[4S1-OS-30a] OS-30

[4S1-OS-30a-02] Development of Hierarchical Spatial Cognition through Visuo-Motor Experience by Shared-Module RNN

Wataru Noguchi¹, 〇Hiroyuki Iizuka¹ (1. Hokkaido University)

Keywords:Spatial Cognition, Hierarchical Cognition, Deep Learning

We propose a neural network model in which shared RNN modules are hierarchically connected, which can acquire hierarchical spatial cognition through visuomotor experience. The proposed model takes the visual and motor sensations of mobile agents as input and modulates its neural connections through learning to predict future visual sensations. The proposed model was trained on the visuomotor sequences of a mobile robot agent in a simple simulation environment with multiple rooms connected by corridors. As a result, the representations of the local spatial position of the robot in the room and the global scale spatial arrangement of the rooms were self-organized in the shared RNN of the lower and higher layers of the model, respectively. The result indicates that the shared module structure encourages the development of hierarchical and structuralized spatial representations.