Abstract
INTRODUCTION: Humans can perceive social interactions from natural as well as from highly impoverished stimuli, as shown by the classical experiments by Heider and Simmel. The neural circuits underlying this visual function remain completely unknown, and it has been suggested that the recognition of such stimuli is based on sophisticated probabilistic inference. We present a simple neural model, which is consistent with basic facts known about neurons in the visual pathway, that recognizes social interaction from naturalistic as well as from abstract stimuli. In addition, we present an algorithm for the generation of classes of naturalistic and abstract interaction simuli with full parametric control. Such stimuli are critical for electrophysiological experiments that clarify underlying mechanisms. METHODS: The model consists of a hierarchical shape-recognition pathway with incomplete position invariance that is modelled using a deep neural network (VGG16). The top-levels of the architecture compute the relative motion, speed and acceleration of moving agents in the scene, and classifies the interactions. Relative position is computed using a gain-field mechanism. The stimulus synthesis algorithm is derived from dynamic models of human navigation which are combined with methods for computer animation of quadrupedal animals. RESULTS: Classifying abstract interaction stimuli consisting of moving geometrical figures generated by the algorithm, we found reliable classification of 12 interaction categories. The model reproduces this classification and recognizes also interactions from real movies showing interacting animals. The model proposes a variety of neuron classes that are presently being searched for in onoing electrophysiological experiments. CONCLUSION: Simple neural circuits combined with learning are sufficient to account for simple forms of social interaction perception in real and artificial stimuli.