Domain: Human-Computer Interaction (HCI), Explainable Artificial Intelligence (XAI), and Machine Learning (ML).
Persona: Senior Research Scientist in Human-Centered AI and Interactive Visualization.
2. Reviewer Recommendation
A suitable group to review this topic would be the Program Committee for the ACM CHI Conference on Human Factors in Computing Systems or a Senior Review Board for Responsible AI (RAI) at a major technology firm (e.g., OpenAI, Microsoft, or Google).
3. Abstract and Summary
Abstract:
This PhD defense, titled "Democratizing Human-Centered AI with Visual Explanation and Interactive Guidance," details a multifaceted research framework designed to bridge the gap between complex AI systems and diverse stakeholders. The research is structured into three primary thrusts: explaining AI mechanisms to various audiences, enabling human-led guidance of AI behavior, and integrating these practices into standard development workflows.
The work introduces several high-impact tools and datasets. For explanation, it presents CNN Explainer, a browser-based visualization for learning deep learning concepts, and DiffusionDB/WizMap, a massive dataset and scalable visualization system for analyzing generative AI prompts. For guidance, it introduces GAM Changer and GAM Coach, which allow developers and end-users to edit model parameters and seek actionable recourse. Finally, for democratization, the Farsight tool integrates responsible AI envisioning directly into the prototyping phase. The research emphasizes the use of on-device computation to ensure privacy and scalability, aiming to shift the AI development paradigm from technical-centric to human-centric.
Summary of Research and Defense:
0:00 Research Mission: The core objective is to enhance human-AI interaction by building bridges through accessible explanation and human guidance, moving beyond purely algorithmic solutions to empower the "people part" of the AI equation.
3:52 CNN Explainer (Explaining to Novices): This tool addresses the difficulty students face in connecting mathematical mechanisms with model structures. It uses a live, browser-based simulation of a Convolutional Neural Network (CNN) to visualize neuron activations and sliding window convolutions.
9:14 DiffusionDB & WizMap (Explaining to Policy Makers): This section introduces the first large-scale text-to-image prompt dataset (14 million pairs) to identify potential harms like deepfakes and misinformation. WizMap is presented as a scalable system capable of visualizing millions of data embeddings directly in a browser to help practitioners find clusters of bias or specific usage patterns.
14:24 GAM Changer (Guiding AI with Human Values): Focusing on high-stakes domains like healthcare, this tool allows experts to perform "model editing" on Interpretable Boosting Machines (EBMs). It enables users to manually correct illogical model behaviors—such as a drop in predicted pneumonia risk for patients over age 100—by dragging and adjusting shape functions.
23:13 GAM Coach (Actionable Recourse): This system provides end-users with "algorithmic recourse." Unlike static counterfactual explanations, GAM Coach allows users to specify which constraints are actionable for them (e.g., changing a credit score vs. changing income) to achieve a favorable AI decision.
27:11 Farsight (Democratizing Responsible AI): To address "responsible AI" (RAI) awareness, Farsight is introduced as a widget that monitors prompts during the prototyping phase. It uses AI to envision potential harms, stakeholders, and second-order effects (e.g., a medical translation error leading to patient mistrust) before a model is deployed.
34:35 Evaluation of Farsight: A study with 42 AI prototypers demonstrated that the tool significantly increased the number of potential harms users could envision independently, suggesting a strong educational and preventative value.
38:40 Impact and Future Directions: The candidate highlights the integration of these tools into major platforms (Microsoft, Google) and academic curricula. Future work focuses on "interactive AI alignment" and leveraging on-device computation (WebGPU/WASM) for lower latency and better privacy in AI development tools.
41:51 Professional Trajectory: The candidate concludes by announcing a transition to a Research Engineer role at OpenAI, focusing on the continued advocacy for human-centered AI methodologies.