Zining Zhu (朱子宁)

zining@cs.toronto.edu

I am a PhD candidate at University of Toronto. I am interested in understanding the mechanisms and abilities of neural NLP systems. In the long term, I try to build reliable, interpretable an trustworthy NLP systems. These include developing methods to communicate complex decisions to humans, and empowering real-world applications with the advance of the models.
Following are some topics I am working on:
  • Interpretability: develop methods that reveal the mechanisms of complex neural network systems. I work on solidifying probing methods [C8, C4] and exploring the utility of probing analyses [C9, C7, C5].
  • Trustworthiness: control the models along multiple linguistic levels. The goal is to allow the models to make the correct decisions with the available knowledge in diverse situations.
  • Datasets: understand the datasets that ML and NLP systems are trained on [I4].
  • Applications: using ML and NLP methods to address some real-world problems in the healthcare and society contexts [C6, C3].

The publication page contains a complete list of my publications. I happily collaborate with many researchers on the above topics. Please click here if you want to chat about collaboration or related ideas.

Education

University of Toronto 2019 - present
Ph.D student in Computer Science
Advisor: Frank Rudzicz
University of Toronto 2014 - 2019
Bachelor in Engineering Science, Robotics option.

Work / Research

Amazon, Applied Scientist Intern, 2022
Search - Query Understanding. Advisors: Haoming Jiang, Jingfeng Yang, Sreyashi Nag, and Chao Zhang
Tencent Jarvis Lab, Machine Learning Engineering Intern, 2019
Neural language models and pre-training techniques. Advisor: Ruihui Zhao.
Winterlight Labs, Research Software Engineer, 2017 - 2018
Automatic detection of dementia from narrative speeches. Advisor: Jekaterina Novikova.
TripAdvisor, Software Engineering Intern, 2017
Android applications and Java API.
Dynamic Systems Lab at UTIAS, Research Assistant, 2016
Enhancing drone controllers using deep neural networks. Advisor: Angela Schoellig.

Misc