Paper ID: | 7343 |
---|---|

Title: | Variational Bayes under Model Misspecification |

he paper characterizes the asymptotic behavior of the variational Bayes approximation under model misspecification. I found none of the results particularly surprising since they are intuitive and expected, especially based on the results of [17] & [29]. (The proof also seems to be a straightforward extension.) I also find that the authors make too big of a deal about the model misspecificaiton error asymptotically dominating the variational approximation error, making this old observation look like a new contribution. (As an approximation becomes more accurate asymptotically, obviously the model misspecificaiton is going to dominate.) In particular, I find the section 2.2 verbose and repetitious in explaining the obvious intuition. The authors should be more explicit about the limitation of the result discussed in Section 2.3. I believe the local asymptotic normality assumption (Assumption 5) often fails under the type of models considered in Section 2.3 (e.g. hyper-parameters of the Gaussian process regression does not concentrate under the in-fill aymptotics). All that said, the result is a welcome addition to the literature on the theoretical properties of the variational Bayes approximation under model. I also appreciate that the authors kept the intuitions very clear (without making things unnecessarily complicated) and the paper is generally very easy to read. Minor comments: - Assumption 4 & 5 in the supplement are not just analogous but are essentially identical to Assumption 2 & 3. In this case, why not make the assumption for Section 2.3 more clear in the main manuscript? - Line 127, `tests $\phi_n$`: a `test` is undefined. I suppose it is a compactly supported smooth function, but is certainly not in the standard vocabulary of stats/ML audience. - Line 166, `limiting exact posterior`: this terminology threw me off and was very confusing to me because `limiting` and `exact` are contradictory. I suggest to call it just `limiting posterior`. - Line 166, `\theta` vs `\tilde{\theta}`: does the parameter with tilde play a different role? If not, it is just confusing. - Line 172, `\mathcal{Q}^d`: is the same as `\mathcal{Q}`? I suppose it is meant to emphasize the dependency on $d$, but the dependency was always there and the sudden change of notation is just confusing. - Line 289, `simulation corroborates... the limiting VB posterior coincide with the limiting exact posterior.`: I don't think this claim is true. Just looking at RMSE does *not* establish that that the two distributions (VB and MCMC) are close. Response to author feedbacks: Straitening out the main contributions and clarifying the limitations will certainly make the paper more worthwhile to the readers. With a successful revision, the paper will deserve the score of 7 (though there is no 2nd round review unfortunately).

The paper is clearly written and very well presented. Although inherently technical, the results are explained both precisely and in plain language, with proof sketches to convey intuitive understanding. This paper is a great model of clear communication of technical results. The results are novel to my knowledge and well situated in terms of previous literature. I found no obvious technical errors, although I wasn't able to closely check the proofs in the supplement. My impression is that the results themselves don't involve significant new technical ideas and are more or less straightforward extensions of previous theorems. Nonetheless, actually doing this technical work is a valuable contribution. My main concerns about significance, which largely apply to Bernstein-von Mises theorems more generally, is that by focusing on the asymptotic regime the work assumes away essentially all of the practically relevant structure in Bayesian inference problems. Behind all the technical machinery, the intuition behind these proofs (which, to its credit, the paper does a good job of conveying) is that for identifiable models in the iid asymptotic regime, the likelihood dominates the prior, and the posterior concentrates at a normal shrinking to a point mass, so we can ignore the prior and we can mostly ignore posterior uncertainty. But if you really believe you're in this regime, why not save yourself the trouble of VB and just fit an MLE? The argument that the MLE minimizes KL between the true data distribution and a misspecified model is so trivial that it's more of an observation (that the non-constant part of KL is just the expected model log likelihood) than an argument. This work dresses up that argument with substantially more mathematical machinery, but not (as far as I can tell) much more insight. It tells us that if you run VB in the setting where there is no uncertainty to quantify, it preserves the properties of a point estimate. This is well and good -- it's always possible that something could have gone wrong, and there's some pedantic value in checking that it doesn't -- but it's also kind of not the point of VB. Practical Bayesian inference involves quantifying uncertainty; without that, why are we here? We only get to do so much with our wild and precious lives, and it's not my place to question the authors' choices, but I can't help but view this as something of an example of math for math's sake with limited takeaways for the broader field. All that said, theoretical papers are in scope for NeurIPS, and this one is well done within (as far as I'm qualified to judge) the standards of the community.

