Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Paradise of Forking Paths: Revisiting the Adaptive Data Analysis Problem

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that when an analyst selects counting queries by a Pólya-tree posterior utility, the resulting adaptive distribution estimate has lower mean squared error than a non-adaptive histogram with the same query budget and does…

desk verdict A coherent constructive-ADA framework built on Pólya trees, with a standard conjugacy theorem and a plausible greedy utility, but the empirical comparison does not isolate adaptivity from Bayesian smoothing. read the letter →

arxiv 2501.11804 v1 pith:ZQIF3K3Q submitted 2025-01-21 stat.ME

Amir Hossein Hadavi , Mohammad M. Mojahedian , Mohammad Reza Aref This is my paper · ORCID
classification stat.ME MSC 62G0762C1062L05
keywords adaptivedataanalysisPólyatreesBayesiannonparametricsdistributionestimationcountingqueriesqueryselectionsequentialupdatingoverfitting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper revisits the adaptive data analysis (ADA) problem, where an analyst queries a dataset round by round, and asks whether adaptivity can be a blessing rather than a curse when the analyst's goal is inference rather than adversarial query answering. It claims that when the analyst holds a Pólya-tree prior over the unknown distribution and selects each new counting query to maximise the expected reduction in mean squared error of the resulting density estimate, the estimate beats a non-adaptive histogram using the same number of queries and, unlike the histogram, does not overfit after many rounds. The argument is made concrete through a closed-form utility function and a conjugacy theorem for finite Pólya trees that lets the analyst update beliefs with simple additions to the prior's alpha and beta parameters. The paper's significance is that it offers a principled and interpretable way to convert subjective prior knowledge into an adaptive query strategy, with implications for human-in-the-loop data analysis and for understanding when adaptivity helps rather than hurts.

What carries the argument

The load-bearing mechanism is a finite Pólya tree (FPT) prior over densities, built on a dyadic partition of the domain, together with the utility function $u(I)$ that assigns each leaf of the current analysis tree a score reflecting how much branching it would reduce the MSE of the piecewise-constant estimate. The conjugacy theorem (Theorem 1) and its corollary ensure that a single counting answer $n_L$ suffices to update the parent hyperparameters, making the algorithm computationally light. The two approximations that keep the utility tractable are the replacement of ancestor mass by its posterior mean and the plug-in of the expected count $\rho n$ for the unseen child count; these enter the derivation of the closed-form score.

What would settle it

Run the algorithm on a distribution where $n$ is small (e.g., $n=20$) and the prior is broad, so the posterior variance of the ancestor masses is not negligible, and compare the MSE of the adaptive method against a non-adaptive histogram; if the adaptive estimate no longer beats the histogram, the claimed benefit depends on the neglected variance. Alternatively, compare the algorithm's query rankings with an exhaustive search of actual MSE reductions for a large-$n$ dataset: if the rankings reverse whenever $\mathrm{Var}[M_b]$ is sizable, the utility function misranks informative queries.

Watch

Extended reading notes

Core claim

The central discovery is that, under the proposed non-parametric Bayesian model, the optimal next query is the one that maximizes a simple utility score $u(I) = 2^{l_I} n_I \nu_I^2 \rho_I (1-\rho_I) / ((1+\eta_I)(1+\eta_I+n_I))$ for a candidate interval $I$ in the current analysis tree. This score is derived from the expected decrease in the mean squared error of the piecewise-constant density estimate if the analyst branches that interval, after approximating the posterior variance of ancestor masses as negligible. The paper proves that finite Pólya trees are conjugate to counting queries, so the entire adaptive procedure reduces to maintaining an analysis tree and updating the hyperparameters at the queried node via $(\alpha + n_L, \beta + n_R)$. Simulations with Gaussian mixtures show that this adaptive estimate achieves lower MSE and avoids the overfitting behaviour of a standard histogram at large query counts.

Load-bearing premise

The selection rule assumes that after each update the analyst's uncertainty about already-explored parts of the distribution is negligible, so the true count in a child can be replaced by its prior expected value without changing which query is most useful; if this fails, the chosen queries may be far from optimal.

