Generative Multitask Learning Mitigates Target-Causing Confounding

Explore a groundbreaking approach to causal representation learning for multitask learning in this lecture by Kyunghyun Cho. Discover how a simple and scalable method can improve robustness to prior probability shift by mitigating unobserved target-causing confounders. Learn about the limitations of conventional multitask learning approaches and how the proposed method addresses these issues by considering the dependency between targets. Gain insights into the practical implementation of this technique, which requires minimal modifications to existing machine learning systems. Examine the results from experiments on the Attributes of People and Taskonomy datasets, demonstrating the conceptual improvement in robustness to prior probability shift. Delve into the speaker's background as an associate professor at New York University, CIFAR Fellow, and senior director at Genentech, and understand how his diverse experience informs this innovative approach to multitask learning.