Using Echo State Networks for modelling musical improvisation

Echo State Networks (ESNs) constitute an emerging field within Machine Learning research. As opposed to other Recurrent Neural Networks, they are compu-tationally cheap, and thus effectively offer a much greater repertoire of dynamics. Owing to their ability to retain information about long previous history as the " echo " states, coupled with their mathematical simplicity and ease of use, ESNs outperform other blackbox modeling techniques by orders of magnitude. Their application ranges from pattern recognition, through communication channel equi-libration, to detection of repeated motifs in melody-like sequences. This thesis extends and further explores the use of ESNs to pick up and reproduce melodic patterns. By greatly increasing the memory size and input range, we construct an ESN capable of detecting a " backbone " of a melodic pattern presented consecutively in a few variations, which then, depending on the setup, continues to produce either the backbone or random variations thereof, resulting in a kind of minimal music (represented by composers like Phillip Glass). We investigate the impact of the characteristic of the presented melody on the network's ability to pick up the " backbone " , and the influence of various parameter setting on melody generation. As a result, we suggest a biologically plausible model mimicking, in a restrictive sense, the compositional and improvisational abilities of musicians and investigate its properties, providing insights into the kind of processes that may be going on in human brains.