Stanford CS329H: Machine Learning from Human Preferences | Autumn 2024 | Human-centered Design

Key Concepts

• Human-Centered Design (HCD): Designing technology with a focus on human needs, capabilities, and behaviors.
• Human-Computer Interaction (HCI): A discipline focused on the design, evaluation, and implementation of interactive computing systems for human use.
• Norman Doors: Doors with confusing or non-intuitive designs that make it difficult to understand how to open them.
• Affordance: A property of an object that indicates how it can or should be used.
• Design Thinking: A problem-solving approach that emphasizes empathy, experimentation, and iteration.
• Double Diamond Method: A design process model consisting of two diamonds: one for problem discovery and definition, and one for solution development and delivery.
• Prompt Engineering: The process of crafting effective prompts for large language models to elicit desired responses.
• User Experience (UX): A person's perceptions and responses resulting from the use or anticipated use of a product, system, or service.
• Automation Bias: The tendency to over-rely on automated systems, even when they are incorrect.
• Mixed-Initiative System: A system where both the user and the computer can take the initiative in carrying out tasks and making decisions.

Human-Centered Design and Human-Computer Interaction

• HCI is a discipline that designs technology from a human perspective, contrasting with traditional engineering and computer science approaches that often prioritize solving engineering problems directly.
• The core idea is to apply the scientific method to the art of design, focusing on eliminating obstacles that hinder user interaction.
• Example: Norman doors illustrate the importance of intuitive design. Users shouldn't struggle to understand how to operate a door (push, pull, slide).
• The lecture argues that difficulties in using technology are often due to bad design, not user incompetence.
• Learning from human preferences is a powerful tool for building AI technologies that work better for people.
• Analogy: Just as a door's design should make it obvious how to enter a room, AI tools should be designed to make it easy for users to accomplish their tasks.
• The current state of AI, where users must craft "magic prompts" to get desired results, is considered a design failure.
• Well-designed technology should be so intuitive that users are barely aware of the effort that went into its creation.
• Quote: "The most profound technologies are those that disappear. They weave themselves to the fabric of everyday life until they're indistinguishable from it." - Mark Weiser

Bridging the Gap Between Humans and Computers

• HCI aims to bridge the gap between human capabilities and computer functionality.
• This involves both extending computer capabilities and adapting technology to better suit human needs.
• Evolution of Interfaces: From punch cards and command lines (requiring high expertise) to graphical user interfaces (GUIs), pointing devices, and touchscreens (more intuitive).
• The goal is to make technology ubiquitous and easy to engage with.
• Examples: Icons that resemble the actions they perform, intuitive phone interfaces.
• Current efforts in VR/AR are aimed at making computing interfaces even easier to use.
• In the AI world, the focus is shifting from requiring expert coding skills to using prompting and chatting as interfaces.
• The chasm between AI capabilities and user-friendliness is wider than in general computing.

Technology-Centric vs. User-Centric Design

• Technology-Centric Design: Focuses on developing technology for its own sake, often with an abstract notion of users.
• User-Centric Design: Centers on the user's needs and problems, aiming to make it easy for them to solve their problems efficiently.
• Technology-centric design can lead to searching for applications for existing tools, while user-centric design starts with the problem and designs a solution.
• Chatbots are a mix of both, adapting technology to solve user problems but sometimes requiring complex prompting.

Design Thinking and the Double Diamond Method

• Design thinking challenges traditional engineering approaches by asking "why" a problem needs solving, rather than immediately seeking solutions.
• Quote: "If I had asked people what they wanted, they would have said faster horses." - Henry Ford
• Design thinking encourages deep introspection to identify the root cause of a problem.
• The Double Diamond Method is a popular framework:
  • First Diamond (Problem):
    • Discover: Research and gather information to understand the problem broadly.
    • Define: Synthesize the information to identify the root cause of the problem.
  • Second Diamond (Solution):
    • Develop: Ideate and prototype potential solutions.
    • Deliver: Implement and refine the final solution.
• Discovery Phase: Involves field studies, interviews, surveys, and market research to understand the context and needs.
• Definition Phase: Uses techniques like participatory design and affinity diagrams to narrow down the root cause.
• Development Phase: Employs storytelling and prototyping to explore different solutions.
• The process is iterative, involving observation, idea generation, prototyping, and testing.

