Zory Zhang

Email: [email protected]

Website: zoryzhang.notion.site

Nationality: China

Pronoun: he/him

Education

University of Illinois, Urbana-Champaign (UIUC) Aug 2023 - May 2026 (Expected)

• Pursuing Bachelor's degree of Science in Mathematics and Computer Science, CGA=3.96/4.00

Hong Kong University of Science and Technology (HKUST) Sep 2021 - July 2023

• Former undergraduate majoring in Computer Science and Mathematics

Research Interest

I’m interested in how we can replicate on machines the amazing learning ability of humans by looking at how humans manage to do so (which is a complicated process with lots of practice).

Through reverse-engineer human learning, I attempt to join the modern attempt to answer the age-old question: how much innate knowledge and concept (”inductive bias”) is likely needed to give rise to a certain amount of learning ability in certain aspects, and what is the right way to use them to help with learning. I have a board interest in borrowing insights from research in computational modeling and rational analysis, language acquisition, developmental psychology, philosophy of cognitive science and mind, traditional and deep learning, and cognitive architecture, …

A byproduct of this mission is a system that can learn sample-efficiently, incrementally, and continually (never-ending learning, without catastrophic forgetting), as well as flexibly generalize during performance. I believe that reasoning is not only a performance task but also helps learning, i.e. Reason2Learn. Given prior beliefs, humans can learn so much from so little exposure to a novel domain (e.g. ARC). I also believe perception and reasoning are bidirectional.

My toolkit includes hierarchical concept formation (COBWEB), schema induction (LISA) and explanation-based learning (EBL) for human-like learning, analogical reasoning and “neuro-symbolic” architecture for human-like inference, using knowledge graph as less expressive but cheap factual (as opposite to rule-like or procedural) knowledge representation, as well as employing formal mathematics theorem proving as a playground to study cognition.

📄 Publications

Accepted, in press:

• Dylan Zhang, Curt Tigges, Zory Zhang, Stella Biderman, Maxim Raginsky, and Talia Ringer (2024). Transformer-Based Models Are Not Yet Perfect At Learning to Emulate Structural Recursion. Transactions on Machine Learning Research. Arxiv.

Experience

Research Intern

May 2024 - Aug 2024. Institute for the Study of Learning and Expertise & Georgia Institute of Technology.

• Supervised by Pat Langley and Christopher J. MacLellan. Worked as project leader.
• Work in progress: Developing a human-like learning algorithm that can learn from patterns in knowledge graphs and gain the ability to predict missing information in incomplete unseen knowledge graphs.

Research Project Leader, as a volunteer

Feb 2024 - Present. University of Illinois, Urbana-Champaign.

• Supervised by Jiaxuan You. Worked as project leader.
• Work in progress: Integrate Analogical Abduction, Connectionism, and Explanation-based Learning in a theorem prover.