Pith. sign in

REVIEW 2 major objections 6 minor 61 references

A Dirichlet-process weight prior lets Gaussian-splat mixtures choose their own component count in closed form and match fixed-budget fits with far fewer Gaussians.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 08:20 UTC pith:ANU25AQW

load-bearing objection Clean, honest nonparametric complexity control for VBGS-style splat fitting; the math and deconfounded experiments hold up, and the deliberate scope limits are stated clearly. the 2 major comments →

arxiv 2607.10912 v1 pith:ANU25AQW submitted 2026-07-12 cs.CV

DP-Splat: Bayesian Nonparametric Complexity Control for Gaussian Splatting

Aqi Dong This is my paper
classification cs.CV
keywords 3D Gaussian SplattingDirichlet process mixturesvariational inferencecomplexity controlstick-breaking priorsparse finite mixturespredictive uncertaintycoordinate ascent
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

3D Gaussian splatting usually decides how many Gaussians to use with hand-tuned growth and pruning rules, or a user-chosen cap. This paper replaces the fixed finite weight prior in conjugate variational Gaussian splatting with a truncated stick-breaking Dirichlet-process prior (and a sparse overfitted Dirichlet as a theory-backed alternative), so the number of occupied components becomes an output of inference while every update stays a closed-form coordinate-ascent step. A stochastic natural-gradient variant keeps per-step cost independent of the number of points. The authors show that effective complexity adapts to the data, recovers the true count within one component on well-separated synthetics under a suitable concentration, and matches or exceeds a fixed-K baseline’s held-out color prediction on 3D scenes with several times fewer components. They also give a rigorous truncation-error bound, calibrated predictive color variance, and document that under mean-field coordinate ascent the usual asymptotic ordering of Dirichlet-process versus sparse finite mixtures reverses.

Core claim

Switching only the mixture-weight prior of conjugate variational Gaussian splatting to a truncated stick-breaking Dirichlet process (or a sparse overfitted finite Dirichlet) makes the number of occupied components adapt to the data while preserving closed-form CAVI and SVI updates. The prior’s practical contribution is complexity selection, not better per-component efficiency: at matched effective budgets a converged Dirichlet-process fit ties a converged fixed-K fit, yet on 3D scenes it matches or exceeds the fixed-K baseline’s held-out color prediction with 5.9–7.6× fewer components. Under mean-field coordinate ascent at practical sample sizes the Dirichlet-process prior resists over-split

What carries the argument

The weight-prior switch: replace the finite symmetric Dirichlet over mixture weights with a truncated stick-breaking Dirichlet-process prior (primary) or a sparse overfitted Dirichlet (secondary), so Beta/Categorical/Normal–Inverse–Wishart (and optional Gamma) coordinate updates remain closed-form while the occupied-component count ˆK is an inference output.

Load-bearing premise

The whole argument rests on treating scene complexity as a mean-field mixture of colored points from depth-lifted RGB-D data, with coordinate ascent that permanently kills empty components and with rendering treated only as a downstream map that carries no guarantees.

