Data fusion in neural networks via computational evolution

Pattern recognition systems commonly employ a single representation of the sensor data. For hard classification problems it is unlikely that a single representation will be able to capture all the relevant information in the sensor field. For a given input, the goal is to fuse the information contained in multiple representations to compute the associated pattern class. For each representation, the learning vector quantization network is first used to establish a transformation to an associated feature space. A recurrent network is then used to fuse the information generated by each of the representations. The weights for the recurrent network are learned using an evolutionary strategy. This network is multi-stable and its equilibrium states are associated with different pattern classes. For a specified input, the system relaxes to an equilibrium state associated with an underlying pattern class. The class decision boundaries generated by the recurrent neural network are compared to the boundaries generated by nearest neighbor recall