PIE: Pseudo-Invertible Encoder
This addresses the need for invertible compression in machine learning, offering a novel method for high-dimensional data, but it appears incremental as it builds on existing autoencoder frameworks.
The paper tackles the problem of information compression from high-dimensional data by emphasizing invertible compression, introducing Pseudo Invertible Encoders (PIE) as a new class of likelihood-based autoencoders with pseudo bijective architecture. It shows that a Gaussian Pseudo Invertible Encoder outperforms WAE and VAE in sharpness of generated images on MNIST, though no concrete numbers are provided.
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.