Candidate: Shayan Shirahmad Gale Bagi
Title: Causal Representation Learning for Out-of-Distribution Generalization
Date: July 11, 2024
Time: 1:30 PM
Place: REMOTE ATTENDANCE
Supervisor(s): Crowley, Mark - Czarnecki, Krzysztof
Abstract:
Conventional supervised learning methods primarily rely on statistical inference and assume identically and independently distributed (i.i.d) data. However, this assumption often fails in real-world applications, where environments or domains frequently exhibit changing factors, leading to challenges in model robustness and generalization. Additionally, statistical models are typically treated as black boxes, with their learned representations remaining opaque and difficult to interpret.
Addressing these issues from a causality perspective, my research aims to develop methods that enhance the interpretability and adaptability of machine learning models in dynamic and uncertain environments.
I have proposed innovative methods for learning causal models that can be effectively applied to various machine learning tasks, including transfer learning, out-of-distribution generalization, reinforcement learning, and action classification. The first method introduces a generative model designed to learn causal variables in applications where the causal graph is known, such as Human Trajectory Prediction. This approach leverages domain knowledge to model the underlying causal mechanisms, leading to improved performance on both synthetic and real-world datasets. The results demonstrate the generative model's ability to outperform traditional statistical models, particularly in scenarios involving out-of-distribution data.
The second method focuses on learning causal models in applications where the causal structure is unknown, a more challenging yet common scenario. I have explored various conditions and assumptions that enable the discovery of causal relationships without prior knowledge of the causal graph. This method integrates advanced techniques in causal inference and machine learning to construct models that can adapt to new, unseen environments. The evaluation on real-world and synthetic datasets indicates that the proposed method not only surpasses existing approaches but also provides a robust framework for causal discovery, paving the way for more reliable and interpretable AI systems.
Overall, my research contributes significant advancements in the field of causal learning, offering novel solutions that enhance model interpretability and robustness. These methods provide a solid foundation for developing AI systems capable of adapting to diverse and evolving real-world conditions, thereby expanding the applicability and effectiveness of machine learning across various domains.