Object Localisation Using Laterally Connected “What” and “Where” Associator Networks

Abstract

We describe an associator neural network to localise a recognised object within the visual field. The idea extends the use of lateral connections within a single cortical area to their use between different areas. Previously, intra-area lateral connections have been implemented within V1 to endow the simple cells with biologically realistic orientation tuning curves as well as to generate complex cell properties. In this paper we extend the lateral connections to also span an area laterally connected to the simulated V1. Their training was done by the following procedure: every image on the input contained an artificially generated orange fruit at a particular location. This location was reflected – in a supervised manner – as a Gaussian on the area laterally connected to V1. Thus, the lateral weights are trained to associate the V1 representation of the image to the location of the orange. After training, we present an image with an orange of which we do not know its location. By the means of pattern completion a Gaussian hill of activation emerges on the correct location of the laterally connected area. Tests display a good performance with real oranges under diverse lighting and backgrounds. A further extension to include multi-modal input is discussed.