acceptodds
Under review as a conference paper at ICLR 2027

Your VAE Knows Where to Sample, Not What It Doesn't Know: Merged Aggregated Posteriors as Deterministic Priors and Cluster Maps

Abstract

The aggregated posterior of a trained VAE, the -component Gaussian mixture of its per-sample posteriors, can be compressed to components by moment-preserving hierarchical merging: one deterministic pass, with no seed and no re-fitting, yields the entire nested family, every from to . We evaluate this object in three roles. As a latent density for out-of-distribution detection it fails: a pre-registered study (MNIST, CIFAR-10; with CelebA-64 added later) refutes it in every form we could construct — post-hoc, co-trained, IWAE-trained, as a discrete codebook, substituted into the pixel likelihood, and at every along the path — across three dataset regimes. On CelebA-64, latent density sits near or below chance against SVHN, but pixel based density (with a calibrated decoder) is even worse. The easy-regime (MNIST) verdict turns on the training recipe, not on the object: ordinary geometric augmentation with a convolutional architecture makes MNIST latent density beat that same model's pixel likelihood at ( seeds), while the CIFAR-10/SVHN inversion barely moves. As a structure recovery device it succeeds: co-evolving the merge partition during training raises the clustering ceiling decisively over the best plain pipeline (NMI vs. ; plain ELBO plus a KL floor suffices), though not to the of DEC, which gives up the generative model. A controlled comparison shows the hierarchy itself carries this structure ( NMI over post-hoc merging on the same encoder), while density depends only on the encoder ( AUROC on MNIST). As a generative prior it succeeds too: drop-in sampling beats on seeds on every dataset, a frozen merged codebook beats learned codebooks in an SQ-VAE-style model, and, unlike an equal- EM prior, the mixture cannot catastrophically corrupt the model's likelihood. Against EM, which samples as well, the merge's advantage is this safety, determinism and nesting, not accuracy. One principle organizes all three verdicts: moment-preserving compression preserves where the data lives, its support and cluster structure, not how likely it is.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
variational autoencoders, gaussian mixture models, aggregated posterior, deep clustering, out-of-distribution detection, OOD, generative models, unsupervised, probabilistic methods

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_cQJeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Starting from the aggregated posterior q(z)=(1/N)∑iN(μi,diag(σi2))q(z) = (1/N)\sum_i N(\mu_i, diag(\sigma_i^2)) obtained from training a VAE, we perform K-means-like reduction to K components by repeatedly merging pairs according to the moment matching rule of Eqs 1-2 in successive rounds of mutual nearest neighbours where each round constitutes a perfect matching reducing the number of components by half and is defined via scoring all pairs with diagonal Bhattacharyya distance. We thus get a deterministic hierarchical nested mixture for any choice of K in the range [1,N]. A co-training version of the model re-partitions the training data into K parts every epoch and introduces a cluster reconstruction loss together with a free bits constraint. The performance of this mixture is assessed in three separate tasks: As a latent density for OOD detection, our hypothesis gets falsified as the pre-registered. It underperforms on MNIST compared to both EM and kNN (AUROC 0.821 vs 0.922 and 0.932 respectively), it does worse than 0.5 on CIFAR-10 vs SVHN, like any latent score and on CelebA-64 (post-registration) it scores near chance with respect to SVHN. For clustering it has a smaller NMI variation across beta and seeds than EM (largest across seed std 0.04 vs 0.12) and the co-trained partition achieves 0.740 NMI on conv-MNIST (below DEC at 0.861). As a generative prior it outperforms N(0,I) in sampling and is roughly on par with EM. When substituted in as the prior within the pixel likelihood it has no effect on that score (whereas an EM mixture can sometimes cause it to collapse). As a frozen codebook in an SQ-VAE style model it outperforms a codebook learned from scratch.

