Feature Alignment as a Generative Process

TL;DR

This paper introduces feature alignment, a technique for approximating reversibility in neural networks, enabling image recovery and generation without decoders, with applications in memory-efficient training.

Contribution

It proposes a novel feature alignment method that approximates reversibility, allowing image reconstruction and generation in neural networks without explicit decoders.

Findings

Can recover images from latent representations.

Generates images statistically similar to training data.

Improves image quality with GAN coupling.

Abstract

Reversibility in artificial neural networks allows us to retrieve the input given an output. We present feature alignment, a method for approximating reversibility in arbitrary neural networks. We train a network by minimizing the distance between the output of a data point and the random output with respect to a random input. We applied the technique to the MNIST, CIFAR-10, CelebA and STL-10 image datasets. We demonstrate that this method can roughly recover images from just their latent representation without the need of a decoder. By utilizing the formulation of variational autoencoders, we demonstrate that it is possible to produce new images that are statistically comparable to the training data. Furthermore, we demonstrate that the quality of the images can be improved by coupling a generator and a discriminator together. In addition, we show how this method, with a few minor modifications, can be used to train networks locally, which has the potential to save computational memory resources.