Editorial extensions

If this is right

  • The adaptive query-selection rule attains a given density-estimation accuracy with fewer counting queries than a fixed histogram, so analysts can get the same quality estimate from a smaller interaction budget.
  • Because the procedure updates a coherent posterior, the MSE stays roughly flat as the query count grows, while a non-adaptive histogram eventually overfits the empirical partition.
  • Counting answers suffice for exact Bayesian updating of finite Pólya trees: after querying the left child of a leaf, the analyst only needs to add the observed count to $\alpha$ and the complement to $\beta$ of that leaf's prior.
  • The utility score $u(I)$ provides an interpretable ranking of candidate queries, linking each candidate to the analyst's prior mean, confidence, current mass estimate, and sample size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the false premise about negligible ancestor variance is the weak spot, a natural fix would be to add a variance-penalty term to $u(I)$ that accounts for the residual uncertainty in $M_b$; a simulation study with small $n$ could test whether this preserves the advantage.
  • The utility score's dependence on $\rho(1-\rho)$ suggests a deeper connection to information-theoretic query selection; comparing this rule to an entropy-reduction criterion could reveal whether the two coincide in ranking.
  • The constructive framing suggests a broader conjecture: for inference-oriented analysts with well-calibrated priors, adaptivity reduces sample complexity or query complexity under general losses, not just MSE; this could be tested in a minimax or regret framework.
  • The model's alignment with cognitive science invites an experiment where human analysts' next queries are compared to the algorithm's choices, testing whether the Pólya-tree rule describes actual human adaptive belief updating.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a revised Adaptive Data Analysis model in which a Bayesian analyst sequentially selects counting queries, guided by a Pólya-tree prior, in order to estimate an unknown distribution with as few queries as possible. The main theoretical contribution is a conjugacy result for finite Pólya trees (Theorem 1) and a derived utility function u(I) (Eq. 16) used in Algorithm 1 to choose which node to split next. The authors argue that this constructive, inference-oriented view of ADA contrasts with the adversarial query-answering literature. Simulations compare the proposed adaptive Pólya-tree estimator ('junc-ADA') with a non-adaptive histogram ('NADA') and with a variant using midpoint junctions ('mid-ADA') in terms of mean squared error and total variation distance, and report favorable performance for small and medium query budgets.

Significance. If the central claim were established, the paper would offer a useful counterpoint to the worst-case ADA literature by showing that adaptivity can be constructive for inference when the analyst has a reasonable prior. The algorithmic structure is interpretable, and the derivations behind the utility function are clearly laid out. However, the significance is currently limited by two load-bearing gaps: the utility derivation relies on unquantified approximations, and the empirical comparison does not isolate adaptivity from the choice of estimator. The paper does not provide code or machine-checked proofs, but Algorithm 1 is sufficiently specified to be reimplemented. Overall, the idea is promising, but the evidence presented does not yet support the headline conclusion that adaptivity itself reduces the number of queries needed.

