Philip Mirbach

Abstract

The information a neural network can learn about data depends on the choice of data representation. Different representations can clarify or hide different descriptive factors of the data. We try to improve the classification accuracy of images by using different representations of the images with specialized neural networks. We hope that the different information obtained from the networks will complement each other and that an ensemble of all networks will lead to improved performance. This will be investigated using the example of the Fourier representation of the images.