My research focuses on probabilistic circuits, investigating how the geometry of the log-likelihood landscape affects generalization. I develop curvature-based measures for PC log-likelihood surfaces to relate sharpness to overfitting. My work integrates theory, algorithm design, and density-estimation experiments on structured PCs. Currently, I develop sharpness-aware training and geometry-based regularization efficiently for probabilistic models.

ckn

https://seekayan.github.io/