What would settle it

On well-separated synthetic mixtures with known true K and a regime-appropriate fixed concentration, check whether recovered effective component count stays within ±1 of truth across N from 10^3 to 10^5; on the same 3D scene point clouds, check whether a single-pass Dirichlet-process fit at matched truncation still matches the fixed-K baseline’s held-out point-color PSNR while using several times fewer occupied components.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Scene fits can report an automatic effective Gaussian budget instead of requiring a user cap or densification heuristics.
  • Held-out color prediction quality comparable to fixed large-K variational fits becomes available at roughly 6–8× fewer occupied components on the tested scenes.
  • Posterior predictive color variance is available in closed form and is well calibrated when the generative model is correct.
  • Truncation level T need only be a few times larger than the eventual ˆK; the corrected bound shows truncation error is negligible at typical scene budgets.
  • Practitioners should not read ˆK as a consistent estimator of a true K; it is a pragmatic complexity control shaped by coordinate-ascent dynamics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Uncertainty-aware next-best-view or densification could be driven directly by the closed-form Var[c|s] maps without extra networks.
  • Adding merge/split or reordering moves would likely close the documented gap between variational ˆK growth and exact-posterior asymptotics.
  • The same weight-prior switch could be dropped into other conjugate mixture scene models (e.g., continual SLAM variants) without redesigning the rest of the pipeline.
  • If rasterized novel-view metrics track the point-level gains, automatic complexity control would become a practical alternative to post-hoc pruning of over-parameterized 3DGS models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper replaces VBGS’s fixed-K finite Dirichlet weight prior with a truncated stick-breaking Dirichlet-process prior (primary) and a sparse overfitted finite Dirichlet (secondary), keeping the colored-point mixture fully conjugate so that CAVI and natural-gradient SVI remain closed-form. It supplies an exact ELBO monotonicity guarantee (Prop. 1), a rigorous finite-α truncation bound that corrects the commonly quoted anti-conservative large-α Ishwaran–James approximation (Prop. 2), and an explicit account of what the effective component count ˆK does and does not estimate under mean-field coordinate ascent. Empirically, ˆK recovers true K within ±1 on well-separated synthetics under regime-appropriate α; a deconfounded image comparison isolates the DP prior’s contribution to complexity selection rather than per-component efficiency (Table 1); on NeRF-synthetic RGB-D scenes DP-Splat matches or exceeds VBGS held-out point-color prediction with 5.9–7.6× fewer components; predictive color variance is well calibrated on model-matched data; and the paper documents a variational inversion of exact-posterior asymptotics across three orders of magnitude in N.

Significance. If the results hold as stated, this is a careful and useful bridge between classical Bayesian nonparametric mixture methodology and modern Gaussian-splatting scene representations. The contribution is deliberately statistical rather than leaderboard-oriented: closed-form nonparametric complexity control, a corrected truncation bound of practical relevance at the α regime these models use, calibrated posterior-predictive color variance, and a documented gap between variational practice and posterior asymptotics. Strengths that raise confidence include machine-checked update oracles, Monte-Carlo ELBO-term checks, cross-code regression against reference VBGS, multi-seed config-driven experiments with machine-readable records, and an unusually honest treatment of what ˆK estimates. These are genuine methodological assets for the community working on principled complexity control in splat-style models.

major comments (2)
  1. [§6.8, Table 2, Abstract] §6.8, Table 2, and the abstract’s 3D claim: the reported match/exceed of VBGS with 5.9–7.6× fewer components is on a rasterizer-free held-out point-color metric (E[c|s] on an unseen frame), not on novel-view rasterized PSNR/SSIM. The manuscript scopes this correctly in §3, §5 and §7, but the abstract and introduction still frame the work as complexity control for Gaussian Splatting practice. For the applied claim to land with a CV audience, either (i) add at least a limited rasterized quality-vs-budget comparison on the same fitted models, or (ii) rephrase the abstract/intro so that the 3D result is unambiguously a point-mixture predictive comparison, with rendering-grade evaluation deferred. As written, a reader can over-read the scene numbers as rendering results.
  2. [§6.2, Figure 2] §6.2 and Figure 2: recovery of K_true within ±1 requires a regime-appropriate fixed α (α=0.1 for K=3, α=1 for K=10, α=5 for K=30), and variational learning of α tracks the local optimum rather than correcting it. This is reported honestly, but it is load-bearing for the claim of automatic complexity control: without a usable α schedule or a merge/split/MFM mechanism, the practitioner still faces a concentration choice that plays a role analogous to choosing K. The paper should state more prominently (abstract or contributions list) that ˆK adaptation is automatic only conditional on a regime-appropriate concentration, and that learning α does not remove that requirement under plain CAVI.
minor comments (6)
  1. [§6.1, Figure 1] Figure 1 caption and §6.1: the 10% of DP seeds that land in visibly lower local optima are reported but not quantified in ELBO or ˆK terms. A one-line summary of the ELBO gap and resulting ˆK for those seeds would help readers assess practical risk.
  2. [Table 1] Table 1: the full-budget VBGS column is given only as a range (18.4–33.5 dB) while other columns report means and min/max. Adding a mean for the full-budget column would make the −6.47 / −4.79 dB gaps easier to interpret.
  3. [§4] §4, SVI paragraph: the schedule parameters (τ0=64, κsched=0.7, |B|=2^16) are stated without a short sensitivity note. Even a single sentence that conclusions are stable under modest schedule changes would strengthen the scaling claim.
  4. [§4, Effective complexity] Notation: n_min for the effective-component threshold and N_k for soft counts are easy to confuse on a first read; a brief reminder at first use of ˆK would help.
  5. [§2] Related work: BOGausS and PUP 3D-GS are cited; a one-sentence contrast with post-hoc pruning methods (that they act after a fixed-budget MLE fit, whereas DP-Splat builds parsimony into the prior) is already present but could be moved earlier for readers coming from the engineering compression line.
  6. [§4, Eq. (7)] Appendix E is thorough and useful; consider pointing to it more explicitly from Eq. (7) in the main text for readers who want the product-of-Student-t conditioning subtlety.

