Genetic selection of biologically inspired receptive fields for computational vision

The paper presents the genetic selection of biologically inspired receptive fields classifiers to improve pattern recognition in neural networks. A genetic algorithm is employed to select the x and y dimensions of the receptive fields in a two plane per layer configuration with two hidden layers. Networks were ranked based on the fitness criterion: best generalization performance on handwritten digits. Results show a strong correlation between the neural network performance and the receptive field x and y dimensions. The best receptive field configuration results outperformed those of the perceptron based models. Best receptive field configurations consist of a small aspect ratio in x and y direction in each plane of the two hidden layers