Generative Modeling

Deep generative classifier uses Short Run Markov Chain Monte Carlo inference, Langevin dynamics, and backpropagation through time to achieve similar classification accuracy as an analogous discriminative classifier, i.e., a convolutional neural network, while it has the advantages that it can generate data, it can learn unsupervised with additional unlabeled data, and it exhibits robustness to adversarial attacks due to the stochasticity of the Langevin equation and the top-down architecture of the underlying generator network.

Short-run MCMC residual network outperforms a variational autoencoder in regard to image reconstruction error and image synthesis quality, while not requiring the design of a separate inference network

T-snake deformable models, which can segment complex-shaped structures, improve generated masks passed to GAN for image inpainting

Formulation of evidence lower bound (ELBO) for variational autoencoder and application to synthesizing binary images