Image data compression and generalization capabilities of backpropagation and recirculation networks

A comparison is made of the image data compression and generalization capabilities of both the backpropagation and recirculation networks. The convergence speed of the network is also examined. Simulation results show that the recirculation network has a better performance compared to the backpropagation network when used for image data compression application. The internal representation of the neural network by the concept of basis images of the weight matrix, which is helpful toward a better understanding of the principle of data compression and generalization capabilities of the neural networks, is also discussed.<<ETX>>