Thesis code: 20005
Thesis Type: M.Sc. thesis in Machine Learning, Data Science, Computer Science, Mathematics, or equivalent
Research Area: Data Science for Industrial and Societal Application
- Knowledge of Python
- Software development skills
- Basic concepts on data science, concerning data analysis, processing and machine learning
- Basic concepts on generative models
Nowadays we have seen generative adversarial networks very successful in creating images, in fact they have proven themselves to be capable of creating hyper-realistic faces, animating paintings, colorizing sketches, etc… However, these models cannot handle only images but also text and audio. Anyway, the latter is a field that remains largely unexplored. For this reason, we want to try to build generative models capable of creating new and unpublished melodies given a set of sounds as input data. The constrain is that sounds created must also have a melody that is harmonious, not just a random sequence of the input sounds. As a bonus we would like to consider the creation of melodies based on a target feeling that we decided to arouse.
The objective of this thesis consists in the study and implementation of machine learning algorithms useful for generating new ad unreleased melodies starting from a set of basic sounds. The proposed algorithms will be trained using a collection of open dataset and private data provided by a company in the sector, with which the candidate will have the opportunity to interact during his thesis work. The candidate will have both the task of creating the training dataset and studying and evaluating the best algorithms to apply for the case of study.
The candidate is required to implement deep adversarial learning algorithms using popular framework (TensorFlow, PyTorch, Keras, etc..).
Contact: send a resume with attached the list of exams to firstname.lastname@example.org specifying the thesis code and title.