Gestalt Perception of Biological Motion: A Generative Artificial Neural Network Model

Gestalt perception is one of the hallmarks of our perceptual-cognitive apparatus. The perception of a Gestalt relies on properly binding and mapping visual features onto a matching template, possibly adjusting the perspective meanwhile. In this paper we propose a generative artificial neural network model, which implements such Gestalt perception mechanisms. The model is combining principles of predictive coding and active inference with suitable inductive learning and processing biases. The design of the model enables the inference of types of biological motion upon observation, tackling the challenges of (i) feature binding into Gestalten, (ii) perspective taking, and (iii) behavior interpretation. Ablation studies reveal three critical architectural components: First, spatial input encodings improve Gestalt perception when being separated into relative positional, directional, and motion magnitude pathways. Second, the implicit parallel testing of alternative Gestalt and perspective taking hypotheses becomes possible when inputs are translated into neural population encodings. Third, a temporal predictive processing module allows the retrospective inference of the type of unfolding behavior given autoencoder-compressed stimuli. We believe that similar modularizations, inductive biases, and inference mechanisms should also be implemented in other architectures where the temporal selective integration of possibly partially scrambled information sources is useful.