Research (Selected Topics)
Causal discovery is the problem of inferring the causal relationships from observational data. It is a fundamental task in various disciplines of science and engineering. We are working
on computational methods to discover causal relationships from real-world complex data, with theoretical guarantees.
Causal Machine Learning
Current machine learning is mostly driven by statistical correlations. However, relying on pure statisitical correlations makes machine learning models hard to generalize to new domains and tasks. Some argue that this requires learning not only the statistical correlations, but also the causal generative process behind the data. We are exploring causal systems' properties and how they could benefit knowledge transfer in machine learning. In addition, we are interested in improving the fairness, robustness, and interpretability of machine learning models from a causal perspective.
3D Computer Vision
3D computer vision aims to model and understand the visual world by inferring 3D structures from 2D images. We are interested in deep learning methods for depth estimation, 3D reconstruction, 3D model retrieval, etc. In addition, we are leveraging the causal process from the 3D world to 2D images to improve the generalization of 3D reconstruction models.