A Matter of Perspective: From Interpreting Models to New Horizons
On 21 May 2026, Lim Jia Peng successfully defended his thesis, titled "Perspectives on Interpretability for Neural Text Representations."
The main thesis of this dissertation is that the interpretability of neural text representations can be enhanced by examining three fundamental elements of the learning process: inputs, models, and outputs.
- The inputs part investigates the interpretability of subword representations and explores how to construct token vocabularies that better capture textual meaning.
- The models part develops novel approaches to interpreting neural topic models and disentangling superposed concepts in large language models, drawing on classical graph optimization problems.
- The outputs part examines the relationship between automated coherence metrics and human judgments, providing insights into how the interpretability of text representations can be evaluated.
Jia Peng has a long association with Preferred.AI. He got involved in research as an undergraduate pursuing his bachelor's degree in Information Systems, before joining us as a PhD candidate in 2021.
His research has resulted in publications at leading natural language processing venues, including ACL, EMNLP, COLING, and Computational Linguistics, as well as the premier machine learning conference NeurIPS. His academic achievements have also earned him numerous accolades, including the SMU Presidential Doctoral Fellowship, Singapore Data Science Consortium Dissertation Research Fellowship, AISG Research PhD Fellowship Award, PhD Dean's List, and DSO-AISG Incentive Award.
After graduation, he has embarked on the next chapter of his research career as a Research Scientist in Machine Learning at Meta in the United States. We congratulate him on this exciting milestone and wish him continued success as he brings his research expertise to new challenges and opportunities.