Circularity Check

0 steps flagged

No significant circularity: claims rest on classical conjugacy, stick-breaking, and external asymptotics, with empirical results not forced by definition or self-citation.

full rationale

The derivation chain is self-contained against external classical results and does not reduce predictions to fitted inputs by construction. Proposition 1 is standard conjugate mean-field CAVI monotonicity (Bishop §10.2). Proposition 2 derives a rigorous finite-α truncation bound from Ishwaran & James (2001) and corrects their large-α exponential approximation with an elementary inequality; the bound is not fitted to data. The predictive color moments (Eq. 7, Appendix E) follow from NIW conjugacy → Student-t (Murphy 2012) plus mean-field factorization and the law of total variance; they are verified against independent numerical integration, not against a fitted target. ˆK recovery on synthetics is explicitly conditioned on regime-appropriate fixed α and is reported as failing under a single α or under variational α-learning (mode-following); that is an honest empirical qualification, not a self-definitional loop. The deconfounded image comparison (Table 1) isolates complexity selection by tying against a converged fixed-K Dirichlet baseline at the same ˆK, so the +2.7 dB gap over single-pass VBGS is not re-labeled as a prior efficiency gain. Scene results use a priori α=100 and a rasterizer-free held-out point-color metric. Citations to Rousseau–Mengersen, Miller–Harrison, Nguyen, Blei–Jordan, Hoffman et al. are external classical literature; VBGS (Van de Maele et al.) is the baseline being modified, not a self-authored uniqueness theorem. Cross-code regression against the VBGS reference is verification, not load-bearing circular support. No step reduces a claimed prediction to its own inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper inherits the entire conjugate NIW + Categorical + Beta machinery and the stick-breaking construction from the classical literature; its own free choices are the concentration/truncation hyper-parameters, the effective-component threshold, and the decision to treat rendering as a pure deployment map. No new physical entities are postulated. The load-bearing modeling assumptions are the mean-field factorization and the sufficiency of the colored-point mixture likelihood for complexity control.