Critical strengths
  1. The experimental technique is thorough. In particular, the VaDE and DEC reproductions (see Appendices G and H) stand out on their own merits. For example, the discovery that DEC's claimed accuracy relies on a dropout procedure at finetune time, which the DEC paper states does not apply.

  2. The negative OOD result is presented in full with explanations of part of it. Under calibrated Gaussian likelihood, the reconstruction error, reconstruction probability, and IWAE score deliver very similar AUROCs on SVHN (0.174, 0.174, and 0.172, respectively; Table 14). While the pixel score correlates with PNG file size (Spearman's ρ=−0.60\rho = -0.60), the latent score shows no correlation (ρ=−0.08\rho = -0.08). Changing the CelebA-64 likelihood model from a Bernoulli to a calibrated Gaussian reduces the pixel AUROC on SVHN from 0.769 to 0.153 (Table 16).

  3. The authors state where the method loses: EM samples as well, k-means and DEC cluster MNIST better, and GPU EM is faster.

Critical weaknesses

The merged mixture beats EM in some settings, but each time a simple alternative does as well or the control that would settle the comparison is missing.

  1. Novelty. The moment-preserving merge of Eqs. 1-2 is standard Gaussian mixture reduction (Salmond, 1990; Runnalls, 2007; Goldberger & Roweis, 2004). Ward's (1963) method already gives a deterministic, nested hierarchy, and merging mutual nearest neighbours is the reciprocal-nearest-neighbour scheme (Murtagh, 1983). None of this is cited or compared against, and the new parts, the Bhattacharyya cost and the forced perfect matching, are not shown to beat Runnalls' cost or Ward linkage.

  2. The perfect matching forces near-equal weights: true cluster sizes 490, ..., 6 come back as 256, ..., 124 (Section 3.2). All clustering results are on class-balanced data, so the gains over EM may come from this built-in balance. As a latent density the mixture is also far worse than EM (held-out log-density −19.3-19.3 against −4.0-4.0 at K=50K = 50; Table 28).

  3. The density-safety comparison is confounded. The EM prior drops the MNIST pixel score from 0.9997 to 0.85 or 0.80 on some seeds because its components shrink to σ=0.011\sigma = 0.011. But EM is fit to the means only, while the merge uses the posterior scales σi\sigma_i (Appendix K.3). EM on the full aggregated posterior, or with a variance floor, may well be just as safe, and until that is run the advantage is not established.

  4. The baselines are weak. EM samples as well or better, and beating N(0,I)\mathcal{N}(0, I) repeats Ghosh et al. (2020). k-means on the posterior means reaches 0.700 NMI on conv-MNIST, above the merge (0.594) and EM (0.641) (Table 35); the authors attribute the co-trained partition's 0.740 to "the co-trained encoder, not the mixture fitter". The CIFAR-10 gap in Table 11 comes from fixed σ=1\sigma = 1 models that reconstruct each image as its mean colour (Appendix D.10). The codebook study has no frozen k-means or EM codebook as a control (cf. Łańcucki et al., 2020).

  5. The unifying principle is not tested. Every latent score is below 0.5 on CIFAR-10 vs. SVHN (0.27 to 0.30 in Table 1), and the abstract says density "depends only on the encoder", so the failure belongs to the representation rather than the compression. "Preserves where the data lives ... not how likely it is" summarizes the results; no experiment in the paper could have contradicted it.

  6. Presentation. Most claims rest on 35 appendix tables. Table 1 uses one seed and the degenerate CIFAR-10 models, and the KK-path results use a different merger from Section 3.2 (Appendix D.9). There are many typos ("disentengle", "algoirthm", "wehre fittet"), a garbled contribution bullet (lines 105-106) and leftover template text in the Acknowledgments.

Questions for authors
  1. How does the merge compare with Ward linkage on the posterior means and with greedy mixture reduction under Runnalls' cost? Both are deterministic and nested, so they can be compared on clustering, sampling, and use as a prior density.

