Abstract:Wasserstein distances greatly influenced and coined various types of generative neural network models. Wasserstein autoencoders are particularly notable for their mathematical simplicity and straight-forward implementation. However, their adaptation to the conditional case displays theoretical difficulties. As a remedy, we propose the use of two paired autoencoders. Under the assumption of an optimal autoencoder pair, we leverage the pairwise independence condition of our prescribed Gaussian latent distribution to overcome this theoretical hurdle. We conduct several experiments to showcase the practical applicability of the resulting paired Wasserstein autoencoders. Here, we consider imaging tasks and enable conditional sampling for denoising, inpainting, and unsupervised image translation. Moreover, we connect our image translation model to the Monge map behind Wasserstein-2 distances.
What problem does this paper attempt to address?
The main problems that this paper attempts to solve are the theoretical and practical application challenges of existing Wasserstein Autoencoders (WAE) in conditional generative modeling. Specifically:
1. **Theoretical Difficulties**: Although the standard Wasserstein Autoencoder performs well on unconditional distributions, it has theoretical difficulties when adapting to conditional generation. In particular, the bounds of the unconditional Wasserstein distance cannot be directly extended to conditional distributions, and the dependence of the model on conditional or observational data cannot be ensured [3].
2. **Implementation of Conditional Generation**: In order to enable the Wasserstein Autoencoder to perform conditional sampling, that is, to generate samples under specific conditions based on certain observational data, it is necessary to overcome the above - mentioned theoretical difficulties and provide an effective implementation method.
To solve these problems, the authors propose Paired Wasserstein Autoencoders. By using two paired autoencoders, assuming the optimal autoencoder pair, and utilizing the pairwise independence condition of the specified Gaussian latent distribution, these theoretical obstacles can be overcome and conditional sampling can be achieved.
### Specific Problems and Solutions
- **Theoretical Basis**: Based on the Optimal Transport Theory, especially the Wasserstein distance, the Paired Wasserstein Autoencoder framework is constructed.
- **Model Design**: Two paired autoencoders are introduced, which respectively map images of two different distributions to a partially shared latent distribution, which is described by a standard Gaussian distribution. Through the pairwise independence of latent variables, the conditional sampling \((X_1 | X_2 = x_2)\) can be approximated.
- **Experimental Verification**: Through a series of experiments, the feasibility of Paired Wasserstein Autoencoders in practical applications is demonstrated, including:
- Image Denoising
- Image Inpainting
- Unsupervised Image Translation
These experiments not only verify the effectiveness of the model but also show its advantages in dealing with inverse problems, such as explaining the principles behind the image translation model through Monge mapping.
### Conclusion
The paper successfully extends the Wasserstein Autoencoder to conditional generation scenarios, solves theoretical problems, and verifies its practicality in various image processing tasks through experiments. However, it also points out the limitations of the traditional autoencoder structure in the reconstruction process, and suggests that future research directions may consider hierarchical or over - complete autoencoders.
### Related Formulas
The key formulas involved in the paper are as follows:
- **Wasserstein - p Distance**:
\[
W_p^p(X_1, X_2) := \inf_{\pi \in \Pi(\mu_{X_1}, \mu_{X_2})} \mathbb{E}_{(X_1, X_2) \sim \pi} \| X_1 - X_2 \|_p^p
\]
- **Derivation of Conditional Distribution**:
\[
(X_1 | X_2 = x_2) = D_1(Z_1, z_2) \quad \text{where} \quad E_2(x_2) = (z_2, z_3), \quad Z_1 \sim \mathcal{N}(0, I)
\]
These formulas ensure the mathematical rigor and readability of the model.