Supervision: Nathanaël Perraudin,Michaël Defferrard

Project type: Semester project (master)


Deep learning has become one of the most important tools in Machine Learning as it has imposed itself as the solution for many applications. The emergence of several libraries has tremendously simplified the implementation of a neural network. Nevertheless, in order to have accurate results and a tractable training session, one still need a clever architecture design, a good data preprocessing and a sufficient amount of data.

In this project, the student will face these challenges through the practical implementation of a score generator. The student will first have to build a dataset of scores and face the challenges related to real data, e.g. the quality versus quantity trade-off or issues related to data cleaning. He will then have to transform the score into a neural network friendly format, i.e. a distributed representation will have to be devised and tested. Once he reached this milestone, he will design a recurrent neural network able to capture the essence of the score. He will finally train this network to produce new scores and evaluate them qualitatively.

For evaluation, the student is asked to present two web pages: (1) a blog-like entry will explain the technical aspect of his project and (2) a live demonstration will generate new scores on demand.