  2. Can you fit EM to the full aggregated posterior (using the σi\sigma_i), or with a variance floor, and repeat the experiment in which the EM prior collapses the MNIST pixel score?

  3. Since the perfect matching forces near-equal weights, do the clustering results hold on class-imbalanced data such as long-tailed MNIST or CIFAR-10? Does re-estimating the weights after merging change them?

  4. In Table 27, how does a frozen k-means or EM codebook on the same latents compare with the frozen merged codebook?

  5. Could Table 1 be recomputed with the calibrated CIFAR-10 models and 10 seeds, the configuration the appendix treats as the reference?

  6. Where is the pre-registration recorded, and what did it fix (hypothesis, metrics, analysis plan, values of KK)? An anonymized link or timestamp would help.

2 comments

We thank the reviewer for an accurate summary and a review in which every weakness is paired with the experiment that would address it. We agree with almost all of it. The revised PDF has not been uploaded yet, as some of the results are still in the finish line (next few hours); it will be followed by a list of changed numbers when the last runs finish. Numbers below are from a re-run on the full training split (10 seeds, calibrated decoders, one evaluator). Two answers go against the paper.

Q1. Ward on the posterior means: run, and it is as good as or better than anything of ours. 230 plain VAEs, NMI at K = 10, pooled over β, paired by encoder:

textbook Ward, meansour Ward-type pass, posteriorsk-meansp (Ward vs ours)
conv-MNIST0.7540.7350.7024e-6
MLP-MNIST0.6760.6690.6140.12
STL-100.1630.1650.1600.55
CIFAR-100.1240.1170.1253e-8

Our Bhattacharyya merge is lower still (0.607 on conv-MNIST). For clustering, the posterior scales, the moment-preserving update and the matching add nothing over the 1963 algorithm. We withdraw the clustering claim for the method.

Runnalls' cost: implemented, not finished. We implemented the exact closest-pair reduction under Runnalls' bound and verified it against brute force on small inputs, but it was too slow on the real posteriors to finish in time. So the reviewer's statement stands: the Bhattacharyya cost and the perfect matching are not shown to beat Runnalls' cost. Sampling and prior-density comparisons for both reductions were not run either. We will, of course, run it, and we do plan to incorporate that in future versions.

Q2. Not run in the rebuttal period. The comparison is confounded, as the reviewer says: EM saw only the means. Elsewhere we added EM on one posterior draw per training point, but not yet as the prior in the pixel likelihood. Until then, the revision does not call the merged prior safer than EM.

Q3. Not run, for lack of time. We agree it matters most: the perfect matching forces near-equal cluster sizes and all our data are balanced. Long-tailed MNIST and CIFAR-10, with and without re-estimated weights, are planned; we make no claim about imbalanced data and state the balance as a limitation.

Q4. Not run (needs training). Without a frozen k-means or EM codebook the study only shows that a frozen data-derived codebook beats one learned from scratch, as Łańcucki et al. (2020) suggest. The revision says so.

Q5. Yes. β = 1, K = 50, 10 seeds, AUROC:

scoreCIFAR-10 vs SVHNvs CIFAR-100MNIST vs FashionMNISTvs KMNIST
pixel likelihood0.0320.5171.0001.000
EM on means0.0790.5210.8820.915
EM on draws0.0850.5220.8610.898
merged mixture0.0900.5230.8090.852

The picture is unchanged: every score inverts against SVHN, and on MNIST the merge trails EM. kNN and Mahalanobis are being recomputed.

Q6. It is not an external registration, and we should not have used a word that suggests one. It is a design document in our own repository, first committed on 3 June 2026, before the result files (the earliest is dated 8 June). It fixed the claim ("matching or beating EM-GMM latent density at equal K on OOD AUROC"), score, protocol, baselines and failure modes. Deviations: Omniglot and Textures were never run; CelebA was added later. The revision says "fixed in advance in a dated design document" and provides it.

