Machine Learning Processes As Sources of Ambiguity: Insights from AI Art

Title of the Paper

Machine Learning Processes as Sources of Ambiguity: Insights from AI Art

Paper Information

• Domain: Artificial Intelligence Art and Human-Computer Interaction Design
• Keywords: Ambiguity, Machine Learning, Artificial Intelligence, Art, Computer Vision, Generative Art

Research Background and Problem

• Issues and Challenges:

  • Current efforts in the Human-Computer Interaction (HCI) field to transform Machine Learning (ML) into a design material have shown limited success.
  • The design challenges posed by the "capability uncertainty" and "output complexity" of machine learning create difficulties for industry designers.
  • HCI still requires new frameworks and design standards to better define the user experience of ML.

• Significance of the Research:

  • Machine learning is at the core of many future technologies, making it crucial to understand how its uncertainties impact design.
  • The rapid rise of Artificial Intelligence Art (AI Art) offers HCI a new perspective to observe and learn how ML can serve as a creative medium.

• Research Motivation and Related Work:

  • AI artists create works using computer vision and image generation technologies, showcasing types of ambiguity that trigger multiple interpretations by users.
  • Ambiguity is regarded in both HCI and traditional art fields as a valuable design resource that stimulates user interpretation, innovation, and reflection.

Solution

• Core Methods and Contributions:

  • Analyzed nine AI art pieces to explore how artists utilize various stages of ML (dataset management, model training, and application) to evoke ambiguity.
  • Applied the ambiguity theory proposed by Gaver et al. to AI art and extended it with a new category of "process ambiguity."
  • Challenged mainstream demands in HCI for ML, such as the focus on reliability and explainability, and proposed an art-based alternative design perspective.

• Innovations:

  • Emphasized treating the process of machine learning itself as a design element, rather than merely a technical detail.
  • Introduced "process ambiguity" as a new category, shifting the focus from artifacts to their creation and design processes.

• Implementation Steps and Key Techniques:

  • Adopted a humanistic interpretive analysis and art critique approach to conduct a detailed examination of the ambiguity features in nine AI art pieces.
  • Explored how artists inject ambiguity at each of the three core stages of the ML pipeline (dataset management, model training, and application) to provoke diverse interpretations from the audience.

Research Findings

• Specific Findings:

  • Identified eight main techniques used by artists when creating with ML, such as selecting datasets with problematic labels, deliberately halting at suboptimal model training stages, and applying models to different domains.
  • Proposed "process ambiguity" as an additional category of ambiguity, demonstrating the uncertainty and openness of ML art experiences at both technical and interpretative levels.

• Advantages Over Existing Solutions:

  • Compared to traditional product-centered frameworks, process-oriented ML design better reveals the ethical, cognitive, and creative potentials behind the technology.
  • Rejected the traditional design assumption that ML systems should completely eliminate errors, proposing instead that errors can be embraced and even leveraged to create new experiences through design.

• Experimental or Evaluation Results:

  • All nine analyzed art pieces demonstrated applications of ambiguity within deep neural network models and ML data flows.
  • For instance, ImageNet Roulette highlighted the challenges of data uncertainty by questioning the racial biases embedded in datasets.
  • Learning to See transformed common objects into natural phenomena perceived by humans, successfully exploring novel forms of ambiguity in datasets and application contexts.

• Limitations and Future Directions:

  • The study analyzed a limited number of art pieces, primarily focusing on works by established artists, without addressing the more diverse experiments of amateur creators.
  • Future research is advised to expand the sample scope, exploring potential patterns among artists from more regions and cultural backgrounds.
  • Further examination is needed on how ambiguity strategies can be applied to commercial design fields, such as gaming, home entertainment, and educational platforms.

Conclusion

By combining novel perspectives from HCI and AI art, this research not only enriches HCI's understanding of ML but also proposes the potential to uncover design value within uncertainty. This study contributes to rethinking the user experience design of ML, opening new discussions at the intersection of technology and creativity.