TL;DR
This paper bridges the gap between Generative Adversarial Networks (GANs) and Variational Inference (VI) by showing how to use GAN-like techniques to perform approximate inference in latent variable models, even when distributions are "implicit" (hard to evaluate density, but easy to sample from). They introduce a unifying framework and practical algorithms for this task, enabling more flexible and powerful inference.
What Problem Does This Paper Address?
Variational Inference (VI) is a powerful technique for approximating intractable posterior distributions in latent variable models. Traditionally, VI relies on models where probability densities can be easily computed. However, many modern generative models, like GANs, use "implicit distributions" – models from which we can easily draw samples and compute gradients with respect to model parameters, but whose probability density function is intractable. This makes it difficult to apply standard VI techniques. While GANs have revolutionized generative modeling, their application to inference has often lacked a clear theoretical connection to VI. This paper aims to unify existing GAN-like inference methods and provide a principled framework for developing new ones that leverage implicit distributions within VI.
Methods
The paper proposes a framework for VI with implicit distributions by reformulating the Evidence Lower Bound (ELBO) in terms of density ratios. They introduce two main formulations:
- Prior-Contrastive (PC): This formulation contrasts the approximate posterior distribution ($q_\psi$) with the prior distribution ($p_\theta$). This is analogous to how Variational Autoencoders (VAEs) work.
- Joint-Contrastive (JC): This formulation contrasts the joint distribution of observed data and latent variables ($p_\theta(x, z)$) with the joint distribution implied by the approximate posterior and prior ($q_\psi(z|x)p_\theta(z)$). This is more general and can handle fully implicit models.
For estimating the necessary density ratios or their gradients when distributions are implicit, they present two main algorithmic approaches:
- Adversarial Methods (PC-Adv, JC-Adv): These methods use a discriminator (similar to GANs) trained to distinguish between samples from different distributions (e.g., approximate posterior vs. prior for PC-Adv, or joint distributions for JC-Adv). This allows for estimating density ratios or their gradients.
- Denoising Methods (PC-Den, JC-Den): These methods leverage the insight that a Bayes-optimal denoiser learns the score function (gradient of the log density) of a distribution. By training a denoising autoencoder, they can estimate these gradients, which are then used to update the variational parameters.
The paper also discusses how these methods can be combined and highlights the challenges of learning model parameters ($\theta$) when the likelihood is implicit.
Datasets
Not discussed in the paper. The paper focuses on theoretical frameworks and algorithmic development, using toy examples for illustration rather than standard datasets.
Key Insights and Ideas
- GANs as Inference Tools: The core insight is that the powerful sampling and gradient estimation capabilities of GANs can be directly repurposed for approximate inference in latent variable models, not just for generative tasks.
- Unifying Framework: The paper provides a clear theoretical lens (prior-contrastive vs. joint-contrastive ELBO formulations) to understand and connect various existing GAN-like inference algorithms.
- Implicit Distributions are Powerful for VI: Just as implicit distributions excel in generative modeling, they can offer more expressive and potentially sharper approximations for posteriors in VI, especially when explicit densities are intractable or too restrictive.
- Denoising as an Alternative to Adversarial: Denoising autoencoders offer a complementary approach to adversarial methods for estimating score functions, which are crucial for VI with implicit models. This can be more flexible in choosing which distributions are explicit vs. implicit.
- Learning vs. Inference Distinction: The paper clearly distinguishes between variational inference (finding approximate posterior parameters $\psi$) and variational learning (finding model parameters $\theta$). While their methods excel at inference, learning with implicit likelihoods remains a significant challenge.
Key Results
The paper doesn't present quantitative results on large-scale datasets but rather focuses on theoretical contributions and algorithmic development.
- Algorithmic Frameworks: They propose and detail three core algorithms: PC-Adv, JC-Adv, and PC-Den, along with their joint-contrastive denoising counterparts (JC-Den).
- Theoretical Connections: They establish clear connections between existing methods like Adversarial Variational Bayes (AVB), ALI, and BiGAN, and their proposed framework.
- Experimental Validation (Toy Model): Figure 1 demonstrates the effectiveness of PC-Adv, JC-Adv, and PC-Den in approximating the posterior of a "continuous sprinkler" toy model. The results show that all methods can produce reasonable approximations, with the quality primarily influenced by the architecture of the approximate posterior network ($q_\psi$).
Limitations and Future Work
- Learning Implicit Models: A significant limitation is that most of the presented algorithms (except for those leveraging Jensen-Shannon divergence, which are not strictly VI) do not allow for fully variational learning of model parameters ($\theta$) when the likelihood ($p_\theta(x|z)$) is implicit.
- Reverse Mode Differentiation (RMD): The authors suggest RMD as a promising avenue to enable fully variational learning for implicit models, by algorithmically differentiating the SGD process itself. They mention JC-Adv-RMD as a theoretical algorithm for this.
- Denoising Precision: While denoising methods are flexible, their gradient estimates can be imprecise, potentially leading to less sharp approximations compared to adversarial methods. Future work could explore score matching techniques to improve denoising-based methods.
- Prior-Contrastive vs. Noise-Contrastive: The paper poses a question about whether the prior-contrastive approach is too similar to noise-contrastive estimation, hinting at potential areas for further theoretical exploration.
- Experimental Validation: While a toy example is shown, more extensive empirical validation on challenging real-world problems would be beneficial to fully assess the practical advantages of these methods.
Summary updated