Machine Learning Uncertainty as a Design Material: A Post-Phenomenological Inquiry

Title of the Paper

Using Machine Learning Uncertainty as a Design Material: A Postphenomenological Inquiry

Paper Information

• Subject Area: Human-Computer Interaction (HCI), Design Research, Postphenomenology, Machine Learning
• Keywords: Postphenomenology, Machine Learning, Design Research, Uncertainty, Pattern Leakage, Futures Creep, AI Design

Research Background and Issues

• Issues and Challenges:

  1. Current design research lacks sufficient understanding and exploration of machine learning (ML), especially in viewing ML uncertainty as a design opportunity rather than a barrier.
  2. Most studies focus on improving user experience and algorithmic transparency, rather than directly investigating how ML uncertainty shapes human experience.
  3. Existing design approaches (e.g., Explainable AI, XAI) tend to explain ML uncertainty rather than leverage it.

• Significance:

  1. Machine learning is rapidly integrating into daily life, and its technical opacity and statistical decision-making complexity demand design-level interventions.
  2. Uncertainty is an inherent attribute of ML systems, and neglecting its design potential may limit the emergence of novel design methods and interaction paradigms.

• Research Motivation and Related Work:

  1. The authors highlight how ML's statistical inference characteristics, through its data models and decision-making processes, shape human daily experiences in opaque ways.
  2. The current technical community focuses on reducing uncertainty through algorithmic interpretation rather than embracing it as a source of innovation.

Solution

• Proposed Approach/Solution:

  1. The authors adopt postphenomenological theory to redefine ML uncertainty, considering it as an essential and inevitable design material within machine learning technologies.
  2. By analyzing four research cases across different design methodologies, the authors derive three core concepts: thingly uncertainty, pattern leakage, and futures creep.

• Innovations:

  • Reframing ML uncertainty as part of design research, transforming it from an "obstacle to be explained" into a creative and socially meaningful design material.
  • Focusing on uncovering phenomena from technological uncertainty to deepen the understanding of the relationship between ML-driven systems and human experience.

• Implementation Steps and Key Techniques:

  1. Conduct postphenomenological analysis of four design research cases to reveal the unique characteristics of ML technologies and the uncertain phenomena they generate.
  2. Summarize relevant experiences and design observations, constructing a conceptual framework to describe and operationalize these phenomena.
  3. Propose "thingly uncertainty," "pattern leakage," and "futures creep" as theoretical and practical guides for design research.

Research Outcomes

• Specific Outcomes:

  1. Introduction of three key concepts:

     • Thingly Uncertainty: Describes how ML-driven devices interact with the environment and humans, serving as carriers of uncertainty and triggering new interaction relationships.
     • Pattern Leakage: Refers to the overflow of probabilistic patterns learned by ML models into the world, reshaping how things are perceived and interpreted.
     • Futures Creep: Explores how ML-driven technologies alter human relationships with time through prediction.

  2. Provides an analytical framework suitable for design inquiry, particularly extending the postphenomenological "human-technology-world" relationship.

• Comparison with Existing Solutions:

  • This study differs from XAI methods that emphasize rational explanation and transparency, as well as engineering approaches that solely optimize user experience. It focuses on intervening in ML uncertainty through design methods, opening up new spaces for interaction and reflection.
  • Views uncertainty as a potential source of creativity rather than an error or a factor to be avoided.

• Experimental or Evaluation Results:

  • Analysis of the four cases demonstrates that ML uncertainty can be both a manifestation of technical limitations and a driver of design innovation. For example:
    • In the "smart home camera" case, uncertainty prompted users to reflect on data pattern recognition and social privacy.
    • In the "high-water pants predicting future sea level rise" case, ML uncertainty altered participants' understanding of geographic environments through future time predictions.

• Limitations and Future Directions:

  1. Limitations:
     • The scope of case studies is limited and does not fully encompass the diverse applications of ML technologies in broader design fields.
     • Ethical and socio-political reflections are not deeply explored.
  2. Future Directions:
     • Expand design experiments on "pattern leakage" and "futures creep" to explore their application potential across different industries and environments.
     • Investigate how ML uncertainty can be integrated into human-centered design methods, combining ethical considerations to promote responsible design practices.