major comments (3)
  1. [IV.B.5, Eqs. (9)-(16)] The utility u(I) that drives Algorithm 1 is derived using two approximations: first, the posterior mass of ancestor nodes is treated as known, Var[M_b] ≈ 0, in Eq. (9); second, the unknown child count n_I is replaced by its prior mean, \hat n_I = ρ_{par(I)} n_{par(I)}, in Eqs. (13)-(14). No error bounds are given for either approximation, and no sensitivity analysis shows that the ranking of candidate nodes is stable when these approximations break down. Because the selected queries are exactly those maximizing u(I), the claim that Algorithm 1 chooses near-optimal informative queries is not supported without further analysis.
  2. [V.B.2, Figs. 7-10] The main empirical claim that junc-ADA outperforms a non-adaptive approach at the same query budget is confounded: junc-ADA and NADA differ in both the query-selection rule and the estimator. NADA is a classical histogram with no shrinkage, while junc-ADA is a smoothed Pólya-tree posterior whose hyperparameters in Eq. (20) encode the known smoothness scale σlow of the simulation. The reported MSE/TV gains could therefore be produced by Bayesian smoothing alone, with no adaptive querying. mid-ADA is not a non-adaptive control because it uses the same adaptive partition. A non-adaptive Bayesian Pólya-tree estimator with identical hyperparameters and the same number of empirical counts is required to separate the effect of adaptivity from the effect of the estimator. Additionally, the binning specification for NADA is not given, so the baseline is not pinned down.
  3. [V.A.2, Eq. (20)] The prior hyperparameters are chosen using the true smoothness scale σlow of the simulation through lmax = ⌈log2(xlim/σlow)⌉, and the simulations assume equal split probabilities ρ = 1/2. The paper does not report how performance degrades when the prior is misspecified. Since the method is motivated by the assumption that the analyst has a reasonably calibrated prior, a misspecification or sensitivity experiment is needed to judge the robustness of the claimed advantages.
minor comments (6)
  1. [I] The sentence 'We assume that A progresses in a Bayesian manner within a non-parametric structure' appears twice in the introduction; one occurrence should be removed.
  2. [V.A.2, Eq. (20)] The condition in Eq. (20) uses 'x ≤ lmax' and 'x > lmax', but x is the domain coordinate and l is the tree level; this should be 'l ≤ lmax' and 'l > lmax'.
  3. [IV.D, proof of Theorem 1] The proof refers to a step '(g)' in Eq. (19), but no displayed equation is labeled (g); either label the corresponding line or remove the reference.
  4. [IV.B.5, after Eq. (15)] The sentence 'The calculation of (15) depends entirely on par(I), rather than on I itself' is confusing because Eq. (16) subsequently redefines I to be the parent node; please clarify the notation.
  5. [V.B.1] There is a typo in the sample size notation: 'n = 10 4' should read 'n = 10^4'.
  6. [V.B.2] The text says 'Figs. 8 and 9 reveal that junc-ADA excels NADA in total variation distance', but Fig. 8 displays MSE and Fig. 9 displays TV distance; the cross-reference should be checked.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the adaptive query-selection utility is derived from a Bayesian MSE decomposition with explicit approximations, and the empirical evaluation is an external simulation comparison.

full rationale

The paper's central claim, that Algorithm 1 with utility (16) estimates P with lower MSE than a non-adaptive histogram, is not equivalent to its inputs by construction. The utility u(I) = 2^{l_I} n_I ν_I^2 ρ_I(1-ρ_I)/((1+η_I)(1+η_I+n_I)) is derived in Eqs. (8)-(16) from a mean-squared-error decomposition of a finite Pólya tree, using two explicitly stated approximations: Var[M_b]≈0 after ancestral updates (Eq. (9)) and replacement of the unknown count n_I by its prior mean ρ n (Eq. (14)). These are modeling approximations, not circular definitions: the ranking is computed from the current prior hyperparameters, and the actual empirical answer n_IL from the dataset is used to update the posterior in Algorithm 1 lines 7-8. The hyperparameters in Eq. (20), including 0.4 and 1.02, are declared heuristic prior choices rather than fitted to the evaluation data, so no fitted input is relabeled as a prediction. The only self-citation, [27], is cited in the introduction as an example of physical ADA considerations and plays no role in the derivation or in the empirical comparison. The conjugacy result (Theorem 1) is proved in the paper via induction and is standard PT updating; no uniqueness theorem is imported from the authors' prior work. The strongest skeptical objection is that the junc-ADA vs NADA comparison varies both the query rule and the estimator, but that is a question of experimental isolation rather than circularity: the gains are measured against an external Gaussian-mixture simulation, and the NADA baseline is an ordinary histogram rather than a construct defined by the paper's own fitted values. Therefore no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central algorithm relies on four hand-chosen numerical parameters (0.4, 1.02, lmax, ρ=0.5) and several approximations that are not proven. The strongest assumptions are the analyst's access to a correct minimum peak width and the posterior-concentration approximation used in the utility derivation. No new physical or latent entities are introduced.

