Martin Wattenberg: ML Visualization and Interpretability

Interaction Models

Martin Wattenberg discusses the challenges and opportunities in designing interaction models for language systems. He emphasizes the importance of transparency, noting that even partial transparency can significantly enhance user understanding and trust in AI systems 1. Wattenberg believes that future interfaces should allow users to not only observe but also influence AI behavior, creating a two-way interaction process. He states, "The ultimate thing is going to be to tune the model, to work with the experience that we want" 1.

   

Dashboard Design

In discussing AI dashboard design, Wattenberg highlights the potential of novel interfaces to improve user interaction with AI systems. He explores the idea of AI as aggregators of statistics versus builders of internal world models, a concept that remains scientifically intriguing 2. Wattenberg reflects on past and present conversation interfaces, noting that while visual elements have been integrated, the fundamental design has seen little evolution. He remarks, "It's not clear that we have an improvement" when comparing modern interfaces to older designs 3.