Abstractnet: a generative model for high density inputs

Abstract

This paper introduces AbstractNet, a generative model for high density inputs. The model suggests a method that uses unsupervised learning to generate feature maps. The model drastically improves the performances of raw audio generation by reducing the required amount of input data and computing power necessary to achieve a similar result when compared to the state of the art.