Perceptual Generative Autoencoders

By Zijun Zhang, Ruixiang Zhang, Zongpeng Li, Yoshua Bengio, Liam Paull. 2019

32 citations recorded by the source corpus.

Modern generative models are usually designed to match target distributions directly in the data space, where the intrinsic dimension of data can be much lower than the ambient dimension. We argue that this discrepancy may contribute to the difficulties in training generative models. We therefore propose to map both the generated and target distributions to a latent space using the encoder of a standard autoencoder, and train the generator (or decoder) to match the target distribution in the latent space. Specifically, we enforce the consistency in both the data space and the latent space with theoretically justified data and latent reconstruction losses. The resulting generative model, which we call a perceptual generative autoencoder (PGA), is then trained with a maximum likelihood or variational autoencoder (VAE) objective. With maximum likelihood, PGAs generalize the idea of reversible generative models to unrestricted neural network architectures and arbitrary number of latent dimensions. When combined with VAEs, PGAs substantially improve over the baseline VAEs in terms of sample quality. Compared to other autoencoder-based generative models using simple priors, PGAs achieve state-of-the-art FID scores on CIFAR-10 and CelebA.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Generative Models, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Ting-Li Chen
2. Geoffrey E. Hinton
3. Ruixiang Zhang
4. Quoc V. Le
5. Yoshua Bengio
6. Pieter Abbeel
7. Trevor Darrell
8. Ruslan Salakhutdinov

In Computer Vision and Pattern Recognition

1. Ting-Li Chen
2. Luc Van Gool
3. Geoffrey E. Hinton
4. Dacheng Tao
5. Trevor Darrell
6. Andrew Zisserman
7. Ming-Hsuan Yang
8. Xiaogang Wang

In Generative Models

1. Ting-Li Chen
2. Geoffrey E. Hinton
3. Ruixiang Zhang
4. Yoshua Bengio
5. Ruslan Salakhutdinov
6. Aaron C. Courville
7. Li Deng
8. Kyunghyun Cho

In Signal Processing

1. Ting-Li Chen
2. Geoffrey E. Hinton
3. Ruixiang Zhang
4. Li Deng
5. Dong Yu
6. Yoshua Bengio
7. Ruslan Salakhutdinov
8. Jinyu Li

Citations in the corpus

• Dual Contradistinctive Generative Autoencoder (2020, neutral): “PGA [68] adds a constraint to the latent variables.”
• Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning (2022, background): “Several works [31, 63, 64] also explore normalizing flows on discrete data under the Variational Autoencoders [29] framework by enriching the prior.”
• NCP-VAE: Variational Autoencoders with Noise Contrastive Priors (2020, unknown): “ance on CelebA-64 Model FID# NCP-VAE (ours) 5.25 NVAE (Vahdat & Kautz, 2020) 13.48 RAE (Ghosh et al., 2020) 40.95 2s-VAE (Dai & Wipf, 2018) 44.4 WAE (Tolstikhin et al., 2018) 35 Perceptial AE(Zhang et al., 2020) 13.8 Latent EBM (Pang et al., 2020) 37.87 COCO-GAN (Lin et al., 2019) 4.0 QA-GAN”