Integrating expert modules by local receptive neural network

摘要

In this paper, the task of integrating conflicting experts' opinions is achieved through a training of local receptive gating network using backpropagation. The front layers of the gating network consist of local-receptive fields which form feature maps of the input that enable modulations to experts' output. The resulting network achieves accurate modelling of the solution mapping through the efficient combination of existing experts. Experimental results on a histogram thresholding problem show the superior performance of the modular network over classical algorithms.