Roxie (Ruoxi) Jiang

I am a Computer Science PhD student at the University of Chicago advised by Professor Rebecca Willett.

Previously, I received my master's degree in Operations Research at Columbia University and bachelor's degree at Xi'an JiaoTong University.

Email  /  Github

My research interests lie in machine learning for dynamical systems and its applications in scientific computing. In particular, I work on learning structural hidden representations of the high-dimensional data. My work has been applied to addressing inverse problems with uncertainty quantification and designing practical algorithms to achieve data efficiency in decision-making problems (i.e., bandits). Currently, I am interested in predicting high-dimensional chaotic systems with deep learning.

We introduce a novel likelihood-free inference method based on contrastive learning that efficiently handles high-dimensional data and complex, multimodal parameter posteriors.

We introduce nested diffusion models, an efficient and powerful hierarchical generative framework that substantially enhances the generation quality of diffusion models, particularly for images of complex scenes.

We introduce a novel neural network parameterization that induces low-rank simplicity bias and enhances semantic feature learning in generative representation learning frameworks. By incorporating a weighting factor to reduce the strength of identity shortcuts within residual networks, our modification increases the MAE linear probing accuracy on ImageNet from 67.8% to 72.7%, while also boosting generation quality for diffusion models.

We introduce two novel approaches: Optimal-transport(OT) based method with prior knowledge of the phsycial property; and Contrastive learning (CL) based method in absence of prior knowledge, to match long-term statistics of chaotic dynamical systems with noisy observations.

We develop a novel deep-learning-based approach to solve time-evolving Schrödinger equation, with inspiration from stochastic mechanics and diffusion models.

A novel contrastive framework for learning feature embeddings of observed dynamics jointly with an emulator that can replace high-cost simulators for parameter estimation.

To overcome the curse of dimensionality, we propose to adaptively embed the feature representation of each arm into a lower-dimensional space and carefully deal with the induced model misspecifications.

Source code from Dr.Jon Barron.