Dissertation Proposal by ZHENG Xiaosen | Understanding Behaviors of ML Models from Different Aspects

Modern Machine Learning (ML) models have demonstrated impressive capabilities in various fields including both computer vision and natural language processing. However, the inner workings of these models remain obscure, creating potential concerns for their use in various applications. Therefore, understanding these models' behavior is crucial. In this dissertation, we try to understand ML models' behaviors from different aspects. The first aspect is Data Memorization, which aims to understand the memorization behavior of deep neural network models. This helps us to understand models' generalization. The second aspect is Data Attribution, which aims to trace models' behavior back to the training data. This makes models' predictions more transparent by highlighting the training data that the models rely on.

Xiaosen Zheng is a fourth-year PhD Candidate in Computer Science at the SCIS, supervised by Prof. Jing Jiang. His research focuses on Post-hoc Interpretability on ML models.