Human goal-directed behavior unfolds in complex, dynamic environments in which perception, action, and cognition continuously interact. Understanding how the brain supports such naturalistic behavior requires experimental designs that capture this com...
Human goal-directed behavior unfolds in complex, dynamic environments in which perception, action, and cognition continuously interact. Understanding how the brain supports such naturalistic behavior requires experimental designs that capture this complexity while remaining analytically tractable. In this thesis, I introduce an interactive, Minecraft-based shepherding task that integrates long-term planning, proactive environmental restructuring, and complex multi-agent interaction. I develop a large set of feature spaces describing perception, spatial structure, action, and cognitive variables, enabling high-dimensional modeling of behavior and neural responses.
Among these features, eye movements are one of the most essential for understanding the neural basis of goal-directed behaviors in naturalistic cognition. However, achieving high-precision eye tracking in fMRI remains a challenge. Even slight head shifts from the initial calibration position can introduce drift in eye tracking data, leading to substantial gaze inaccuracies. To address this, I develop Motion-Corrected Eye Tracking (MoCET), which corrects drift using head motion parameters derived from fMRI preprocessing. MoCET requires no additional hardware and can be applied retrospectively to existing datasets and achieves greater spatial and temporal precision for modeling of behavior and neural responses in naturalistic cognition.
Using these tools, I apply voxelwise encoding modeling to characterize how different cortical regions encode the computations required for each stage of the task. Model connectivity analysis identifies three large-scale functional networks optimized for action planning, spatial and environmental reasoning, and multi-agent interaction, respectively.
To investigate why certain features are preferentially represented in cortical activity, I develop a computational model that estimates the behavioral relevance of each feature by training recurrent neural networks to predict participants’ future actions. Comparing these behavioral relevance scores with variance explained of cortical activity reveals brain–model correspondence in task-relevant functional networks. Cortical networks preferentially encode the features most useful for generating future actions, and this prioritization shifts with the current behavioral goal.
Together, these findings provide a comprehensive account of how the human brain flexibly reorganizes to support naturalistic goal-directed behavior. By integrating an interactive virtual environment, high-precision eye tracking, Voxelwise Encoding Modeling, and computational analysis of behavior, this thesis establishes a generalizable framework for studying the neural basis of goal-directed behaviors in naturalistic cognition.