Criticism, Culture, and the Automatic Generation of Artworks

Lee Spector, Adam AlpernAAAI 1994 · OpenReview

Abstract

Researchers wishing to create computational systems that themselves generate artworks face two interacting challenges. The first is that the standards by which artistic output is judged are notoriously difficult to quantify. The larger AI community is currently involved in a rich internal dialogue on methodological issues, standards, and rigor, and hence murkiness with regard to the assessment of output must be faced squarely. The second challenge is that any artwork exists within an extraordinarily rich cultural and historical context, and it is rare that an artist who is ignorant of this context will produce acceptable works. In this paper we assert that these considerations argue for case-based AI/Art systems that take critical criteria as parameters. We describe an example system that produces new bebop jazz melodies from a case-base of melodies, using genetic programming techniques and a fitness function based on user-provided critical criteria. We discuss the role that such techniques may play in future work on AI and the arts.

Introduction

The field of AI-generated art faces two significant hurdles: the subjectivity of aesthetic beauty and the necessity of cultural context. Because there is no universal, quantifiable standard for "good" art, evaluating AI systems objectively is difficult. Furthermore, art is deeply rooted in historical context; an AI "artist" lacking this background is unlikely to produce meaningful work.

Rather than attempting to solve the philosophical problem of what constitutes "art," this paper proposes "factoring out" the critic. By building a system that takes a critic (a fitness function) and a culture (a case-base of prior works) as parameters, researchers can focus on the capability of the system to conform to specific constraints, regardless of whether any specific critic is "correct."

The "Constructed Artist" Framework

The proposed framework separates the creative engine from the subjective judgment process. As shown in Figure 1, the system utilizes a Genetic Programming (GP) engine to evolve programs that satisfy user-defined criteria while drawing from a library of existing works.

System architecture for the 'Constructed Artist' framework, illustrating the separation of the critic function and cultural case-base from the generation system.

This approach treats culture as a library of successful artworks, allowing the system to be architecture-independent but output-sensitive to specific cultural contexts. By evolving programs (processes) rather than just static melodies (products), the system discovers complex, multi-step transformations that a human programmer might not have considered.

Methodology: Evolving Musical Programs

The system employs Genetic Programming to evolve actual LISP programs rather than static data strings. In this context, the "artist" is a program that takes a melody as input and produces a new melody as output.

The Critic and Case-Base

The user provides formalizable criteria, such as the balance between novel and familiar tones, which serve as the fitness function. The system also has access to a "case-base"—a small library of five four-measure fragments from Charlie Parker songs. The GP function set includes musical transformations like AUGMENT (slowing down), INVERT (flipping intervals), and RETROGRADE (reversing), as well as operations that allow evolved programs to compare output to the case-base.

Musical Representation

Melodies are represented as lists of 64 numbers, corresponding to pitches for 16th notes. This "weak" representation, while limited in its ability to handle complex rhythms like triplets, provides a structured environment for testing the evolutionary process.

Evolutionary Results

The system demonstrated significant progress in satisfying the provided critic. Starting from a random population with an average fitness score of 7.43, the system evolved a "best-of-run" program by generation 19 with a fitness score of 2.82 (where lower scores indicate better performance).

Evolutionary progress showing the improvement of fitness scores for both the best individual and the population average over 20 generations.

The winning program was a sophisticated nested function that performed complex manipulations, including case-sensitive transpositions and interleaving slowed-down versions of the input melody. In one instance, the system produced a "call and response" pair that scored 0.19, very close to a perfect fitness of zero.

Generalization and Robustness

To test robustness, the evolved program was applied to three Charlie Parker melodies not included in the training set. The system produced fitness scores between 0.81 and 1.66. While these scores were higher (less fit) than those for the training data, they indicate that the system learned a robust strategy for pleasing its critic rather than simply memorizing the case-base.

Conclusion and Future Work

The "Constructed Artist" approach offers a clean separation between the science of building generative systems and the philosophy of aesthetic judgment. By treating the critic as an external parameter, researchers can focus on the system's ability to conform to diverse constraints.

Future work should address the limitations of the current musical representation and the small scale of the case-base. The authors suggest that future systems might even involve the automatic generation of critics by extracting patterns directly from cultural case-bases, further reducing the reliance on human-coded rules.

Summary updated

Abstract

Researchers wishing to create computational systems that themselves generate artworks face two interacting challenges. The first is that the standards by which artistic output is judged are notoriously difficult to quantify. The larger AI community is currently involved in a rich internal dialogue on methodological issues, standards, and rigor, and hence murkiness with regard to the assessment of output must be faced squarely. The second challenge is that any artwork exists within an extraordinarily rich cultural and historical context, and it is rare that an artist who is ignorant of this context will produce acceptable works. In this paper we assert that these considerations argue for case-based AI/Art systems that take critical criteria as parameters. We describe an example system that produces new bebop jazz melodies from a case-base of melodies, using genetic programming techniques and a fitness function based on user-provided critical criteria. We discuss the role that such techniques may play in future work on AI and the arts.
BibTeX
@inproceedings{spector1994criticism,
  author = {Lee Spector and Adam Alpern},
  title = {{Criticism, Culture, and the Automatic Generation of Artworks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
  url = {https://openreview.net/forum?id=HyVicCeuWr},
}
Figure