Zining Zhu (朱子宁)

• Interpretation: Develop methods to probe deep neural networks -- unveiling what happens among the structures and the neurons, and query the sources of their strong capabilities [W4, C5]. Then, we can incorporate the learned insights into developing better models [C9]. We should also make these probing methods solid -- reliable and valid [C8, C4].
• Explanation: AIs can be more helpful if they can explain to humans, and they have promising potential. Along this topic, we solidify the problem setting of "machine explanation": How to evaluate the explanations? How can we make explanations faithful and useful? We try to leverage the reasoning abilities and the stored commonsense of large language models and facilitate "machine teaching", to diverse audiences.
• Safety, control, alignment: To harness the powerful abilities of AIs, we need better approaches to control their behavior. AIs make errors, hallucinate, and demonstrate biases. We try to quantify these improvement area, and develop methods to address them.
• Applications: Use AI tools to address societal challenges, for example, healthcare and fairness, ideally following a human-AI collaboration paradigm. [C6, C3].

The publication page contains a complete list of my publications.