ABSTRACT
A variant of the encoder architecture, where units at the input and output layers represent nodes on a graph, is applied to the task of mapping locations to sets of neighboring locations. The degree to which the resulting internal (i.e. hidden unit) representations reflect global properties of the environment depends upon several parameters of the learning procedure. Architectural bottlenecks, noise, and incremental learning of landmarks are shown to be important factors in maintaining topographic relationships at a global scale.
