Particle Semi-Implicit Variational Inference

Semi-implicit variational inference (SIVI) enriches the expressiveness of variational families by utilizing a kernel and a mixing distribution to hierarchically define the variational distribution. Existing SIVI methods parameterize the mixing distribution using implicit distributions, leading to intractable variational densities. As a result, directly maximizing the evidence lower bound (ELBO) is not possible, so they resort to one of the following: optimizing bounds on the ELBO, employing costly inner-loop Markov chain Monte Carlo runs, or solving minimax objectives. In this paper, we propose a novel method for SIVI called Particle Variational Inference (PVI) which employs empirical measures to approximate the optimal mixing distributions characterized as the minimizer of a free energy functional. PVI arises naturally as a particle approximation of a Euclidean--Wasserstein gradient flow and, unlike prior works, it directly optimizes the ELBO whilst making no parametric assumption about the mixing distribution. Our empirical results demonstrate that PVI performs favourably compared to other SIVI methods across various tasks. Moreover, we provide a theoretical analysis of the behaviour of the gradient flow of a related free energy functional: establishing the existence and uniqueness of solutions as well as propagation of chaos results.

Paper

References (62)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer KYHY7/10 · confidence 3/52024-07-09

Summary

This paper introduces Particle Variational Inference (PVI), a novel method for semi-implicit variational inference (SIVI) that employs empirical measures to optimize the mixing distribution without making parametric assumptions. Unlike existing SIVI methods that face challenges with intractable variational densities and rely on costly computational techniques, PVI directly optimizes the evidence lower bound (ELBO) using a particle approximation of an Euclidean–Wasserstein gradient flow.

Strengths

1. The paper is well-organized and presents the complex concepts of the proposed method with clarity. 2. The Particle Variational Inference (PVI) algorithm, which discretizes the Euclidean–Wasserstein gradient flow to accommodate general mixing distributions, is both practical and innovative. This algorithm is not only grounded in solid theoretical foundations but also implemented effectively, showcasing its real-world applicability. 3. Empirical comparisons of PVI with other SIVI approaches across a variety of toy and real-world experiments demonstrate their good performance.

Weaknesses

I do not identify any significant weaknesses in the paper. However, there are a few points that require clarification, which I have addressed in the questions section.

Questions

1. Figure 1: The current caption makes it difficult to interpret the figure. It would be helpful to explicitly mention that lighter shades represent smaller μ\muμ values, while darker shades represent larger μ\muμ values. 2. Figure 2: The second sentence of the caption is somewhat confusing. Referring to Table 1 for additional clarification might improve the reader's understanding. 3. Minor Typo: In line 303, "mu=2" should be corrected to "$\mu=2$"

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

The authors have adequately discussed the limitations of their approach.

Reviewer JYYZ9/10 · confidence 5/52024-07-11

Summary

The authors introduce a novel algorithmic approach to fitting semi-implicit variational approximations. This method is based on discretizing a suitable gradient flow, and the paper provides a comprehensive theoretical analysis to support it. The approach, called Particle Variational Inference (PVI), directly optimizes over the space of probability distributions for the mixing distribution rather than optimizing a specific parametric form for the mixing distribution, as in previous SIVI approaches. Numerical examples demonstrate improvements compared to existing semi-implicit variational inference methods. The paper provides a solid theoretical foundation and a practical algorithm with beneficial properties over existing methods in the literature.

Strengths

This paper is a very strong submission, and I believe it is worthy of acceptance at this conference. The strengths that I would like to highlight are the following: - **High degree of novelty:** Semi-implicit variational inference (SIVI) methods have existed for some time. Existing approaches faced challenges due to the nature of their design, often leading to optimizing bounds on the ELBO or difficult algorithms. This paper's essential contribution is constructing an objective function that can be directly optimized by its gradient flow. This is a significant advancement in the SIVI literature, enabling direct optimization of the ELBO. - **Theoretical foundations:** The paper provides a solid theoretical analysis of the proposed method. This includes the study of a related gradient flow, establishing existence and uniqueness of solutions, and providing propagation of chaos results. These theoretical underpinnings give a rigorous basis for the practical algorithm. - **Practical algorithm:** The particle variational inference (PVI) algorithm is derived as a practical implementation by discretizing the gradient flow of the proposed objective function. This direct link between theory and practice is a strength of the paper, as it provides a clear path from the mathematical formulation to a computationally feasible method.

