{"id":"3a6421b0-62c9-4b9d-8c57-00e34633aac0","arxiv_id":"1909.01839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A VAE with a Beta-Bernoulli process (Indian Buffet Process) prior achieves better disentanglement and reconstruction than Gaussian-prior VAEs on several benchmarks and improves supervised clinical tasks when combined with labels.","lead":"This paper introduces a VAE with an Indian Buffet Process prior, so the latent representation can grow in complexity while keeping dimensions independent, and tests whether that improves disentanglement. It reports better disentanglement than standard and VampPrior VAEs on image benchmarks and shows that adding the model to supervised clinical tasks improves accuracy and interpretability.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ELBO in Eq. (6) pairs a Concrete variational distribution with a Bernoulli prior; the KL term is not finite as written, so the stated training objective and the theoretical mechanism are unverified.","rationale":"The reader's weakest assumption focused on whether the Concrete and Kumaraswamy relaxations preserve independence and whether truncation limits growth. My concern is more fundamental: the KL divergence between the Continuous Concrete distribution and the discrete Bernoulli prior in Eq. (6) is not finite, so the ELBO as written is not a well-defined objective. This is an internal inconsistency rather than a disagreement with consensus. It is load-bearing because the paper's central mechanism—richer independent posterior density reducing competition in the ELBO—relies on the derived objective being the actual one being optimized. The concern is addressable: the authors can specify the relaxed prior, the surrogate KL, or release code showing the actual loss. Because this is a fixable exposition gap and the empirical results might survive under the correct objective, I keep the reader's CONDITIONAL verdict rather than moving to REJECT or UNVERDICTED. I mark agreement as 'disagree' because the reader's stated weakest assumption was not this KL mismatch, although it does involve the Concrete relaxation.","tokens_in":14618,"tokens_out":10396,"duration_ms":112255,"concrete_test":"Request the exact expression for the KL term in Eq. (6), or the released code's loss function. Independently compute, for K=1, tau=1, pi=0.5, the KL between the Concrete(pi,tau) density and Bernoulli(pi). If the integral is infinite, the implemented objective must differ from Eq. (6); then re-derive the ELBO with the actual prior/relaxation and re-run the dSprites MIG comparison under that objective to see whether the reported gap persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C defines the generative model with z_nk ~ Bernoulli(pi_k) (Eq. 4) and the variational family with q(Z|nu,X) = Concrete (Eq. 5). Eq. (6) then includes KL(q(Z|nu,X) || p(Z|nu)) as part of the ELBO. Since q is absolutely continuous on [0,1] while p has point masses at 0 and 1 (w.r.t. the appropriate base measure), this KL diverges; as written, the objective is not the ELBO. The paper does not state that p(Z|nu) is relaxed to a Concrete prior or that a surrogate KL (e.g., straight-through with Bernoulli KL, or KL between Concrete densities) is used. Without this clarification, the empirical improvements in MIG and the TC-D analysis cannot be attributed to the stated IBP objective—the actual objective may include a different prior or a heuristic gradient estimator. This is the weakest point in the argument because the central claim rests on the identifiability of the mechanism (reduced competition via richer independent density), and that mechanism is only meaningful if the optimized objective is the one derived.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IBP-VAE, in which the latent representation is generated from a Beta-Bernoulli process via Y = Z ⊙ A, and cIBP-VAE, which augments this model with a supervised classification loss for task-relevant representations. The central hypothesis is that a richer but independent posterior density improves disentangled representation learning by reducing competition between the reconstruction and independence terms in the ELBO. Experiments on colored MNIST, 3D Chairs, dSprites, skin lesion images, and 12-lead ECG report qualitative and quantitative improvements, including a MIG of 0.5477 at β=5 on dSprites versus 0.4786 for β-VAE and 0.4848 for β-VampPrior, and improved clinical classification over supervised