Representation Learning for Drum Loops with a Variational Autoencoder

A representation is learned for MIDI drum loops, using a variational autoencoder. The aim is to create a representation which will be useful as a component in human-computer interfaces and in music generation systems. A large library of MIDI drum loops is described and used to train an autoencoder neural network in an unsupervised fashion. The result is a low-dimension representation which captures essential dimensions of variation in the data, and can be used to generate new drum loops and interpolate between pairs of loops