Update After Rebuttal ---------------------- After reading rebuttal and other reviews, I continue to argue for acceptance and leave my original score (good paper, accept) unchanged. I thank the authors for willingness to discuss issues like local optima in the VB result in a revision, and also for willingness to describe simulations more carefully and share complete simulation code. I was glad to see comments from other reviewers about relevance of the LAN assumption or the Bernstein-von-Mises approach, and I hope a revision addresses these issues in more depth (as the rebuttal hints). In particular, I'd encourage a thoughtful response to the question R2 raises: "... if you really believe you're in this regime, why not save yourself the trouble of VB and just fit an MLE?". This which isn't really addressed in the rebuttal, and I think it's important to both raise and answer the point in the paper. Review Summary -------------- I appreciate the paper's focus on determining what happens to the optimal approximate posterior in the inevitable case that the model is "wrong", and thus the contribution of providing theorems to guide our understanding of how tractable approximations like VB behave in the asymptotic limit will be of interest at the conference. I wish the paper had a bit more to say about how local optima fit into this story (at least acknowledging the practical problems) and I'd like to see more details about the simulations, but overall this seems like interesting work and would lead to productive discussions as an accepted poster. Technical Comments ------------------ ## Comment on local optima? One inevitable issue with the practical outputs of variational Bayes iterative optimization is that we almost surely return a local optima rather than a global one (the variational ELBO is usually non-convex for any model of interest, such as the LDA topic model in the examples). While the theorems rather nicely govern the behavior of the global optima, I'm not sure they can say anything about local optima. Given this, I'm a bit surprised that the Simulations in Sec. 3 seem to gloss over the practical issues of local optima in VB as well as the practical mixing issues of MCMC on real finite datasets. For example, I'd be very reluctant to say that HMC has ever really converged to an "exact posterior" for a model like LDA, especially given only the results of one chain (often mixing issues aren't apparent until one chain does much better than 50 others). Perhaps with simple 15-dimensional observations it's possible to avoid serious problems (but still, if you look closely the curves don't match perfectly, so maybe there is a complication here). Would be interesting to see results from a much larger topic model fit (perhaps with several hundred observations), where local optima might play a larger role. ## Comments on rates of convergence? For practitioners that will never have access to "infinite" data, I wonder if the analysis here provides insight about when to expect that the convergence conditions are "close enough" to being satisfied (e.g. as argued in the Fig 2 plots, where 20000 examples is a proxy for "infinite" data). ## Tail condition assumption: How strong is it? When providing the key assumptions required for Thm. 1, while I agree it's likely true that many common priors satisfy the required tail condition -- second derivatives of log p(\theta) do not grow faster than exp(\theta^2) -- I wonder if there are any known (or easy-to-construct) counterexamples? It might help to give a bit more intuition for what kind of smoothness this implies. Comments on Simulations ----------------------- * Code to perform the simulation study is missing (seems that only the stan model specification is provided, not the code to produce the plots). Sharing this would be crucial to help readers reproduce results and understand their practical import. * How are the errorbars/intervals in Fig 2b calculated? Is this showing the full spread of the KL across topics, as well as average (mean) across topics? * Likewise, how are the errorbars in the remainder of Fig. 2 calculated? Should we be bothered that they do not seem to go to zero with more data (N)? * I might recommend that each figure show something like a 10x higher maximum N values than currently used. Doesn't seem you're showing real convergence (e.g. the intervals in Fig 2c at N=20000 don't quite line up yet). Presentation Comments --------------------- I found Sec. 1, especially the Main Idea, quite easy to follow. Well done. Fig. 1 does a nice job illustrating the main idea, with perhaps the exception that visually, the change in distribution over \theta from observing one to infinite data is larger than the whole space of possible factorized distributions Q.