Abstract by Paul Bodily
Pop*: Using Concept Learning to Compose Lyrical Music
Recent advances in human-level concept learning have achieved new levels of performance for one-shot learning in classifying and generating hand-written characters. Can this approach be used as a general model for creating computer creativity? To test this question we have designed a concept learning approach to the task of lyrical song-writing. Using principles of decomposition and learning-to-learn, our model learns the concept of a lyrical song with sufficient specificity and generalizability to create new songs that exhibit novelty, value, and surprise. The concept learning approach is well-suited for computational creativity given that variables representing inention and meaning are explicitly modeled.