Automatic Music Generation Using Variational Autoencoders

The aim of the thesis is the design and evaluation of a generative model based on deep learning for creating symbolic music. Music, and art in general, pose interesting problems from a machine learning standpoint as they have structure and coherence both locally and globally and also have semantic content that goes beyond the mere structural problems. Working on challenges like those can give insight on other problems in the machine learning world. Historically algorithmic music generation focused on structure and was achieved through the use of Markov models or by defining, often manually, a set of strict rules to be followed. In recent years the availability of large amounts of data and cheap computational power led to the resurgence of Artificial Neural Networks (ANN). Deep Learning is machine learning based on ANN with many stacked layers and is improving state of the art in many fields, including generative models. This thesis focuses on Variational Autoencoders(VAE), a type of neural network where the input is mapped to a lower-dimensional code that is fit to a Gaussian distribution and then tries to reconstruct it minimizing the error. The distribution can be easily sampled allowing to generate and interpolate data in the latent space. Autoencoders can use any type of network to encode and decode the input, we will use Convolutional Neural Network (CNN) and Recurrent Neural Netowrks (RNN). Since the quality of music and art in general is deeply subjective and what seems pleasing to one may not be for another we will try to determine the “best” model by conducting a survey and asking the participants to rate their enjoyment of music and whether or not they think each sample to be composed by a human or AI.