Novelty. Agreed. The revision cites Salmond, Runnalls, Goldberger & Roweis, Nielsen et al., Ward and Murtagh, and describes the method as classical moment-preserving reduction applied to a VAE's aggregated posterior with a different merge order. For a reducible criterion like Ward's, reciprocal-nearest-neighbour merging gives the closest-pair tree; with our cost it does not, and the closest-pair order chains under the Bhattacharyya distance. That is the only algorithmic difference we claim.

Principle. Agreed: it is a summary, not a tested principle, and the OOD failure belongs to the representation (EM inverts too). The remaining claim is small: the known failure also holds for the reduced aggregated posterior.

Baselines, presentation. Beating N(0, I) confirms Ghosh et al.; EM samples as well. Tables are regenerated by one script; typos and the template text are fixed.

Thanks for your reply and for doing the comparison of Ward with the suggested merge. It proves that all my concerns were true: Ward linking of the means is not worse, but at least as good as the suggested merge; the only algorithmic difference between the two is the order of merging; the claim of safety in relation to EM is retracted; and the "pre-registration" was just an internal design document.

The comparison to Runnalls' cost has not been done, and neither has the experiment with EM with the full aggregated posterior as the prior on the pixel likelihood, and class imbalanced data and a frozen k-means or EM codebook. All of these issues have not been addressed. The only thing left standing now is a deterministic hierarchy where the sampling prior is equal to EM (FID 65.7 on CelebA-64), and an OOD negative result, which also applies to EM.

Reviewer_SWBGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the aggregate posterior of a VAE, i.e. the mixture of variational posterior approximations conditioned on NN inputs, and compresses it to K<NK<N components via a deterministic hierarchical clustering approach. The VAE's OOD capabilities are examined using different image datasets. Judging from the Conclusions section, the proposed compressed method doesn't outperform simple baselines in any of the experiment / "regimes".

Critical strengths

The studied problem, OOD detection, is important, and the probabilistic nature of VAEs make them interesting models for trying to tackle this challenging task. Also, the idea of merging the mixture of Gaussians centered in the data points seems interesting.

Critical weaknesses

Presentation Unfortunately I found the paper difficult to follow starting from the title: a prior is a deterministic function, so "deterministic prior" makes little sense to me, and "Your VAE knows where to sample, not what it doesn't know" reads awkwardly. Next I could not parse the last part of the first sentence in the abstract: "yields the entire nested family, every KK from NN to 1."

Semantically, the abstract confuses me. What is the "object" being evaluated? The aggregate posterior, or its compressed version? The use of a number of non-standard terms and concepts -- co-training, latent density, at every KK along the path", pre-registered study, pixel-based density -- contribute to the confusion, and by the end of the abstract I am sorry to say that I honestly don't know what the paper proposes.

Regrettably, the introduction is of similar presentation character, and after multiple tries I was still not sure what is being proposed or studied in the submission.

The presentation unfortunately made it difficult for me to penetrate the paper, and I have no choice but to give the lowest scores in informativeness and soundness.

Questions for authors

I am sorry for providing such a crude review, but I highly suggest the authors to work on the exhibition of the method and the results produced in order to improve the presentation. I recommend thinking about the following: are you proposing a method for compressing the aggregate posterior? What parts of this contribution are then your own, and why is it a good idea to compress this posterior?

You seem to have many interesting areas where you want to evaluate the VAE's OOD capabilities, so I hope the authors gives the project another try after polishing the presentation.

1 comment

We thank the reviewer for the frank review. Its main message is correct: the submitted title, abstract and introduction did not say what the paper proposes. That is a failure of our writing. The revised PDF is not uploaded yet (controls for another reviewer are still running, and abstract and introduction are being rewritten as described here).

Are you proposing a method for compressing the aggregate posterior? Yes. A trained VAE maps every training point xix_i to a Gaussian q(z∣xi)q(z | x_i). The aggregated posterior q(z)=(1/N)Σiq(z∣xi)q(z) = (1/N) Σ_i q(z | x_i) is a mixture with one Gaussian per training point. We reduce it to KK Gaussians by repeatedly replacing two components with the single Gaussian of the same weight, mean and variance. Pairs are chosen by a distance between components (Bhattacharyya); in every round, all mutually nearest pairs are merged at once. There is no trainable parameter and no seed, and one run gives the reduced mixture for every KK.

What parts are your own? Not the merge of two Gaussians: that is classical mixture reduction (Salmond 1990; Runnalls 2007; Goldberger & Roweis 2004), which we should have cited and now do. Ours are: (1) applying it to a VAE's aggregated posterior at full size (tested up to 162,770 components) with an exact GPU schedule that never stores the N×NN × N distance matrix; we found no earlier work that reduces this object by merging, earlier reductions are learned (VampPrior and successors). (2) Merging all mutually nearest pairs per round instead of one closest pair per step; on VAE posteriors the closest-pair order collapses into one large component. (3) The measurement of what the reduced mixture is good for, including the negative results.

Why compress this posterior? For a fixed encoder, it is the prior that maximises the ELBO (Hoffman & Johnson 2016), and the standard normal prior does not match it. But with N components it costs O(N) to evaluate and memorises the training set. A KK-component reduction is cheap. Whether it keeps what one needs depends on the use: yes for sampling, no for detecting out-of-distribution inputs.

"Doesn't outperform simple baselines in any regime." Close to correct, and even more so after the controls for the other reviewers. For clustering, classical Ward linkage on the posterior means beats our merge; we no longer claim a clustering result. The OOD result is negative. As a prior, the reduced mixture improves on N(0,I)N(0, I), and EM of the same size samples as well. Against EM the method offers: no fitting, no seed, every K from one run. The revision states the contribution in these terms.

Title. "Deterministic prior": agreed, the construction is deterministic, not the prior; the subtitle is changed. We would like to keep the main title, which states the finding (the mixture tells you where to sample, not which inputs are unfamiliar), and accept that it only works if the abstract then says the same plainly.

"Nested family, every KK from NN to 1." The merge starts with NN components and ends with one; every intermediate mixture comes from the previous one by merging. So all K come from one run and form a tree that can be cut at any KK. The revision says it this way.

Which object is evaluated? The compressed KK-component mixture; it is now named every time.

Terms. Latent density: the density of the KK-component mixture at the encoder mean of an input, used as an OOD score. Pixel-based density: the model's marginal likelihood p(x)p(x), estimated by importance sampling. Co-training: updating the partition during VAE training; now a control, not a contribution. Pre-registered: removed from the abstract; in the body, the hypothesis and protocol were fixed in a dated design document before the experiments, which we provide.

Introduction. Reordered: the aggregated posterior and why to reduce it; the method in three sentences; the three uses and findings; contributions; related work on mixture reduction.

The reviewer notes that the soundness and informativeness scores follow from the presentation. We hope these answers, and those to the other reviewers on the experiments, allow a second look.

Reviewer_4L6wOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript introduces a moment merging based approach to compress any general VAE's aggregated posterior. The authors test the effectiveness of the compression across three axes, for OOD detection where it fails, for clustering, where it works but not as well as k-means, and it they claim it works better against EM as a generative prior. The paper provides experiments on several low-res vision datasets like CIFAR/SVHN/MNIST/CelebA. The main claims are not unified or supported by the provided evidence, be it the experimental design or results or the attached code. Further the numbers don't match across different parts of the paper, e.g. CIFAR10/SVHN scores as in Line 361 (0.03-0.04), Tab. 1 says 0.578 for SVHN, Tab. 30, Tab. 13 also have different numbers.

The contributions are also up in the air, it feels more like an heavily AI-driven investigative study than an impressive technical contribution.