free parameters (6)
  • DP concentration α (or Gamma hyper-parameters a0,b0 when learned)
    Controls expected number of components; must be chosen in a regime-appropriate range (α≈0.1 for K=3, α≈1 for K=10, α≈5 for K=30) or the recovery claim fails; learning α by VB simply tracks the local mode.
  • truncation level T
    Inference-time Gaussian budget; held fixed at 3Ktrue or 2000 in experiments; theory shows truncation error is negligible once T≳2.5 ˆK but local optima appear when T=Ktrue.
  • sparse Dirichlet concentration e0
    Set ≪1 (0.1, 0.01, 0.001) to invoke Rousseau–Mengersen emptying; empirically ˆK saturates T independently of the precise e0 under CAVI.
  • effective-component threshold nmin
    ˆK = #{k : Nk > nmin}; default 1; sensitivity checked on {0.5,1,2,5} and does not change conclusions, but remains a free reporting choice.
  • NIW hyper-parameters (κ0=1e-3, ν0=Dm+2, Ψ0 from data variance)
    Default conjugate settings inherited from VBGS; affect posterior predictive scale and therefore calibration numbers.
  • SVI schedule (τ0=64, κsched=0.7, |B|=2^16)
    Controls stochastic natural-gradient mixing; full-batch limit recovers CAVI but finite-batch trajectories depend on these knobs.
axioms (5)
  • standard math Conjugacy of Normal–Inverse–Wishart, Beta, Categorical and Gamma families under the mean-field variational family yields exact coordinate-ascent maximizers of the ELBO.
    Invoked for Proposition 1 and all CAVI updates in §4; textbook (Bishop 2006).
  • standard math The truncated stick-breaking construction with vT:=1 approximates the infinite DP with L1 error bounded by 2N(α/(1+α))^{T−1}.
    Proposition 2; first inequality from Ishwaran & James (2001), second derived in the paper.
  • domain assumption Mean-field factorization q(π)∏k q(θk,s)q(θk,c) is an adequate variational family for the claims about ˆK, calibration and predictive moments.
    Stated in §3–§4; implies the product-of-Students predictive and the absence of within-component spatial–color dependence.
  • domain assumption The mixture density on colored points (RGB-D or depth-lifted) is the correct likelihood for complexity control; rasterized rendering is only a deployment map.
    Explicitly scoped in §1, §3, §5 and §7; all theory and most metrics are point-level.
  • ad hoc to paper Component death without revival under plain CAVI is an acceptable dynamics for pragmatic complexity control.
    Acknowledged in §6.4 and §7 as the driver of the variational–asymptotic inversion; the paper deliberately does not add merge/split or reassignment heuristics.

pith-pipeline@v1.1.0-grok45 · 23443 in / 3613 out tokens · 36870 ms · 2026-07-14T08:20:43.993126+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of DP-Splat: Bayesian Nonparametric Complexity Control for Gaussian Splatting." pith.science (2026). https://pith.science/paper/ANU25AQW

@misc{pith2026260710912,
  author       = {Pith},
  title        = {Pith review of: DP-Splat: Bayesian Nonparametric Complexity Control for Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ANU25AQW}},
  note         = {Machine review of arXiv:2607.10912}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussian Splatting represents scenes as finite mixtures of anisotropic Gaussians whose number of components $K$ is set by heuristic density control or user caps. Variational Bayes Gaussian Splatting (VBGS) recast splat fitting as conjugate variational inference, but $K$ remains fixed. We replace the finite symmetric Dirichlet over mixture weights with a truncated stick-breaking Dirichlet-process prior -- and, as a theory-backed alternative, a sparse overfitted finite Dirichlet -- so that the number of occupied components adapts to the data while every update remains a closed-form coordinate-ascent step; a natural-gradient stochastic variant makes the per-step cost independent of the number of points. We give an exact monotonicity guarantee, a rigorous truncation-error bound correcting an anti-conservative large-$\alpha$ approximation in common use, and an honest account of what the fitted number of components estimates. Empirically: (i) the effective complexity $\hat{K}$ adapts to scene complexity and recovers the true $K$ within $\pm 1$ on well-separated synthetic data with regime-appropriate concentration; (ii) a deconfounded comparison shows the DP prior's contribution is complexity selection, not per-component efficiency -- converged DP fits exceed single-pass fixed-$K$ VBGS by +2.7 dB at matched budgets yet tie an equally converged fixed-$K$ baseline, and on 3D scenes DP-Splat matches or exceeds VBGS's held-out color prediction with 5.9-7.6x fewer components; (iii) the posterior-predictive color variance is well calibrated on model-matched synthetic data; and (iv) the ordering suggested by exact-posterior asymptotics reverses under mean-field coordinate ascent: the DP prior resists over-splitting while the sparse finite mixture saturates its truncation, a gap between variational practice and posterior asymptotics documented across three orders of magnitude in $N$.

