Generative Adversarial Networks and Variational Auto-Encoders provide impressive image generations from Gaussian white noise, which are not well understood. We show that such generations do not require to learn a discriminator or an encoder. They are computed with a scattering transform which preserve the deformation properties of image synthesis. The deep convolutional network generator is calculated as the solution of a regularized inverse problem. We show that this approach also applies to time-series and audio synthesis, thus providing an alternative to recurrent neural networks and wavenets. Numerical results will be shown on images and audio signals.
Joint work with Tomas Angles and Mathieu Andreux, Ecole Normale Superieure, Collège de France
Copyright. All Rights Reserved.