Abstract
Active Inference (AIF) is an emerging framework rooted in neuroscience aimed at elucidating human perception, action, and cognitive processes. Recent advances in machine learning have facilitated the non-invasive study of human cognition through the lens of this framework. This dissertation expands upon this theoretical foundation to address practical challenges in robotic control and human cognitive behavior analysis. We initiate our exploration with a review of the eye-tracking studies that underpin the motivation for AIF research. This provides a background for understanding how AIF can model cognitive processes. First, we present a minimal AIF model within a discrete state space, framed by a Partially Observable Markov Decision Process (POMDP). Through this model, we demonstrate the fundamental mechanisms of an AIF agent and investigate the impact of prior preferences. Subsequently, we shift to a continuous state-space environment, where we apply our codesigned deep AIF framework to a perceptual-motor task, namely an interception scenario involving a vehicle. Here, we compare the performance of our AIF agent with a deep-Q network, shedding light on the agent’s behavior relative to human anticipatory control in such tasks. We introduce a novel approach to local prior preference learning using contrastive learning techniques. This module improves the agent’s ability to adapt its preferences based on context. A novel strategy for robust and efficient learning through a multi-stage sampling method based on free energy evaluation is also proposed. We discuss both the positive and negative results encountered, providing a balanced view of our research efforts. The dissertation concludes with insights into the ongoing work of extending the neural AIF framework to continual learning scenarios. We present initial findings on self-organized memory modules designed to enable agents to adapt to new tasks in dynamic environments. Future research directions are provided from our perspective. Through this comprehensive study, our aim is to contribute to the understanding and application of Active Inference, enhancing its utility in both artificial intelligence and cognitive science research. Our findings not only advance the methodologies of AIF but also provide practical implementations that can be further explored in real-world applications.
Library of Congress Subject Headings
Machine learning--Technological innovations; Reinforcement learning; Cognition; Intelligent agents (Computer software)
Publication Date
12-2024
Document Type
Dissertation
Student Type
Graduate
Degree Name
Computing and Information Sciences (Ph.D.)
Department, Program, or Center
Computing and Information Sciences Ph.D, Department of
College
Golisano College of Computing and Information Sciences
Advisor
Alexander Ororbia
Advisor/Committee Member
Reynold Bailey
Advisor/Committee Member
Cecilia O. Alm
Recommended Citation
Yang, Zhizhuo, "Robust and Efficient Active Inference for Perception, Action, and Learning" (2024). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/12024
Campus
RIT – Main Campus
Plan Codes
COMPIS-PHD