Figures

Figures reproduced from arXiv: 2607.10912 by Aqi Dong.

Figure 1
Figure 1. Figure 1: ELBO traces, 20 seeds per panel. All traces are monotone (worst relative decrease [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Kb (nmin=1) vs. N for Ktrue ∈ {3, 10, 30} (dotted line), T = 3Ktrue. Bars: seed min/max (n=3); series are offset horizontally for visibility. dp(learn) coincides with dp(α=1) and sparse_dir with dir (that coincidence is itself a finding; see text). 6.2 Kb recovery is real but concentration-regime-dependent ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Truncation sweep. Left/middle: held-out log-likelihood and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Kb growth with N (3D, Ktrue=10, T=60; CAVI for N ≤ 105 , SVI beyond; bars: seed min/max). The sparse finite mixture saturates T by N=106 for all three e0; the DP variants grow slowly. This inverts the naive reading of the asymptotic theory (see text). mixtures (Ktrue=10, T=60, N from 103 to 106 ) we observe the opposite ordering: dp(α=1) grows slowly (10 → 13.3 over three decades of N, consistent with log-… view at source ↗
Figure 5
Figure 5. Figure 5: Predictive color-variance calibration: binned realized squared error vs. binned predicted variance [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: W2(fitted, true mixing measure) vs. N (three seeds; atoms in joint spatial–color mean space). Context: contraction theory for exact posteriors is given by Nguyen (2013); no claim is made that CAVI attains those rates [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scene fitting (lego). Left: effective components during continual fitting (matched budget T=2000): VBGS pins ∼1374 components of its budget; DP-Splat self-selects ∼234 (α=100, starting from ∼850 and shedding redundant components) or ∼20 (α=1). Middle: held-out-view predictive color variance— uncertainty concentrates on edges and occlusion boundaries. Right: held-out E[c|s] (denormalized). carries no statis… view at source ↗
Figure 8
Figure 8. Figure 8: Scene fitting (chair), companion to [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 1 canonical work pages

  1. [1]

    , title =

    Ferguson, Thomas S. , title =. The Annals of Statistics , year =

  2. [2]

    Statistica Sinica , year =

    Sethuraman, Jayaram , title =. Statistica Sinica , year =

  3. [3]

    , title =

    Ishwaran, Hemant and James, Lancelot F. , title =. Journal of the American Statistical Association , year =

  4. [4]

    and Jordan, Michael I

    Blei, David M. and Jordan, Michael I. , title =. Bayesian Analysis , year =

  5. [5]

    and Harrison, Matthew T

    Miller, Jeffrey W. and Harrison, Matthew T. , title =. Advances in Neural Information Processing Systems , volume =. 2013 , publisher =

    2013

  6. [6]

    and Harrison, Matthew T

    Miller, Jeffrey W. and Harrison, Matthew T. , title =. Journal of Machine Learning Research , year =

  7. [7]

    and Harrison, Matthew T

    Miller, Jeffrey W. and Harrison, Matthew T. , title =. Journal of the American Statistical Association , year =

  8. [8]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

    Rousseau, Judith and Mengersen, Kerrie , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

  9. [9]

    Model-Based Clustering Based on Sparse Finite

    Malsiner-Walli, Gertraud and Fr. Model-Based Clustering Based on Sparse Finite. Statistics and Computing , year =

  10. [10]

    Generalized Mixtures of Finite Mixtures and Telescoping Sampling , journal =

    Fr. Generalized Mixtures of Finite Mixtures and Telescoping Sampling , journal =. 2021 , volume =

    2021

  11. [11]

    The Annals of Statistics , year =

    Nguyen, XuanLong , title =. The Annals of Statistics , year =

  12. [12]

    Proceedings of the Twentieth International Joint Conference on Artificial Intelligence (