baselines.","tokens_in":14798,"tokens_out":5718,"duration_ms":57212,"significance":"If valid, the paper makes a useful conceptual contribution by separating the independence of a latent density from its modeling capacity, and by demonstrating a nonparametric construction that can in principle grow with data complexity. The interpretability analyses on clinical data, including triggering units and label-swap reconstructions, are informative and address an underexplored application area. The paper also provides architectural details and paired statistical tests for the clinical results. However, the theoretical objective as written is not a well-defined ELBO, and the quantitative disentanglement evidence rests on a single benchmark with limited controls, so the central claim is only partially supported.","major_comments":[{"comment":"The KL term KL(q(Z|ν,X) || p(Z|ν)) in Eq. (6) is not finite as written. q(Z|ν,X) is a Concrete distribution on [0,1]^K (Eq. (5) and Appendix A-A), while p(Z|ν) is a product of Bernoulli distributions with point masses at 0 and 1 (Eq. (4)). With respect to any common base measure, the Concrete density is absolutely continuous and the Bernoulli prior is singular, so the KL divergence is infinite or undefined. The paper never states that p(Z|ν) is relaxed to a Concrete prior or that a surrogate KL (e.g., a straight-through estimator with Bernoulli KL) is used. Because the mechanism claimed in Section III-B rests on optimizing the ELBO in Eqs. (3)-(6), this is a load-bearing gap: the actual training objective may differ from the derived one, and the reported improvements cannot be attributed to the IBP ELBO without specifying the actual objective.","section":"III-C3, Eq. (6)"},{"comment":"The generative model is defined in the K→∞ limit in Eq. (4), and the paper claims that the latent density 'is able to grow with the complexity of the data' (Section I). However, all experiments use fixed finite truncations: K=100 for colored MNIST, K=50 for the skin lesion data, and K=50 for the ECG data, with no sensitivity analysis over K or the concentration parameter α. A finite truncation with fixed hyperparameters does not by itself establish unbounded growth, and the claim that the model grows with data complexity is not directly tested. Please add truncation-sensitivity experiments or substantially temper the nonparametric-growth claim.","section":"III-C1, IV-A, IV-C"},{"comment":"The only quantitative disentanglement metric, MIG, is reported on dSprites, with single runs and no standard deviations or significance tests. The values of β are varied over {1,5,10}, but the selection procedure is not described, creating a risk of best-reported selection. In addition, the 'best result reported in literature' for β-TCVAE in Fig. 4 is taken from another paper, so that comparison is not controlled for architecture, latent dimension, or training budget. These limitations do not invalidate the qualitative results or the internal comparison with β-VAE and β-VampPrior, but they weaken the central empirical claim that IBP-VAE improves over state of the art. Please provide error bars, multiple seeds, and ideally an additional benchmark or metric.","section":"IV-B, Table I, Fig. 4"},{"comment":"The claim that the latent representation is independent across dimensions holds for the generative model conditional on ν, but the variational posterior q(Z|ν,X) in Eq. (5) is specified conditional on the shared Kumaraswamy variable ν, and after marginalizing ν and applying the Concrete relaxation the marginal q(Z|X) is not obviously a product of independent factors. If the relaxed binary variables couple dimensions through ν, the 'reduced competition' mechanism in Eq. (3) may not hold for the actual posterior. Please state explicitly whether independence is intended for the conditional posterior q(Z,A|ν,X) or for the marginal q(Z,A|X), and discuss the implications for the stated mechanism.","section":"III-C1, III-C2"}],"minor_comments":[{"comment":"The index-code mutual information term is written as '-(log N - E_q(y)[H[q(n|y)]]))' with an apparent mismatched parenthesis, and the aggregated posterior q(y) is never explicitly defined as (1/N) Σ_n q(y|x_n). Please define q(y) and correct the formula.","section":"III-B, Eq. (3)"},{"comment":"The caption lists subfigures (a)-(e) in the text but also refers to panel (f), and the list in the caption is incomplete. Please make the caption consistent with the figure.","section":"Fig. 2 caption"},{"comment":"The caption contains a typo: 