free parameters (4)
  • η prior concentration scale (0.4) = 0.4
    Eq. (20) multiplies the depth-squared term by 0.4; it sets how strongly the prior concentrates with tree depth and is chosen heuristically without sensitivity analysis.
  • η priority growth factor (1.02) = 1.02
    Eq. (20) uses (1.02)^i to favor candidate nodes added earlier; the value is heuristic and affects which nodes are queried when several are candidates.
  • Finest resolution depth lmax = 8 in the reported simulations
    lmax = ceil(log2(xlim/σlow)) encodes the analyst's belief about the minimum peak width; in simulations this is aligned with the true σlow, giving the method extra information the non-adaptive baseline does not get.
  • Equal split prior ρ_I = 0.5
    Section V.A.2 assumes left and right children are equally likely under the prior; this is a hand-chosen default, not a fitted value.
assumptions (6)
  • standard math Pólya tree construction with independent Beta splits generates valid random densities on the bounded interval.
    Used throughout Section II.C; properties are cited from references [34]-[37].
  • domain assumption A human analyst can express subjective beliefs as (ρ, η) hyperparameters in a Pólya tree.
    Section IV.A assumes prior elicitation is feasible; the paper provides an intuitive procedure but no validation that humans can produce reliable (ρ, η) values.
  • domain assumption The analyst's prior includes a correct lower bound on the width of peaks, σlow.
    Section V.A.2 sets lmax using σlow and states the analyst's knowledge is aligned with σlow; this is a strong, possibly oracle-like assumption.
  • ad hoc to paper Posterior concentration approximation Var[M_b] ≈ 0 for ancestor masses.
    Section IV.B.5, after Eq. (9), uses ν_b ≈ M_b to simplify the expected loss; no bound is given on this approximation error.
  • ad hoc to paper The unknown future query count can be replaced by its prior mean ρ n.
    Near Eqs. (13)-(14), n̂_I = ρ_{par(I)} n_{par(I)} is plugged in to keep the posterior mean ρ unchanged; this makes the utility myopic and self-referential.
  • domain assumption The residual MSE outside the current finite analysis tree can be ignored in each round.
    Section IV.B.5 says reducing the residual is not the direct concern of the current round; this justifies optimizing only in the current tree.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Paradise of Forking Paths: Revisiting the Adaptive Data Analysis Problem." pith.science (2026). https://pith.science/paper/ZQIF3K3Q

@misc{pith2026250111804,
  author       = {Pith},
  title        = {Pith review of: Paradise of Forking Paths: Revisiting the Adaptive Data Analysis Problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQIF3K3Q}},
  note         = {Machine review of arXiv:2501.11804}
}
read the original abstract

The Adaptive Data Analysis (ADA) problem, where an analyst interacts with a dataset through statistical queries, is often studied under the assumption of adversarial analyst behavior. To decrease this gap, we propose a revised model of ADA that accounts for more constructive interactions between the analysts and the data, where the goal is to enhance inference accuracy. Specifically, we focus on distribution estimation as a central objective guiding analyst's queries. The problem is addressed within a non-parametric Bayesian framework, capturing the flexibility and dynamic evolution of analyst's beliefs. Our hierarchical approach leverages P\'olya trees (PTs) as priors over the distribution space, facilitating the adaptive selection of counting queries to efficiently reduce the estimation error without increasing the number of queries. Furthermore, with its interpretability and conjugacy, the proposed framework allows for intuitive conversion of subjective beliefs into objective priors and their effortless updates to posteriors. Using theoretical derivations, we formalize the PT-based solution as a computational algorithm. Simulations further demonstrate its effectiveness in distribution estimation tasks compared to the non-adaptive approach. By aligning with real-world applications, this structured ADA framework fosters opportunities for collaborative research in related areas, such as human-in-the-loop systems and cognitive studies of belief updating.

Figures

Figures reproduced from arXiv: 2501.11804 by the authors.

