Yiwen Qiu

Scientific discovery often requires uncovering underlying mechanisms that govern complex systems, going beyond mere correlations to identify causal structures. Traditional machine learning methods, while powerful for prediction, typically fail to capture such causal mechanisms, limiting their ability to explanable results and robust generalization across domains. Causal representation learning is a promising paradigm that aims to extract meaningful latent factors that correspond to causal variables, therefore holds the potential to facilitate scientific discovery.

I will investigate the theoretical foundations of causal representation learning, developing methods that clarify its capabilities and boundaries. Furthermore, I will address the challenges posed by real-world applications, such as selection bias, measurement noise, high-dimensional observations, partial observation, and. By bridging theory and practice, this research aims to provide both novel methodological contributions and practical tools for applying causal representation learning in scientific domains. Ultimately, the project seeks to demonstrate how causal representation learning can serve as a rigorous and reliable framework for accelerating discovery in complex real-world systems.