Computational Models of Human-Like Skill and Concept Formation

Presented on April 13, 2023 at 12:30 p.m. in the Technology Square Research Building Ballroom. ; Christopher J. is an assistant professor in the School of Interactive Computing at Georgia Institute of Technology where I run the Teachable AI Lab (TAIL). My work aims to improve our understanding of how people teach and learn and to build AI systems that can teach and learn like people do. ; Runtime: 63:50 minutes ; The AI community has made significant strides in developing artificial systems with human-level proficiency across various tasks. However, the learning processes in most systems differ vastly from human learning, often being substantially less efficient and flexible. For instance, training large language models demands massive amounts of data and power, and updating them with new information remains challenging. In contrast, humans employ highly efficient incremental learning processes to continually update their knowledge, enabling them to acquire new knowledge with minimal examples and without overwriting prior learning. In this talk, I will discuss some of the key learning capabilities humans exhibit and present three research vignettes from my lab that explore the development of computational systems with these capabilities. The first two vignettes explore computational models of skill learning from worked examples, correctness feedback, and verbal instruction. The third vignette investigates computational models of concept formation from natural language corpora. In conclusion, I will discuss future research directions and a broader vision for how cognitive science and cognitive systems research can lead to new AI advancements.