I'm Harry Bendekgey,
a fifth-year computer science Ph.D. candidate at UC Irvine advised by Erik Sudderth. My primary research area is probabilistic machine learning.
I love board games, paddleboarding and sufing, and my dog Claude.
Academic Interests
-
Structured Generative Modeling
Generative models are not just useful for generating synthetic data; they also help us understand the underlying factors of variation in our data. My work aims to build generative models that are interpretable and give us insight into the data we train it on.
-
Inference in Generative Models
How can we leverage trained graphical models to solve new inference questions at test-time? Often times, this requires a careful treatment of multimodality. We aim to improve inference methods to impute missing data, estimate meaningful underlying parameters, and cluster observations.
-
Fair Machine Learning
As decision-making processes become increasingly automated, ethical questions about model fairness are no longer thought experiments. Left unchecked, algorithms can exacerbate disparities between genders, racial groups, and other legally-protected categories of identity.
Publications
| Unbiased Learning of Deep Generative Models with Structured Discrete Representations. H Bendekgey, G Hope, E Sudderth. NeurIPS 2023. ArXiv Link |
| Scaling study of diffusion in dynamic crowded spaces. H Bendekgey, G Huber, and D Yllanes. ArXiv Link Talks and Presentations: March Meeting 2022 |
| Scalable & Stable Surrogates for Flexible Classifiers with Fairness Constraints. H Bendekgey, E Sudderth. NeurIPS 2021. Talks and Presentations: Southern California Machine Learning Symposium 2021 |
| Clustering Player Strategies from Variable-Length Game Logs in Dominion H Bendekgey. AAAI Workshop on Knowledge Extraction from Games (KEG), 2019. ArXiv Link |