Human-AI Interaction (HAI)

• HAI is a field focused on the interaction between humans and AI systems.
• It considers various human stakeholders (AI researchers, developers, domain experts, end-users) and AI technologies (language models, dialogue systems, etc.).
• The interaction can be collaborative or assistive.
• The field aims to bridge the gap between human stakeholders and AI tools to solve real-world problems.
• Key Steps:
  1. Design: Determine why the interaction should occur, what needs to be done, and in what context.
  2. Enable: Use tools and techniques to facilitate the interaction (e.g., personalization, preference learning).
  3. Evaluate: Assess the effectiveness of the interaction.

Prompt Engineering and Hybrid User Interfaces

• Prompt engineering, while freeing, can be challenging as a way to interface with AI.
• It can be difficult to achieve the clarity and reliability of traditional UX design.
• Quote: "Prompt engineers exist to tickle ChatGPT in the right spot so that it coughs up the answers."
• Successful prompting often requires highly stylized and specialized formats.
• Hybrid User Interfaces: Combining GUIs with prompts can improve usability.
• Example: PromptCharming: A system for visual language model that addresses the conceptual gaps when novice users write prompts for image creation.
  • Automatic revision of prompts.
  • Easy ways of finding similar items.
  • Weighing of terms.
  • Support exploration.
  • Quick prototyping.
  • Tracking what they've done.
  • Provide explanations.
  • Direct manipulation.

Building Interactive Systems

• Consider human cognition and perception to make tasks easier.
• Address trust and reliance on the technology.
• Ensure accountability, fairness, transparency, and ethics.
• Personalize and adapt the system to individual users.
• Provide feedback and guide interaction.

Trust and Reliance

• Trust: Belief in the reliability, integrity, and honesty of a system.
• Reliance: Dependence on a system to perform a function, regardless of trust.
• Goal: Achieve appropriate reliance, where users accept correct decisions and reject incorrect ones.
• Challenges: Under-reliance (rejecting correct decisions) and over-reliance (accepting incorrect decisions - automation bias).
• Strategies:
  • Provide explanations.
  • Show uncertainty.
  • Allow user agency.
  • Show the processing steps.

Fairness, Accountability, and Transparency

• Fairness: Ensure the system works equitably for diverse populations, avoiding bias and discrimination.
• Accountability: Address the consequences of mistakes and build that into the system's evaluation.
• Transparency: Explain how the system works to build trust and understanding.
• Ethics: Consider the broader ethical implications of the system's use.

Interaction Initiation

• Human-Initiated: Human uses the computer as a tool to be creative.
• Computer-Initiated: Computer deploys something automatically, human is the audience.
• Mixed-Initiative: Both human and computer take an active role in carrying out tasks and making decisions.
• Mixed-initiative systems require coupling automated services with direct manipulation to encourage creativity.

Evaluating Human-AI Interactions

• Key Questions: How, what, who, and when to evaluate.
• How:
  • Methods: Quantitative (task completion, accuracy) and qualitative (surveys, interviews).
  • Types: Intrinsic (model performance on its own) and extrinsic (model's helpfulness in a real-world task).
  • Metrics: Existing validated metrics or new metrics tailored to the context.
• What: The model, the HCI, or the whole end-to-end system.
• Who: Lay users, domain experts, or automated evaluation (e.g., LLM as a judge).
• When: Instant evaluation, short-term interaction, or long-term longitudinal studies.

Conclusion

• Human-AI interaction design aims to harness AI techniques to solve people's problems while respecting their needs, goals, and values.
• This involves enabling human-AI interactions through tools like preference learning and social choice.
• Evaluation is a critical part of the process, requiring careful consideration of the evaluation setup to best match the context.