Webinar

Abstracts:

Sparse Inversion of Stacked Autoencoder Classification Machines

In many applications of supervised learning it is from interest to generate a number of synthetic high dimensional model input from a high dimensional as well output. Especially in image processing augmentation algorithms like geometric transformations (translation, rotation and scaling fractional linear Möbius transformation), color space augmentations, kernel filters, mixing images, random erasing, feature space augmentation, adversarial training, generative adversarial networks, neural style transfer, and meta-learning have been considered as solutions to the problem of limited data space. Furthermore the classical inverse problem is still of high interest. E. g. the optimization problem: How to achieve a particular desired outcome from a complex high dimensional production process input data. We describe a generative, sparse, stacked autoencoder based model and show its performance on the MNIST data set. The approach can be considered as the alternative to the Generative Adversarial Networks (GAN), conditional GAN’s class approaches, to the variational type of autoencoders, or to the Restricted Boltzmann Machines.