Name: Huadong Xiong

School of Psychology – Ph.D. Dissertation Defense Meeting

Date: Tuesday, May 12th, 2026

Time: 12:00 - 1:30PM

Location: in-person - JS Coon 150

and virtually (https://gatech.zoom.us/my/huadong.xiong?pwd=u8JPjJbQwVLXhhf9RPItRJS7dcFEw4.1)

Dissertation Committee Chair/Advisor:

Robert Wilson, Ph.D. (Georgia Tech)

Dissertation Committee Members:

Dobromir Rahnev, Ph.D. (Georgia Tech)

Apurva Ratan Murty, Ph.D. (Georgia Tech)

Audrey Sederberg, Ph.D. (Georgia Tech)

Marcelo Mattar, Ph.D. (New York University)

Title: An Online Learning Perspective on Cognition

Abstract: How should an intelligent system adapt when the world changes faster than its underlying architecture can be redesigned? Classical symbolic theories have approached this question by describing cognition as computation over structured variables, rules, and formal operations, whereas connectionist theories have approached it through learning rules such as Hebbian plasticity, error-driven updating, and distributed representational change. In my Ph.D. work, I develop an online learning perspective that connects these traditions by treating cognition as the continual adjustment of internal strategies in response to feedback, uncertainty, and resource constraints. This project has three main parts. First, I mathematically show that several influential cognitive models can be re-expressed through a duality with online optimization: although these models differ in surface form and theoretical vocabulary, they can often be understood as optimizing implicit loss functions over internal representations or policies. This provides a way to relate symbolic and connectionist models without reducing one tradition to the other. Second, I apply this framework to human decision making, showing that both value-based and perceptual decisions can be interpreted as adaptive online optimization processes. Alongside this modeling work, I develop faster and more reliable software tools for parameter estimation, enabling more efficient tests of cognitive models at scale. Third, I extend the same perspective to large language models, using online optimization as a lens for studying in-context learning, metacognition, and neural constraints on adaptation. Inspired by questions from human cognition, I further investigate representation interference, Bayesian hypothesis generation, Bayesian evidence integration, and feature binding in LLMs. Together, my work suggests that online learning offers a useful computational perspective for understanding adaptive intelligence across cognitive models, human behavior, and artificial systems.