    Kurihara, Kenichi and Welling, Max and Teh, Yee Whye , title =. Proceedings of the Twentieth International Joint Conference on Artificial Intelligence (. 2007 , pages =

    2007

  13. [13]

    , title =

    Bishop, Christopher M. , title =. 2006 , series =

    2006

  14. [14]

    ACM Transactions on Graphics , volume =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. ACM Transactions on Graphics , volume =. 2023 , doi =

    2023

  15. [16]

    Advances in Neural Information Processing Systems , volume =

    Kheradmand, Shakiba and Rebain, Daniel and Sharma, Gopal and Sun, Weiwei and Tseng, Yang-Che and Isack, Hossam and Kar, Abhishek and Tagliasacchi, Andrea and Yi, Kwang Moo , title =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  16. [19]

    and Tancik, Matthew and Barron, Jonathan T

    Mildenhall, Ben and Srinivasan, Pratul P. and Tancik, Matthew and Barron, Jonathan T. and Ramamoorthi, Ravi and Ng, Ren , title =. Computer Vision -- ECCV 2020 , editor =. 2020 , doi =

    2020

  17. [20]

    Advances in Neural Information Processing Systems , volume =

    Fan, Zhiwen and Wang, Kevin and Wen, Kairun and Zhu, Zehao and Xu, Dejia and Wang, Zhangyang , title =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  18. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Lee, Joo Chan and Rho, Daniel and Sun, Xiangyu and Ko, Jong Hwan and Park, Eunbyung , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  19. [22]

    and Blei, David M

    Hoffman, Matthew D. and Blei, David M. and Wang, Chong and Paisley, John , title =. Journal of Machine Learning Research , year =

  20. [23]

    and Kucukelbir, Alp and McAuliffe, Jon D

    Blei, David M. and Kucukelbir, Alp and McAuliffe, Jon D. , title =. Journal of the American Statistical Association , year =

  21. [24]

    , title =

    Murphy, Kevin P. , title =. 2012 , isbn =

    2012

  22. [25]

    , title =

    Ghahramani, Zoubin and Beal, Matthew J. , title =. Advances in Neural Information Processing Systems 12 , publisher =. 2000 , pages =

    2000

  23. [26]

    Neural Computation , year =

    Amari, Shun-ichi , title =. Neural Computation , year =

  24. [27]

    Proceedings of the 35th International Conference on Machine Learning , series =

    Accurate Uncertainties for Deep Learning Using Calibrated Regression , author =. Proceedings of the 35th International Conference on Machine Learning , series =

  25. [28]

    Jiang, Wen and Lei, Boshu and Daniilidis, Kostas , booktitle =

  26. [29]

    Hanson, Alex and Tu, Allen and Singla, Vasu and Jayawardhana, Mayuka and Zwicker, Matthias and Goldstein, Tom , booktitle =

  27. [30]

    Le, Ya and Yang, Xuan , institution=. Tiny

  28. [31]

    On the Multivariate t Distribution , author=

  29. [32]

    The American Statistician , volume=

    On the Conditional Distribution of the Multivariate t Distribution , author=. The American Statistician , volume=

  30. [33]

    Natural gradient works efficiently in learning

    Shun-ichi Amari. Natural gradient works efficiently in learning. Neural Computation, 10 0 (2): 0 251--276, 1998. doi:10.1162/089976698300017746

    doi:10.1162/089976698300017746 1998

  31. [34]

    Christopher M. Bishop. Pattern Recognition and Machine Learning. Information Science and Statistics. Springer, New York, 2006. ISBN 978-0-387-31073-2

    2006

  32. [35]

    Blei and Michael I

    David M. Blei and Michael I. Jordan. Variational inference for D irichlet process mixtures. Bayesian Analysis, 1 0 (1): 0 121--143, 2006. doi:10.1214/06-BA104

    doi:10.1214/06-ba104 2006

  33. [36]

    Blei, Alp Kucukelbir, and Jon D

    David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe. Variational inference: A review for statisticians. Journal of the American Statistical Association, 112 0 (518): 0 859--877, 2017. doi:10.1080/01621459.2017.1285773

    doi:10.1080/01621459.2017.1285773 2017

  34. [37]