Weaknesses

The following points, particularly the limited experimental validation, prevent me from giving the submission a higher rating (8/9): - **Limited scope of numerical experiments:** The paper's major weakness lies in its experimental section, which represents the minimum set of experiments for an acceptable paper. Specifically: a. Lack of empirical demonstration of expressiveness: Despite claiming in Section 2 that PVI can learn potentially more expressive variational approximations than other SIVI methods, no experiments empirically demonstrate this advantage. b. Insufficient exploration of optimization stability: The paper misses an opportunity to address a key challenge in SIVI methods - the difficulty in getting objective functions/algorithms to converge (or meaningfully work). Experiments showing that PVI provides consistently more stable optimization than other SIVI methods or converges on complex models where other SIVI methods struggle would have significantly strengthened the paper. Additionally, comparing performance accuracy on Bayesian Neural Networks (BNNs) for different SIVI methods doesn't show meaningful advantages to this method. - **Theoretical-practical gap:** The theoretical analysis is for a modified gradient flow (with γ > 0), but the practical algorithm uses γ = 0. However, this discrepancy is mitigated by the authors' justification that this approach is relatively common in the literature, and empirical results did not show obvious discrepancies between theory and practice. - **Clarity in explaining the flexibility of PVI's variational approximation:** The novelty in PVI's approach - not optimizing a specific parametric form of the mixing distribution - was not immediately apparent to me from the explanations in Sections 2 and 3 upon first read. In particular, the last sentence in Section 2 (lines 103-105) is confusing. While it becomes more understandable after reading Section 3, the intuition behind why Q_{YuZ} might not reduce to other parameterizations when fit with PVI is not fully explained.

Questions

1. How sensitive is PVI to the choice of kernel and number of particles? 2. Can the theoretical analysis be extended to cover the γ = 0 case used in practice? 3. How does PVI scale to higher-dimensional problems or larger datasets? 4. Are there specific types of problems where PVI is expected to significantly outperform existing methods? 5. How does the choice of preconditioner affect PVI's performance?

Rating

9

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

The authors have adequately addressed limitations.

Reviewer 38kT7/10 · confidence 2/52024-07-12

Summary

The authors propose a method for SIVI called Particle Variational Inference (PVI) which employs a particle approximation of an Euclidean–Wasserstein gradient flow. PVI directly optimizes the ELBO, and it makes no parametric assumption about the mixing distribution. Their empirical results demonstrate that PVI performs favorably against other SIVI methods across various tasks. The authors provide a theoretical analysis of the behavior of the gradient flow of a related free energy functional.

Strengths

The authors provide extensive theoretical analysis to support the proposed method. The authors provide solid theoretical results along with the proposed method. The paper is well written.

Weaknesses

Experiments on a real-world application could strengthen the paper.

Questions

Line 329: what's the meaning of 'Here the ± denotes the average ...'?

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer o93p6/10 · confidence 4/52024-07-13

Summary

The paper proposes PVI as a new method to conduct variational inference using the semi-implicit distribution. The method is to construct a gradient flow to minimize a regularized ELBO, which is practically implemented as the particle propagations. The empirical studies show the accuracy over density estimation and posterior predictions.

Strengths

- The idea of using Wasserstein gradient flow to optimize the intractable ELBO of SIVI is novel. - The method has good accuracy in the provided simulations. - The techniques in the method derivation may be useful for other areas such as flows and diffusion models.

Weaknesses

A major concern is about the simulations. First, it is unclear whether the comparisons between PVI and the baselines SVI, UVI and SM are fair. The accuracy of SIVI methods can often improve with an increase in the computation such as the number of samples in SVI and the MCMC steps in UVI. For a fair comparison, all methods need to be under the same computation budget. However, the current paper does not provide details of how the SVI, UVI, and SM are implemented. It would be better to include figures that show computation/time versus the accuracies for all the methods. Current simulations are relatively too simple and are only conducted on several 2D toy examples and UCI datasets. First, none of these simulations directly show the accuracy of the posterior inference for a complete Bayesian model. Second, all the simulations are in low-dimension. Does the particle method suffer from the curse of dimension problem? How does the number of particles scale with dimensionality to maintain accuracy? It would be interesting to verify uncertainty estimation in high dimensions. Last, the paper writing is not coherent in some places. For example - The non-coercive is not defined in Proposition 2 - The paper mentions the non-coercive "is closely related to the problem of posterior collapse"; what is the relationship exactly? - How to compute the precondition matrices in Eq 13 is not discussed

Questions

Does the regularization in Eq 4 introduce bias in the posterior inference? Does the theoretical analysis in Sec. 4 provide guidance in designing the algorithm? How to check coercive and l.s.c. in Proposition 3?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

N/A

Reviewer o93p2024-08-10

Thank you for the rebuttal

The further analysis of the computation time and full Bayesian model, together with the discussion on the dimensionality, strengthen the paper and address my major concerns. Thanks the authors for providing code to facilitate results reproducing. I update my score accordingly.

Reviewer 38kT2024-08-11

Thanks for the response. The score from this reviewer is unchanged.

Reviewer JYYZ2024-08-13

Response to rebuttal from authors

Thank you for providing the additional numerical results and answering my questions. The paper is technically flawless and is highly novel. Semi-implicit VI has been around for years, and the construction of the objective plus algorithm solves what I consider an existing open problem in the area (unbiased gradient estimator of the ELBO directly). Thanks for clarifying the existing numerical results. After an additional read, I am super satisfied with the benchmark versus the existing methods provided, and the wide variety of "metrics" used for assessment. While none of the target densities are in the "wow factor" territory, this is not necessary in my opinion. The paper is so strong, that I am sure that the focus of derivative works could be to "apply" or adapt the existing algorithm for high-dimensional and/or more difficult targets. Given that VI is the most popular approximate inference method at the moment, and that such approaches are heavily related to algorithms used to train quite a few of the popular classes of generative models at the moment, this paper has room to be of interest to many in the community, and I would some sort of spotlight or oral for this work. Thanks for the great submission.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC