Operationalizing Essential Characteristics of Creativity in a Computational System for Music Composition

Paul M. Bodily, Dan VenturaAAAI 2024 · DOI

TL;DR

Researchers have developed Pop*, an autonomous music composition system designed around seven core pillars of creativity rather than just "next-token prediction." The study reveals a fascinating psychological insight: human evaluators perceive "creativity" as an emergent property that is greater than the sum of its individual parts, like novelty or technical skill.

The Challenge: Generativity vs. Creativity

While Large Language Models (LLMs) and foundation models are excellent at generating content, the authors argue that "generativity" is not the same as "creativity." In highly structured domains like music, current AI often lacks intentionality, self-evaluation, and a cohesive "aesthetic."

The challenge lies in computational creativity (CC)—moving beyond systems that produce artifacts as a side-effect of training, toward systems that intentionally operationalize the characteristics of creative thought. The authors seek to define these characteristics and test whether a system built on these pillars is actually perceived as creative by humans.

The Pop* System: A Modular Architecture

The authors introduce Pop (Pop Star), a system that composes Western-popular music lead sheets. Unlike a black-box neural network, Pop is a modular architecture designed to address seven specific characteristics of creativity.

The modular architecture of the Pop* system, illustrating the flow from external social media input to final musical output.

The system's workflow is operationalized through several key components:

  1. Generation: Uses Non-homogeneous Markov Models (NHMMs) to handle local transitions while maintaining global constraints.
  2. Knowledge Representation: Employs a Hierarchical Bayesian Program Learning (HBPL) model to represent concepts like structure, chords, and rhythm.
  3. Intentionality: Pop searches Twitter for emotionally charged tweets, using the Empath library* to transform these into a vector of "intentions" (e.g., sadness, optimism) that guide the composition.
  4. Aesthetic: The system maintains a list of "interests" to filter which tweets it chooses to respond to.
  5. Domain Knowledge: It leverages two massive databases (lyrics and lead sheets) and uses a multi-Smith-Waterman (mSW) alignment algorithm to detect and replicate musical motifs and sectional forms.
  6. Autonomy: Pop* independently selects its own themes, training data, and final outputs without human intervention.
  7. Self-Evaluation: Before outputting a song, Pop* scores candidate compositions based on how well they match the intended emotion, lyric variety, and "sing-ability."

Key Insights: Creativity as Emergence

The most striking finding was that participants rated the system's "overall creativity" higher than any of its individual components (like its ability to write lyrics or follow a structure). This suggests that when these seven traits are present together, they create a holistic impression of creativity that exceeds the sum of the parts.

However, evaluators often struggled to attribute "autonomy" or "aesthetic" to the system simply because they knew it was a machine. One respondent noted, "It’s a machine, it has no autonomy or beliefs," highlighting a persistent philosophical bias in how we judge AI.

Evaluation and Results

The authors conducted a survey of 125 people (including 68 musicians) to evaluate Pop*’s output and process. The results showed that the system was perceived as creative across various demographics.

Survey results for questions Q8-Q12, categorized by musicianship and belief in computer creativity.
  • Overall Creativity (Q12): Received a mean score of 3.38/5, which was higher than the scores for individual traits like lyric composition (2.09) or music composition (2.98).
  • Novelty: This was the highest-rated individual artifact trait (avg. 3.16–3.28), suggesting the system is successful at creating "new" things.
  • Musician Critique: As shown in the data, professional musicians were significantly more critical of the system’s technical musicality than non-musicians (3.25 vs 3.53 for overall creativity), yet they still rated its "autonomy" and "knowledge representation" above average.

Limitations and Future Work

The authors identify several areas for improvement:

  • Semantic Cohesion: While the lyrics are "novel," they often lack the deep semantic meaning found in human-written songs.
  • Aesthetic Evolution: The current aesthetic model is relatively static; future versions could allow the system to "evolve" its tastes over time.
  • Closing the Intention Gap: Better methods are needed to ensure the system's internal "mood" is more effectively translated into musical features that humans can readily identify.

By moving toward "holistic CC," the authors argue we can move past "mere generation" and closer to the "final frontier" of artificial intelligence.

Summary updated

BibTeX
@inproceedings{bodily2024operationalizing,
  author = {Paul M. Bodily and Dan Ventura},
  title = {{Operationalizing Essential Characteristics of Creativity in a Computational System for Music Composition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i1.27799},
}
Figure