Improving AI-generated Music with Pleasure Models
Integrating the model into deep learning-based music generation techniques is a promising avenue towards more enjoyable computer-generated music. This is what the proposed research would attempt. In the proposed project, a deep learning architecture that could be trained to generate music with the quantified musical pleasure function as its objective function would be invented. In theory, this model would create music that maximizes the musical pleasure of a specific listener. The desired goal is the creation of an automatic music composition system that can generate music more comparable to human-composed music in terms of genuine likability than the current methods.
The topic evokes several intermediate questions that will be addressed in the course of the research, through a combination of empirical experiments and careful computational modelling. The accuracy of the musical pleasure predictions put forth by the uncertainty-surprise model will be validated empirically. The influence of factors external to the music itself on a listener's musical pleasure will also be investigated and perhaps integrated into the uncertainty-surprise model; mood and time of day, for example, affect the music a listener would prefer at a particular time. Finally, how to measure the success of a system at producing music that is actually pleasurable, rather than elevator-music-pleasurable, is an open question that will need to be resolved before performing an empirical study that compares the actual enjoyability of music created by the proposed model to that of the current music-generation methods.
Exploring these issues would not only aid in the achievement of the proposed project's end goal, but would also contribute to studies involving musical pleasure in general.
If successful, the proposed research will advance the field of autonomous music generation in the way that professionals have predicted. The authors of  "expect AI to blur the lines between music creation and music consumption by making it possible for the user to enjoy musical content that has been specifically produced for him/her, based on past choices and user history." An implementation of the proposed model is a step in this exact direction.
C4DM theme affiliation: