NeurIPS 2020

Fixed-Support Wasserstein Barycenters: Computational Hardness and Fast Algorithm


Review 1

Summary and Contributions: The authors analyze the problem of computing wasserstein barycenters with fixed discrete support. They theoretically demonstrate that while computing the barycenter of m=2 histograms can be cast as a minimum cost flow problem (MCF), when m>=3, it cannot be cast as such, rendering off-the-shelf MCF solvers inefficient. The authors then introduce an algorithm that utilizes entropic regularization to yield an approximate solution to the wasserstein barycenter problem in state of the art complexity.

Strengths: The main strengths of this paper are the theoretical analysis and the fact that the authors introduce a considerably faster algorithm than is previously been presented. The demonstration that the FS-WBP is not an MCF is pedagogical and the combinatorial proof techniques may be of use in other OT problems. The major result of the paper is the algorithmic complexity improvement and that is quantified well in the numerical experiments.

Weaknesses: The main weaknesses are the lack of quantitative and qualitative evaluations. As the authors have pointed out, there are many applications of WSB in image and shape analysis and computer vision in general. The results on MNIST while there are many other applications, leaves something to be desired.

Correctness: To best of my ability, the proofs are sound and correct.

Clarity: Some of the notation in the main text are not self contained i.e. r( ) and c( ) had to be inferred from the supplementary.

Relation to Prior Work: While there have been several attempts to introduce fast algorithms to compute wasserstein barycenters, the authors compare against the best two in terms of complexity in n and \epsilon (IBP and accelarated IBP, respectively) and show that they outperform both.

Reproducibility: Yes

Additional Feedback:


Review 2

Summary and Contributions: This paper studies the FS- WBP, in particular whether it is minimum cost flow (MCF) problem (which is not for m >= 3, n>=3), and proposes a new fast IBP algorithm while studying its complexity and convergence accuracy. The first result is relevant as it justifies the entropic regularized barycenter problem as a computationally efficient algorithm. The algorithm proposal complements the analysis with improved accuracy and convergence results w.r.t. IBP.

Strengths: The discussion about totally unimodular matrices and MCF equivalence is interesting and justifies either looking for other LP efficient reformulations, or approximate methods, which is of big interest to solve the problem. I find the new algorithm proposal also of interest, although more clarity on the derivation of the algorithm and comparisons could improve the paper.

Weaknesses: I find the numerical experiments a bit lacking, even though the authors claim their experiments are extensive (which is a subjective expression and open to criticism). For example, I miss comparisons with [31] and [55], which should be very competitive. Even if [55] is not an entropic regularized formulation, it may still be very competitive. Comparing with [55] may justify why an entropic regularized formulation is needed. Also, why not compare with prox-IBP from [31]? Simulations with Gurobi are nice to see, but they are almost unnecessary. It is known that general solvers scale badly with dimensions and that's why specialized algorithms are useful. Regarding the MNIST visualization, I did not understand why the authors claimed FastIBP provide a "smoother" solution, while I think it is sharper (and sharper should be better, more visually appealing and closer to the unregularized barycenter?). I was surprised in Algorithm 2 by steps 1 and 3, and I would like such steps be better explained in text. Notation should definitely not be in the appendix, which I did not find until specifically looking for something on the appendix.

Correctness: The claims and methods are probably correct, although I did not verify the proofs carefully. The empirical methodology is correct, although I think more comparisons would improve the paper quality.

Clarity: The paper is well written. A few typos: -page 3 line 114: "progress has been, (...)"

Relation to Prior Work: Prior work seems to be thorough and is clearly discussed.

Reproducibility: Yes

Additional Feedback: After rebuttal: Thanks for your response and addressing my concerns. I have increased my evaluation to accept. Please, add the new comparisons you mention in your response to the final version.


Review 3

Summary and Contributions: The paper proves the Wasserstein barycenter problem is not a minimum cost flow problem, and develops a faster regularized algorithm than the currently popular Sinkhorn method for computing the fixed support barycenter.

Strengths: Well supported theoretical claims that form a significant contribution to the theory of Wasserstein barycenters. The authors additionally propose an improved algorithm for computing regularized barycenters that is both faster and more efficient than the Sinkhorn method.

Weaknesses: This paper does not benefit from the short form and limited number of pages of NeurIPS and would be better suited for a journal. The two main contributions of the paper are Theorems 3.5 and 4.3 whose proofs take several pages of the supplemental. The numerical verification is quite weak (again, likely due to space limitations) as only two setups are considered. I do not think these issues are significant enough to recommend rejection, but I would like to see either more intuition, or more empirical verification in a revision.

Correctness: To my understanding, every statement in the paper is correct. The proofs are detailed and easy to follow.

Clarity: The paper is clearly written, but very terse. Much of the interesting content is hidden in the supplemental, and the paper is rife with abbreviations and rushed definitions and statements.

Relation to Prior Work: All relevant prior work is acknowledged.

Reproducibility: Yes

Additional Feedback: After rebuttal: The authors address my concerns on empirical testing. My score will not change, as I already wish to see the paper accepted.


Review 4

Summary and Contributions: This paper considers approximation in the fixed support Wasserstein barycenter problem. For this problem, one wishes to find the barycenter of a set of discrete probability measures. Instead of letting the support of the discrete barycenter vary, instead one fixes a support beforehand and solves an optimization problem over an n-dimensional simplex. While the OT problem can be cast as a min-cost-flow problem, the authors resolve an open question and show that the fixed support Wasserstein barycenter problem in general is not. Therefore, other algorithmic solutions that are not optimal for the MCF problem must be pursued. The authors propose to solve the linear program efficiently using entropic regularization as well as an accelerated iterative Bregman projection (FastIBP) scheme. They justify this algorithm theoretically by showing that it achieves comparable approximation accuracy with some state-of-the-art algorithms. They also give some experiments that show the tradeoff achieved by the method: fast speed with good accuracy. Finally, an experiment on the MNIST data shows that this algorithm converges very quickly for small regularization, which allows them to achieve sharp barycenters on each digit.

Strengths: - The work resolves an important open question and shows that the fixed support Wasserstein barycenter problem is not a minimum cost flow. This is nontrivial, and the authors give a sketch of how they manage to prove this using some classical combinatorial theorems as well as carefully writing out the constraints in the fixed support Wasserstein barycenter problem. They give a specific counterexample in the case of n=3 (number of points) and m=3 (number of measures). - Since other algorithms are needed, the authors propose a new algorithm based on entropic regularization that achieves a novel theoretical complexity bound -- its dependence on the approximation parameter epsilon is better than the plain iterative Bregman projection (IBP) algorithm. - Experiments show that the proposed FastIBP method is faster than existing methods while achieving good approximation accuracy. In the regime tested, the FastIBP algorithm seems to perform better than other standards, such as IBP and BADMM. -The experiments on MNIST show that the algorithm gives sharper barycenters than the non-accelerated iterative Bergman projection algorithm in a fixed amount of time.

Weaknesses: - The setting is restrictive, as it requires a setting where the barycenter support must be pre-specified. In general, the ideas for solving the fixed support problem do not seem to generalize to the general Wasserstein barycenter problem. - The algorithm they propose does not achieve the best complexity bound in all regimes. There seems to be a tradeoff between approximation factor and size of the empirical distributions, since it is obviously faster than IBP (it is an accelerated version), but it does not achieve a better dependence on approximation factor when compared with accelerated IBP and APDAGD. However, it does have a better dependence on the size of the empirical distributions and barycenter, n, than these two methods. This tradeoff is not explored in the paper. Not hints as to interesting theory from the analysis of FastIBP are given. -The explanation of the algorithm is short, and instead of giving heuristic ideas they instead just give the steps of the algorithm. Can anything be learned from this method or is it actually just taking some past work and applying it to this problem? Why is it able to achieve the better dependencies over other methods? - The authors don't compare with some mentioned methods, accelerated IBP and APDAGD, in the experiments. - It is unclear whether the assumption that all measures are supported on the same number of points necessary for these results - is the extension easy based on these results? Does the algorithmic complexity result extend?

Correctness: The claims and methodology appear to be correct.

Clarity: While the paper is generally well written and manages to fit in a lot, there are a few ways in which it could be made more clear. - I think this one way this could be addressed by restructuring the paper - when reading Section 2.3, it is unclear how these results will be used later on especially since Section 3 immediately follows it. I assume that these results are used in the complexity bounds given in Section 4, but only the bound in Lemma 2.1 is referenced there. The connections should be spelled out explicitly, and maybe moved together so that it flows better. - The proof in Section 3 is hard to follow. The matrix R appears to not be defined until (8), but it is referencing the matrix in (7). What does it mean for the "rows of the matrix to be in the set I"? How were these sets I and the j set chosen - did you just search until a counterexample was found? Propositions 3.1 and 3.3 are referenced at the end of the proof, but they should be referenced earlier as they define the notions used within the proof, as language such as "Indeed, columns 1 and 2 imply that rows 1, 4 and 5 are in the same set." -- here you are not explicitly stating it, but it is because you are deriving a contradiction to Proposition 3.2. Therefore, a lot of the language and notation is implied by what is written around it instead of explicitly and directly stating what is meant. Cleaning this up would greatly help with comprehension of this proof. It seems like maybe some things got lost in summarizing/shortening a longer paper. - Please make sure everything is defined within the paper. For example, where is "normalized objective value defined" (as used in the experiments)?

Relation to Prior Work: It is clearly discussed how this work relates to past works.

Reproducibility: Yes

Additional Feedback: Overall, this paper tackles an important problem and makes a step towards a better understanding of a simpler case of the the Wasserstein barycenter problem. While showing the problem to not be a MCF, the authors still give an algorithm with complexity at least somewhat comparable to state of the art. I mentioned the main downsides in the weaknesses section, but in summary, these are: 1) the writing needs to improve, as it feels like much was left out and unclear in the shortening of the paper, 2) the algorithm is not tested in regimes of varying n to see how the algorithm scales compared to other existing algorithms, 3) the algorithm was not compared with other algorithms with good complexity (accelerated IBP and APDAGD) and it is not said why they are not compared. However, even with these shortcomings, I think that this work is important and interesting enough to warrant acceptance into NeurIPS, as it both solves an open question and gives a new algorithmic approach to this problem, which seems to perform well in the limited experiments given. #### POST REBUTTAL #### After seeing the authors response, I am satisfied that they address my main concerns. That is, they give some experiments in the regimes of varying n in the supplementary material, they discuss the novelty of the FastIBP method, and they will improve the discussion of the paper. Since I have already given a high score, I will leave it as it is.