Neuronal Correlates of Semantic Event Classes during Presentation of Complex Naturalistic Stimuli: Anatomical Patterns, Context-Sensitivity, and Potential Impact on shared Human-Robot Ontologies

Abstract

The present study forms part of a research project that aims to develop cognition-enabled robotic agents with environmental interaction capabilities close to human proficiency. This approach is based on human-derived neuronal data in combination with a shared ontology to enable robots to learn from human experiences. To gain further insight into the relation between human neuronal activity patterns and ontological classes, we introduced General Linear Model (GLM) analyses on fMRI data of participants who were presented with complex naturalistic video stimuli comparable to the robot tasks. We modeled four event classes (pick, place, fetch and deliver) attached to different environmental and object-related context and employed a Representational Similarity Analysis (RSA) on associated brain activity patterns as a starting point for an automatic hierarchical clustering. Based on the default values for the Hemodynamic Response Function (HRF), the activity patterns were reliably grouped according to their parent classes of object interaction and navigation. Although fetch and deliver events were also distinguished by neuronal patterns, pick and place events demonstrated higher ambiguity with respect to neuronal activation patterns. Introducing a shorter HRF time-to-peak leads to a more reliable grouping of all four semantic classes, despite contextual factors. These data might give novel insights into the neuronal representation of complex stimuli and may enable further research in ontology validation in cognition-enabled robotics.