Generative adversarial interpolative autoencoding: adversarial training on latent space interpolations encourages convex latent distributions

Tim Sainburg, Marvin Thielk, Brad Thielman, Benjamin Migliori, Timothy Gentner

Abstract: We present a neural network architecture based upon the Autoencoder (AE) and Generative Adversarial Network (GAN) that promotes a convex latent distribution by training adversarially on latent space interpolations. By using an AE as both the generator and discriminator of a GAN, we pass a pixel-wise error function across the discriminator, yielding an AE which produces sharp samples that match both high- and low-level features of the original images. Samples generated from interpolations between data in latent space remain within the distribution of real data as trained by the discriminator, and therefore preserve realistic resemblances to the network inputs.

Keywords: convex, GAN, autoencoder, interpolation, stimuli generation, adversarial, latent distribution

TL;DR: We designed an autoencoder which is trained to learn a convex latent distribution by using an adversarial loss function to discriminate latent space interpolations from real data.