Critical strengths

[S1]: The general approach is interesting.: I personally like the structure of the set of experiments and the authors clearly acknowledging a negative result in terms of the proposed approach not working for OOD samples as discussed in Line 337. The proposed compression approach is relatively simple and seems reasonable to use in practical scenarios.

Critical weaknesses

[W1]: Central claims are unsupported.: The main claims of the paper, for e.g. as in Sec. 3.3 don't agree with the results provided in Tab. 22. In the αmax=0\alpha_{max}=0 control, there is no cluster-reconstruction term, therefore no influence of the partition on the encoder, but yet there is a bump over EM? This seems suspicious, and makes me wonder where the numbers are coming from and if they are even accurate. What is this increase attributed to? Is it a comparison between clustering procedures? It does not seem like a comparison between co-training and post-hoc fitting to me. This renders a lot of the claims untrue. As in the abstract, "co-evolving the merge partition during training raises clustering ceiling", why does this hold true? A standard Ward-type rule can beat the proposed approach, so what even is the main headline? Furthermore, as in Tab. 35, k-means seems to have an NMI of 0.70.7 which is within the variance range of the co-trained partition for conv / MNIST, so what are the implications of these results? More so, the comparison with EM is also not fair given merging uses posterior scales whereas EM uses only means, so they are not compared on equivalent inputs. And on conv-MNIST in Tab. 5, EM beats merged on every β≤4\beta \leq 4 still, so even the merged advantage is not clear despite the unfair comparison grounds. No real imbalanced dataset is used and the reported gains don't provide any signal. There are no other baselines that are compared like VampPrior [1] or TwoStage VAEs and other learned priors?

[W2]: The attached supplementary code and numbers across the papers are contradictory.: There are a lot of working logs with several files (e.g. "results/server_h100/v2_20260914_145416/RECHECK_OLD_CODE.md") containing statements that raise concerns about the reliability of the results. Are the pixel_iwae numbers actually true then? Likewise, "docs/latent_gmm_methodology.md" says Omniglot and Textures were never run, so for Line 291 in Sec. 4, what does that refer to? There are several such eyebrow raising artifacts that were discovered in the provided supplementary code. What did the authors use for final reporting? Further the numbers don't match across different parts of the paper, e.g. CIFAR10/SVHN scores as in Line 361 (0.03-0.04), Tab. 1 says 0.578 for SVHN, Tab. 30, Tab. 13 also have different numbers. So which ones are to be presumed to be the correct one? There are several such instances in the clutters appendix which raise concerns about the veracity of the results. Further line 1905 in App. I says that the trained checkpoints are released amidst other things which are not present in the code, this also highlights an inconsistency within the different statements in the paper in different sections, and what they actually provide.

[1] Tomczak, J., & Welling, M. (2018, March). VAE with a VampPrior. In International conference on artificial intelligence and statistics (pp. 1214-1223). PMLR.

Questions for authors

[Q1]: See [W1] for details. Can you provide strong evidence to back the central claims of the paper. Does the partition influence the training in anyway for αmax=0\alpha_{max}=0? One way would be to provide a control experiment with the same encoder, without co-evolution, using the standard Ward-\k-mean-styled procedure. What would you attribute the clustering gains to? Could you compare with EM on fair grounds, and reconstruct Tab. 5? Do clustering gains hold on a truly imbalanced dataset? Could you clarify what the central claims of the paper are and how they are backed by evidence?

[Q2]: See [Q2] for details. Can you please clarify concisely why there is a mismatch in the statements in the READMEs in the code and the reported numbers and inconsistency across the numbers for the same settings for the same dataset across different parts of the paper? Could you provide some concrete evidence to enhance the reliability of the results?

1 comment

We thank the reviewer for reading the paper and the supplement so closely. Most factual points are correct; we addressed them to improve the paper. We re-ran every MNIST, STL-10, and CIFAR-10 experiment behind the clustering and OOD claims on the full training split with one evaluator, and ran the control the reviewer requested. The revised PDF is not uploaded yet: controls for another reviewer are still running, and we will upload it with a list of every changed number when they finish. (expected within the next few hours). Details follow in further comments.

