PIE: Pseudo-Invertible Encoder

TL;DR

This paper introduces Pseudo Invertible Encoders, a new class of likelihood-based autoencoders with pseudo bijective architecture, emphasizing invertible data compression and demonstrating superior image sharpness on MNIST.

Contribution

The paper proposes a novel pseudo bijective autoencoder architecture for invertible data compression, with theoretical foundations and empirical evaluation.

Findings

Outperforms WAE and VAE in image sharpness on MNIST

Provides theoretical explanation of pseudo bijective architecture

Highlights importance of invertible compression methods

Abstract

We consider the problem of information compression from high dimensional data. Where many studies consider the problem of compression by non-invertible transformations, we emphasize the importance of invertible compression. We introduce new class of likelihood-based autoencoders with pseudo bijective architecture, which we call Pseudo Invertible Encoders. We provide the theoretical explanation of their principles. We evaluate Gaussian Pseudo Invertible Encoder on MNIST, where our model outperforms WAE and VAE in sharpness of the generated images.