'classification accuray' should be 'classification accuracy'.","section":"Table II caption"},{"comment":"The TC-D analysis does not state how the total-correlation term is estimated from the trained models. Without an explicit estimator, the curves in Fig. 5 are not reproducible; please add the estimation procedure.","section":"IV-B, Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The main blocking issue is the undefined KL term in Eq. (6); the authors should be asked to specify the exact objective actually optimized. If a surrogate objective was used, the theoretical mechanism in Section III-B should be reframed accordingly. The quantitative evaluation also needs more rigor before the manuscript can be considered for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth a close read for the clinical nuisance-factor experiments and for raising a point people often skip—posterior richness and independence are different axes. But there is a formal problem in the objective. Eq. (6) writes KL(q(Z|ν,X)||p(Z|ν)) with q Concrete and p Bernoulli. Those are not densities on the same space; the KL is infinite as written. The paper never says it relaxes p to a Concrete prior or uses a surrogate KL, such as straight-through with a Bernoulli KL. That means the training objective is not the ELBO they derive, and the \"reduced competition\" mechanism in Eq. (3) is not what is being optimized. This is load-bearing: the whole story says the IBP prior's independence plus capacity causes the gains, but the actual optimization may be using a different objective or a heuristic gradient estimator.\n\nWhat is new: prior IBP-VAEs [26,27] focused on reconstruction, not disentanglement. Showing that an IBP prior beats Gaussian and VampPrior on MIG for dSprites, and that a supervised variant separates nuisance factors in skin lesion and ECG data, is a legitimate extension. The TC-D analysis is a nice empirical way to visualize the independence-vs-reconstruction tradeoff. I think the empirical results are plausible; they just are not well supported as written.\n\nSoft spots beyond the KL issue: all quantitative comparisons are single runs without error bars; beta is selected and the best reported; the dSprites comparison to β-TCVAE uses a number from the literature rather than a matched run; no code is released. The clinical experiments are suggestive but small (900 training images, 39 patients). The paper also does not engage Locatello et al. on identifiability, which is a real omission for a claim that richer priors improve disentanglement.\n\nBottom line: with the ELBO fixed—Concrete prior or explicitly stated Bernoulli KL surrogate—the paper could be a solid contribution. As it stands, the central mechanism is not demonstrated, even though the empirical direction is plausible. I would send it to a serious referee because the clinical material and the independence-vs-capacity question deserve peer time, but I would not cite it in its current form.","headline":"Worth reading for the clinical nuisance-factor experiments and the independence-vs-capacity framing, but the ELBO as written is not a valid objective—the Concrete q and Bernoulli prior make the KL infinite, so the stated mechanism is unverified.","tokens_in":15374,"tokens_out":3969,"would_cite":false,"duration_ms":38342,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a latent density which stays independent but grows in capacity improves VAE disentanglement, and demonstrates it with an Indian Buffet Process prior.","keywords":["variational autoencoder","disentangled representation learning","Indian buffet process","beta-Bernoulli process","nonparametric latent factor model","posterior modeling capacity","total correlation","clinical ECG"],"falsifier":"Train IBP-VAE on dSprites with the truncation $K$ set to a small value (for example 10) and again with a large value (for example 200), and measure MIG and total correlation at identical $\\beta$. If the small-$K$ model matches the large-$K$ model, the claim that the process grows with data complexity is not load-bearing. A second check: replace the IBP mask with a fixed finite set of independently learned Bernoulli probabilities of the same dimension; if disentanglement remains unchanged, then the nonparametric prior itself is not what drives the improvement.","tokens_in":14389,"feed_emoji":"🧩","tokens_out":5243,"duration_ms":48529,"temperature":0.7,"pith_summary":"The paper argues that part of what limits disentangled representation learning in variational