    On the conditional distribution of the multivariate t distribution

    Peng Ding. On the conditional distribution of the multivariate t distribution. The American Statistician, 70 0 (3): 0 293--295, 2016

    2016

  35. [38]

    LightGaussian : Unbounded 3D G aussian compression with 15x reduction and 200+ FPS

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian : Unbounded 3D G aussian compression with 15x reduction and 200+ FPS . In Advances in Neural Information Processing Systems, volume 37, pp.\ 140138--140158. Curran Associates, Inc., 2024. doi:10.52202/079017-4447

    doi:10.52202/079017-4447 2024

  36. [39]

    Ferguson

    Thomas S. Ferguson. A B ayesian analysis of some nonparametric problems. The Annals of Statistics, 1 0 (2): 0 209--230, 1973. doi:10.1214/aos/1176342360

    doi:10.1214/aos/1176342360 1973

  37. [40]

    u hwirth-Schnatter, Gertraud Malsiner-Walli, and Bettina Gr \

    Sylvia Fr \"u hwirth-Schnatter, Gertraud Malsiner-Walli, and Bettina Gr \"u n. Generalized mixtures of finite mixtures and telescoping sampling. Bayesian Analysis, 16 0 (4): 0 1279--1307, 2021. doi:10.1214/21-BA1294

    doi:10.1214/21-ba1294 2021

  38. [41]

    Zoubin Ghahramani and Matthew J. Beal. Variational inference for B ayesian mixtures of factor analysers. In Advances in Neural Information Processing Systems 12, pp.\ 449--455. MIT Press, 2000

    2000

  39. [42]

    PUP 3D-GS : Principled uncertainty pruning for 3D G aussian splatting

    Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. PUP 3D-GS : Principled uncertainty pruning for 3D G aussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5949--5958, 2025

    2025

  40. [43]

    Hoffman, David M

    Matthew D. Hoffman, David M. Blei, Chong Wang, and John Paisley. Stochastic variational inference. Journal of Machine Learning Research, 14 0 (40): 0 1303--1347, 2013

    2013

  41. [44]

    Hemant Ishwaran and Lancelot F. James. Gibbs sampling methods for stick-breaking priors. Journal of the American Statistical Association, 96 0 (453): 0 161--173, 2001. doi:10.1198/016214501750332758

    doi:10.1198/016214501750332758 2001

  42. [45]

    FisherRF : Active view selection and mapping with radiance fields using F isher information

    Wen Jiang, Boshu Lei, and Kostas Daniilidis. FisherRF : Active view selection and mapping with radiance fields using F isher information. In Computer Vision -- ECCV 2024, Lecture Notes in Computer Science, pp.\ 422--440. Springer, 2024

    2024

  43. [46]

    3D G aussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk \"u hler, and George Drettakis. 3D G aussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 0 (4): 0 1--14, July 2023. doi:10.1145/3592433

    doi:10.1145/3592433 2023

  44. [47]

    3D G aussian splatting as M arkov chain M onte C arlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D G aussian splatting as M arkov chain M onte C arlo. In Advances in Neural Information Processing Systems, volume 37, pp.\ 80965--80986. Curran Associates, Inc., 2024. doi:10.52202/079017-2573

    doi:10.52202/079017-2573 2024

  45. [48]

    Accurate uncertainties for deep learning using calibrated regression

    Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp.\ 2796--2804. PMLR, 2018

    2018

  46. [49]

    Collapsed variational D irichlet process mixture models

    Kenichi Kurihara, Max Welling, and Yee Whye Teh. Collapsed variational D irichlet process mixture models. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence ( IJCAI -07) , pp.\ 2796--2801, 2007

    2007

  47. [50]

    Tiny ImageNet visual recognition challenge

    Ya Le and Xuan Yang. Tiny ImageNet visual recognition challenge. Technical report, Stanford University, CS231N course report, 2015

