Biologically inspired receptive field geometry and orientation for pattern recognition enhancement

Proposes a new method to incorporate biologically inspired receptive fields in a feedforward neural network to enhance pattern recognition performance. Based on a genetic algorithm the method determines the receptive field geometry, orientation, bias, and the number of planes per layer that maximize the pattern recognition performance of the network. The method is tested in the handwritten digit problem. The basic architecture of the neural network is inspired on the Neocognitron model. Resulting network architectures were ranked based on the fitness criterion: best generalization performance on a testing set. Results show a strong correlation between the neural network performance and the receptive field geometry and orientation. Results were compared with those of a fully connected perceptron neural network that does not incorporate receptive fields. Results reached by several networks with receptive field configuration determined by the genetic algorithm outperformed those of the perceptron model