(she/her/hers)
Georgia Institute of Technology
Machine Learning Interpretability, Visualization, Scalable Data Analytics
Haekyu Park is a Computer Science PhD student at Georgia Institute of Technology, working at the intersection of visualization, machine learning, and scalable graph analytics. Her human-centered AI research develops novel interactive scalable interfaces that broaden people's access to AI technologies, help them more easily interpret complex models, gain trust in performant ones, and fix those that malfunction. Her research has resulted in open-sourced interfaces for AI interpretability (e.g., NeuroCartography for interpreting concepts learned in deep learning model, which was invited to present at SIGGRAPH'22 as a top VIS'21 paper, 1% out of 442 submissions), NVIDIA's Accelerated Data Science Teaching Kit (a first-of-its-kind GPU-accelerated data science curriculum), pioneering open-source tools like the viral CNN Explainer that transforms AI education, and the Argo Lite graph visualization used by thousands of students. Haekyu's dissertation is supported by the JPMorgan AI PhD fellowship. She works with Prof. Polo Chau.
Haekyu received her Bachelor's degree in Computer Science and Engineering from Seoul National University. She has worked with researchers, engineers, scientists, and designers at NVIDIA, Microsoft Research, and Stripe.
Human-centered AI: 
Interactive Scalable Interfaces for Trustworthy and Safe AI
Artificial intelligence (AI) technologies, especially Deep Neural Networks (DNNs), are now ubiquitous. Yet, our society still faces fundamental barriers to unleashing AI's full potential: (1) AI's lack of transparency makes people hesitant to trust and deploy them; (2) when models do not perform satisfactorily, people lack actionable guidance for understanding their vulnerabilities and how to fix them. My PhD thesis research addresses these significant challenges in AI through a human-centered approach, by creating novel interactive visualization tools to instill trust in well-performing AI models and to fix malfunctioning models.
Scalable Visual Discovery for Trustworthy and Interpretable AI.
I have been developing novel visual representations that serve as smart microscopes, helping people inspect a model's internals and how they lead to the model's decisions. Our Summit system visualizes and explains how DNNs recognize an object through the novel attribution graph, which summarizes what features are detected by neurons and how those features interact through the neural connections. Our NeuroCartography system further advances the state of the art by capturing the bigger picture of how concepts are collectively encoded by multiple neurons. It introduces two novel scalable summarization techniques to automatically discover and group neurons that detect the same concepts, and describes how such neuron groups interact to form higher-level concepts and the subsequent predictions. NeuroCartography was invited to present at SIGGRAPH'22 as a top VIS'21 paper (1% out of 442 submissions).
Insights to Protect and Troubleshoot Models.
To discover where and why models would fail, my idea is to focus on comparing how the neurons and their connections respond differently to normal inputs and abnormal inputs. For example, to understand DNNs' vulnerability to adversarial attacks, our Bluff system compares the attribution graph of benign inputs and adversarially perturbed inputs, to find out where a DNN model is exploited by the attack and what impact the exploitation has on the final prediction across multiple attack strengths.