Sensor fusion by neural networks using spatially represented information

A neural network model based on a lateral-inhibition-type feedback layer is analyzed with regard to its capabilities to fuse signals from two different sensors reporting the same event ("multisensory convergence"). The model consists of two processing stages. The input stage holds spatial representations of the sensor signals and transmits them to the second stage where they are fused. If the input signals differ, the model exhibits two different processing modes: with small differences it produces a weighted average of the input signals, whereas with large differences it enters a decision mode where one of the two signals is suppressed. The dynamics of the network can be described by a series of two first-order low-pass filters, whose bandwidth depends nonlinearly on the level of concordance of the input signals. The network reduces sensor noise by means of both its averaging and filtering properties. Hence noise suppression, too, depends on the level of concordance of the inputs. When the network's neurons have internal noise, sensor noise suppression is reduced but still effective as long as the input signals do not differ strongly. The possibility of extending the scheme to three and more inputs is discussed.