autoencoders is not only whether latent dimensions are independent, but how much expressive room the latent density has. It claims that when a VAE is regularized toward an independent prior, the prior imposes two separate constraints: independence and a restrictive shape. An independent Gaussian posterior enforces both, and the restrictive shape competes with the reconstruction objective. The paper proposes IBP-VAE, which keeps independence but lets the latent factor model grow with data complexity, and reports that it disentangles more generative factors, including on clinical skin-lesion and ECG data.","feed_headline":"Richer latent density beats fixed priors at disentangling","feed_subtitle":"An Indian-buffet VAE disentangles more factors and reconstructs better than Gaussian-prior VAEs on benchmarks and clinical data.","key_machinery":"The machinery is the Beta-Bernoulli process, also called the Indian Buffet Process, used as a nonparametric prior over an unbounded binary feature matrix $Z$, combined with a Gaussian magnitude matrix $A$ through the elementwise product $Y = Z \\odot A$. This construction keeps latent dimensions independent while allowing a richer marginal density than an isotropic Gaussian. Inference uses a mean-field variational family with a Concrete relaxation for the Bernoulli variables, a Kumaraswamy relaxation for the Beta stick-breaking weights, and a Gaussian $q(A|X)$. The argument runs through the ELBO decomposition into reconstruction, total correlation, dimension-wise KL, and index-code mutual information: richer capacity lowers the dimension-wise KL competition, so independence and reconstruction improve together.","core_discovery":"The central claim is that preserving independence among latent factors while increasing the modeling capacity of the posterior density will improve disentangled representation learning in VAEs, especially when generative factors are complex. The paper instantiates this with a latent representation $Y = Z \\odot A$, where $Z$ is a binary feature-allocation matrix drawn from a Beta-Bernoulli process (the Indian Buffet Process in the infinite limit) and $A$ is drawn independently from an isotropic Gaussian. Because each data point's active latent features are inferred rather than fixed, the model can in principle grow with the data. On dSprites, IBP-VAE reaches mutual information gap (MIG) 0.5477 at $\\beta=5$, above $\\beta$-VAE (0.4786) and $\\beta$-VampPrior (0.4848), and its total-correlation versus distortion trade-off is better than both; on skin-lesion and ECG data, the supervised variant cIBP-VAE improves task accuracy over a supervised CNN and over c-VAE, and learns 'triggering' latent units that localize nuisance factors such as lesion color or pacing artifacts.","pith_inferences":["A direct testable extension is to sweep the truncation level $K$: if the IBP's infinity is doing the work, performance should improve with $K$ up to a point and then plateau, whereas a finite independent Bernoulli mask of the same total capacity should not match it.","The paper's framing suggests that other nonparametric or flexible independent priors, for example normalizing flows with factorized marginals, could reproduce the same gains, which would show that the growth property rather than the specific IBP is what matters.","In clinical settings, the triggering-unit mechanism could be turned into a hypothesis-generation tool: units that localize lesion color or ECG artifacts can be correlated with patient metadata to find novel confounders, though the paper does not test this."],"forward_implications":["IBP-VAE should keep an advantage over Gaussian-prior VAEs as the number or complexity of generative factors grows, because the latent factor count is inferred rather than preset.","Combining unsupervised nuisance disentangling with a supervised task (cIBP-VAE) should help in domains with many unlabeled nuisance factors, such as medical images or signals, where a plain supervised network struggles.","The learned triggering units provide a concrete interpretability mechanism: deactivating one latent unit removes a single semantic factor (font color, leg style, lesion color, ECG pacing artifact) from reconstructions.","The total-correlation versus distortion analysis is a reusable diagnostic for whether a VAE's posterior capacity is limiting disentanglement."],"supporting_citations":[{"why":"Establishes the independence-regularization baseline that IBP-VAE is compared against.","marker":"[2]"},{"why":"Supplies