Abstract
INTRODUCTION: Humans are highly skilled at interpreting intent or social behavior from strongly impoverished stimuli (Heider & Simmel, 1944). It has been hypothesized that such functions might be based on high-level cognitive processes, such as probabilistic reasoning. We demonstrate that several classical observations on animacy and interaction perception can be accounted for by simple and physiologically plausible neural mechanisms, using an appropriately extended hierarchical (deep) model of the visual pathway. METHODS: Building on classical biologically-inspired models for object and action perception (Riesenhuber & Poggio, 1999; Giese & Poggio, 2003), we propose a learning-based hierarchical neural network model that analyzes shape and motion features from video sequences. The model has largely a simple feed-forward architecture and comprises two processing streams for form and object motion in a retinal frame of reference. We try to account with this model simultaneously for a number of experimental observations on the perception of animacy and social interaction. RESULTS: Based on input video sequences, the model reproduces results of Tremoulet and Feldman (2000) on the dependence of perceived animacy on changes in speed and direction of moving objects, on its dependence on the alignment of motion and body axis, and the influence of contact with static barriers along the motion path (Hernik et al. 2013). In addition it accounts for results on the detection of chasing behavior (Scholl & McCarthy, 2012) and of fighting (Heider & Simmel, 1944). CONCLUSION: Since the model accounts simultaneously for a variety of effects related to animacy and interaction perception using physiologically plausible mechanisms, without requiring complex computational inference and optimization processes, it might serve as starting point for the search of neurons that are forming the core circuit of the perceptual processing of animacy and interaction.
Meeting abstract presented at VSS 2018