    2015

  48. [51]

    Compact 3D G aussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3D G aussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 21719--21728, 2024

    2024

  49. [52]

    u hwirth-Schnatter, and Bettina Gr \

    Gertraud Malsiner-Walli, Sylvia Fr \"u hwirth-Schnatter, and Bettina Gr \"u n. Model-based clustering based on sparse finite G aussian mixtures. Statistics and Computing, 26 0 (1-2): 0 303--324, 2016. doi:10.1007/s11222-014-9500-2

    doi:10.1007/s11222-014-9500-2 2016

  50. [53]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF : Representing scenes as neural radiance fields for view synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm (eds.), Computer Vision -- ECCV 2020, volume 12346 of Lecture Notes in Computer Science, pp.\ 405--421. Springe...

    doi:10.1007/978-3-030-58452-8_24 2020

  51. [54]

    Miller and Matthew T

    Jeffrey W. Miller and Matthew T. Harrison. A simple example of D irichlet process mixture inconsistency for the number of components. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc., 2013

    2013

  52. [55]

    Miller and Matthew T

    Jeffrey W. Miller and Matthew T. Harrison. Inconsistency of P itman-- Y or process mixtures for the number of components. Journal of Machine Learning Research, 15: 0 3333--3370, 2014

    2014

  53. [56]

    Miller and Matthew T

    Jeffrey W. Miller and Matthew T. Harrison. Mixture models with a prior on the number of components. Journal of the American Statistical Association, 113 0 (521): 0 340--356, 2018. doi:10.1080/01621459.2016.1255636

    doi:10.1080/01621459.2016.1255636 2018

  54. [57]

    Kevin P. Murphy. Machine Learning: A Probabilistic Perspective. MIT Press, Cambridge, MA, 2012. ISBN 978-0-262-01802-9

    2012

  55. [58]

    Convergence of latent mixing measures in finite and infinite mixture models

    XuanLong Nguyen. Convergence of latent mixing measures in finite and infinite mixture models. The Annals of Statistics, 41 0 (1): 0 370--400, 2013. doi:10.1214/12-AOS1065

    work page doi:10.1214/12-aos1065 2013

  56. [59]

    BOGausS : Better optimized G aussian splatting

    St \'e phane Pateux, Matthieu Gendrin, Luce Morin, Th \'e o Ladune, and Xiaoran Jiang. BOGausS : Better optimized G aussian splatting. arXiv preprint arXiv:2504.01844, 2025

    Pith/arXiv arXiv 2025

  57. [60]

    On the multivariate t distribution

    Michael Roth. On the multivariate t distribution. Technical Report LiTH-ISY-R-3059, Link \"o ping University, Division of Automatic Control, 2013

    2013

  58. [61]

    Asymptotic behaviour of the posterior distribution in overfitted mixture models

    Judith Rousseau and Kerrie Mengersen. Asymptotic behaviour of the posterior distribution in overfitted mixture models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73 0 (5): 0 689--710, 2011. doi:10.1111/j.1467-9868.2011.00781.x

    doi:10.1111/j.1467-9868.2011.00781.x 2011

  59. [62]

    A constructive definition of D irichlet priors

    Jayaram Sethuraman. A constructive definition of D irichlet priors. Statistica Sinica, 4 0 (2): 0 639--650, 1994

    1994

  60. [63]

    Buckley, and Tim Verbelen

    Toon Van de Maele, Ozan C atal, Alexander Tschantz, Christopher L. Buckley, and Tim Verbelen. Variational B ayes G aussian splatting. arXiv preprint arXiv:2410.03592, 2024

    arXiv 2024

  61. [64]

    VBGS-SLAM : Variational B ayesian G aussian splatting simultaneous localization and mapping

    Yuhan Zhu, Yanyu Zhang, Jie Xu, and Wei Ren. VBGS-SLAM : Variational B ayesian G aussian splatting simultaneous localization and mapping. arXiv preprint arXiv:2604.02696, 2026

    Pith/arXiv arXiv 2026