Extending the dual-stream approach to other Gestalt relations

Ascertain whether the dual-stream recurrent neural network that combines glimpse contents and gaze positions to learn spatial structure for numerosity can be extended to account for human judgments of Gestalt relations—including proximity, similarity, enclosure, symmetry, and continuity—when the objects are wholly novel.

Background

The model is proposed as a computational instantiation of a dorsal–ventral factorization of scene structure and content, respectively. It successfully captures enumeration behavior and neural coding motifs associated with posterior parietal cortex, supporting the view that integrating what and where enables zero-shot structural reasoning.

The authors explicitly note that general relational perception in vision encompasses a broader set of Gestalt principles. Whether the same dual-stream mechanism can generalize beyond numerosity to these relations, particularly for novel objects and configurations, remains an open question.