Q1

Does the partition influence training for α_max = 0? No. The partition never enters the loss; cells differing only in a term excluded from the loss reproduce each other exactly (10/10 MNIST, 50/50 CIFAR-10). That row compares clustering procedures on one encoder. Our sentence crediting "the co-trained encoder" was wrong.

Same-encoder control with a standard Ward / k-means procedure. Done on every co-trained cell (10 seeds, NMI at K = 10). One Ward pass on the final encoder is within −0.007 to +0.041 NMI of the co-evolved partition in all 20 β-sweep cells; where significant, it favours the single pass. We withdraw the claim that co-evolving the partition raises the clustering ceiling.

What are the gains attributed to? (1) The clustering rule, which is not ours: textbook Ward linkage on the posterior means of a plain VAE reaches 0.754 NMI pooled on conv-MNIST, against 0.735 for our Ward-weighted pass, 0.702 for k-means, 0.661 for EM on equal inputs and 0.607 for our Bhattacharyya merge. The reviewer's sentence "a standard Ward-type rule can beat the proposed approach" is correct. (2) On MNIST only, the encoder: training with the KL floor and cluster term improves the Ward partition by 0.04 to 0.10 NMI at β ≥ 1 (conv); at β = 10 the KL floor alone explains it. On STL-10 there is no encoder effect.

EM on fair grounds, Tab. 5. We added EM fitted to one posterior draw per training point (the merge's input). Plain VAE, β from 0.5 to 10:

mergeEM, meansEM, drawsk-meansWard pass
conv-MNIST0.551–0.6520.468–0.6700.576–0.6980.616–0.7350.645–0.765
MLP-MNIST0.539–0.5980.381–0.5940.460–0.6280.531–0.6590.560–0.721
STL-100.129–0.1490.140–0.1520.146–0.1590.154–0.1680.155–0.170

On equal inputs EM is above the merge in every conv-MNIST and STL-10 cell. The merge is not the better clusterer.

Imbalanced data. Not run: we ran out of time. This is a real gap, because our merge pairs all components in every round and so produces near-equal cluster sizes by construction. It is the first experiment for the next version; we make no claim about imbalanced data.

Central claims after the re-run. (1) Method: a VAE's aggregated posterior can be reduced to K Gaussians by moment-preserving merging in rounds, exactly, without training or a seed. (2) OOD: the reduced mixture's density is not a usable OOD score; the failure of likelihood-type scores is known, we just show it also holds here. (3) Clustering: no claim for our merge. (4) Prior: sampling improves on N(0, I), confirming Ghosh et al. without fitting; EM samples as well.

Q2

The mismatch has one cause: we fixed two errors in code and main text before submission and failed to regenerate three tables. The notes the reviewer found are our own audit. The three values are one quantity (pixel AUROC, CIFAR-10 vs SVHN) under three scorings: 0.03–0.04 is the model's own calibrated likelihood; 0.578 (Tab. 1) an early model with decoder scale fixed to 1; 0.52–0.58 (Tab. 13, 30) calibrated models wrongly scored under a Bernoulli likelihood. The re-run gives 0.032–0.053. Evidence: one result tree per dataset, every table generated by one script, each cell records commit, split hashes, seed and checkpoint, and two independent evaluators agree within 0.0011 on 360 cells.

Other points

  • VampPrior and learned priors: not added for lack of time; the revision cites them and states the gap.
  • Omniglot, Textures: planned, never run; no reported number uses them. The sentence is corrected.
  • Checkpoints: the sentence should not have been there. They are too large for the 40 MB supplement; we are opening an anonymous Hugging Face repository and will post the link.
  • AI remark: The paper states how AI tools were used. The hypothesis, the design and the errors found are ours. We agree the submission did not make its contribution clear; the four claims above are what the revision states.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.