the ELBO total-correlation decomposition and the MIG metric used to quantify disentanglement.","marker":"[3]"},{"why":"Provides a complex-prior VAE baseline used to separate modeling capacity from independence.","marker":"[9]"},{"why":"Introduces the Indian Buffet Process that defines the unbounded latent feature prior.","marker":"[11]"},{"why":"Provides the rate-distortion-style analysis adapted into the TC-D comparison.","marker":"[20]"},{"why":"Contributes the Kumaraswamy relaxation used for the Beta stick-breaking posterior.","marker":"[24]"},{"why":"Previous IBP-in-VAE work that improves reconstruction but not disentanglement, setting up the gap addressed.","marker":"[27]"},{"why":"Gives the stick-breaking construction used to make the IBP prior tractable.","marker":"[31]"},{"why":"Supplies the Concrete distribution used to relax the Bernoulli feature indicators.","marker":"[34]"}],"fun_headline_variants":["Indian-buffet latents disentangle VAEs better","Richer posterior density improves VAE disentangling","Dynamic latent factors: key to VAE disentanglement","Beta-Bernoulli process lifts VAE disentangling","Flexible latent growth beats fixed priors in VAEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results rest on the assumption that, after the Concrete and Kumaraswamy relaxations, the elementwise product of the binary mask and the Gaussian magnitudes still yields latent dimensions that are truly independent, and that a truncated finite number of features behaves like an unbounded process on the datasets used; if either fails, the improved scores could come from extra capacity alone rather than from the independence-preserving growth the paper describes.","fun_headline_variants_meta":{"raw":{"variants":["Indian-buffet latents disentangle VAEs better","Richer posterior density improves VAE disentangling","Dynamic latent factors: key to VAE disentanglement","Beta-Bernoulli process lifts VAE disentangling","Flexible latent growth beats fixed priors in VAEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000712,"raw_usage":{"total_tokens":3275,"prompt_tokens":1087,"completion_tokens":2188,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":703,"completion_tokens_details":{"reasoning_tokens":2111}},"tokens_in":703,"tokens_out":2188,"duration_ms":17106,"temperature":1.0,"reasoning_tokens":2111,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:22:02.910342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train IBP-VAE on dSprites with the truncation $K$ set to a small value (for example 10) and again with a large value (for example 200), and measure MIG and total correlation at identical $\\beta$. If the small-$K$ model matches the large-$K$ model, the claim that the process grows with data complexity is not load-bearing. A second check: replace the IBP mask with a fixed finite set of independently learned Bernoulli probabilities of the same dimension; if disentanglement remains unchanged, then the nonparametric prior itself is not what drives the improvement.","supporting_citations":[{"cited_title":"beta-vae: Learning basic visual concepts with a constrained variational framework,","cited_arxiv_id":null,"evidence_quote":"Establishes the independence-regularization baseline that IBP-VAE is compared against."},{"cited_title":"Isolating sources of disentanglement in variational autoencoders,","cited_arxiv_id":null,"evidence_quote":"Supplies the ELBO total-correlation decomposition and the MIG metric used to quantify disentanglement."},{"cited_title":"V AE with a vampprior,","cited_arxiv_id":null,"evidence_quote":"Provides a complex-prior VAE baseline used to separate modeling capacity from independence."},{"cited_title":"The indian buffet process: An introduction and review,","cited_arxiv_id":null,"evidence_quote":"Introduces the Indian Buffet Process that defines the unbounded latent feature prior."},{"cited_title":"Fixing a broken elbo,","cited_arxiv_id":null,"evidence_quote":"Provides the rate-distortion-style analysis adapted into the TC-D comparison."},{"cited_title":"Structured varia- tional autoencoders for the beta-bernoulli process,","cited_arxiv_id":null,"evidence_quote":"Previous IBP-in-VAE work that improves reconstruction but not disentanglement, setting up the gap addressed."},{"cited_title":"Stick-breaking construction for the indian buffet process,","cited_arxiv_id":null,"evidence_quote":"Gives the stick-breaking construction used to make the IBP prior tractable."}],"review_version":1}