First of all, let us introduce ourselves, the authors of this blog:
We are two students of the Mathematics, Informatics and Mechanics Faculty at the University of Warsaw, currently pursuing Master's degree. We decided to write a thesis together, due to a year-long successful cooperation in various machine learning projects and similar interest: combining artificial intelligence and art, especially - music.
Journey to the Master’s degree
Our goal during the next two years of the Master's degree studies in Computer Science is to create a system capable of generating music - a skill still mostly recognized as a solely human's ability. As a quite ambitious (in our opinion) Master's thesis, we decided to join our forces and bring the process of research in the project to a public audience. As far as we know, such transparency is quite extraordinary among students and so, we find it an interesting educational experiment as well!
The purpose of this blog
This blog is intended to both document our progress in the project and monitor the latest results in the fields of Music Generation and Music Information Retrieval. We are happy not to be alone in the community - just recently, a few projects, such as Flow Machines or Google's WaveNet and Magenta revealed a growing interest in this rather underdeveloped field. Proved to be quite successful in painting generation, machine learning is yet to provide some decent results in music. We hope to reproduce the current state of the art and wish to improve it further. If, by any chance, you also happen to work on something related to this topic, be sure to contact us! We will be happy to know the community better and perhaps involve in some sort of collaboration.
Our experience in the field
One of our most interesting projects in the field of MIR so far, is a Deep Neural Network used to recognize music genre of a song, trained and evaluated on the well-known GTZAN dataset. During a 24-hour hackathon, along with two other friends, Łukasz Margas and Jakub Królak, we trained the network and developed a nice web frontend. The model takes an MP3 file of a song and displays a live visualisation of its genre - a continuously updated genre probability distribution inferred by our model. Since the computations are quite expensive, we selected a few interesting outputs, which you can see here. Choose a song from the list and see for yourself! The inspiration, research and details on the final model will be described in the next article. We hope you will find it interesting and would love to hear your opinions!
As mentioned before, we are mostly interested in music generation. During the following weeks, after a proper introduction, we shall discuss newest research in generative models in general and music generation in particular, including our experience to this day.
Stay tuned! (pun intended)