A Neural Architecture for Representing and Reasoning about Spatial Relationships

Eric Weiss, Brian Cheung, Bruno Olshausen

Abstract: We explore a new architecture for representing spatial information in neural networks. The method binds object information to position via element-wise multiplication of complex-valued vectors. This approach extends Holographic Reduced Representations by providing additional tools for processing and manipulating spatial information. In many cases these computations can be performed very efficiently through application of the convolution theorem. Experiments demonstrate excellent performance on a visuo-spatial reasoning task as well as on a 2D maze navigation task.

Conflicts: berkeley.edu