data needs to be “remembered” across relatively long stretches of information. Thus, by applying LSTM’s to the task of music composition, we sought to produce coherent, aesthetic pieces generated purely by an artificial neural network trained on music from the Nintendo franchise