Figure 1
Figure 1. Random probability mass allocation by Pólya Tree. The root and two first layers are shown. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Constructive interaction between the responder and the analyst aims to achieve a common goal (in this case, estimating the distribution [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The overall flow of our proposed solution scheme can be illustrated as a five-step iterative process in each round. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of the performance of NADA, mid-ADA, and junc-ADA in approximating the true density for [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Comparison of the performance of NADA, mid-ADA, and junc-ADA in approximating the true density for [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the performance of NADA, mid-ADA, and junc-ADA in approximating the true density for [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Comparing mean of MSE of trials. 0 500 1000 1500 2000 2500 Number of queries 0.00000 0.00025 0.00050 0.00075 0.00100 0.00125 0.00150 0.00175 MSE Median on trials for NADA Interquartile range (5th-95th) Median on trials for junc-ADA Interquartile Range (5th-95th) [PITH…
Figure 8
Figure 8. Figure 8: Comparing median of MSE of trials. method is designed. By this metric, the method is encouraged to give priority to detecting the peaks. But what about the overall shape? Interestingly, Figs. 8 and 9 reveal that junc-ADA excels NADA in total variation (TV) distance for…
Figure 9
Figure 9. Figure 9: Comparing mean of TV error of trials. 0 500 1000 1500 2000 2500 Number of queries 0.000 0.005 0.010 0.015 0.020 0.025 Total variation distance Median on trials for NADA Interquartile range (5th-95th) Median on trials for junc-ADA Interquartile Range (5th-95th) [PITH_F…
Figure 10
Figure 10. Figure 10: Comparing median of TV error of trials. a divide-and-conquer manner represents a typical adaptive approach, aligning well with the constructive ADA model presented and analyzed in our paper. Our scheme provides an intuitive framework for translating subjective beliefs…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [1]

    The future of data analysis,

    J. W. Tukey, “The future of data analysis,” The annals of mathematical statistics , vol. 33, no. 1, pp. 1–67, 1962

    work page 1962

  2. [2]

    50 years of data science,

    D. Donoho, “50 years of data science,” URL http://courses. csail. mit. edu/18 , vol. 337, p. 2015, 2015

    work page 2015

  3. [3]

    The garden of forking paths: Why multiple comparisons can be a problem, even when there is no “fishing expedition

    A. Gelman and E. Loken, “The garden of forking paths: Why multiple comparisons can be a problem, even when there is no “fishing expedition” or “p-hacking” and the research hypothesis was posited ahead of time,” Department of Statistics, Columbia University , 2013

    work page 2013

  4. [4]

    The statistical crisis in science,

    ——, “The statistical crisis in science,” American scientist, vol. 102, no. 6, p. 460, 2014

    work page 2014

  5. [5]

    Preventing false discovery in interactive data analysis is hard,

    M. Hardt and J. Ullman, “Preventing false discovery in interactive data analysis is hard,” in 2014 IEEE 55th Annual Symposium on Foundations of Computer Science . IEEE, 2014, pp. 454–463

    work page 2014

  6. [6]

    Preserving statistical validity in adaptive data analysis,

    C. Dwork, V . Feldman, M. Hardt, T. Pitassi, O. Reingold, and A. L. Roth, “Preserving statistical validity in adaptive data analysis,” in Proceedings of the forty-seventh annual ACM symposium on Theory of computing . ACM, 2015, pp. 117–126

    work page 2015

  7. [7]

    The algorithmic foundations of differential privacy,

    C. Dwork, A. Roth et al. , “The algorithmic foundations of differential privacy,” Foundations and Trends® in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211–407, 2014

    work page 2014

  8. [8]

    Generalization in adaptive data analysis and holdout reuse,

    C. Dwork, V . Feldman, M. Hardt, T. Pitassi, O. Reingold, and A. Roth, “Generalization in adaptive data analysis and holdout reuse,” in Advances in Neural Information Processing Systems , 2015, pp. 2350–2358

    work page 2015

Show all 43 references
  1. [9]

    A minimax theory for adaptive data analysis,

    Y .-X. Wang, J. Lei, and S. E. Fienberg, “A minimax theory for adaptive data analysis,” arXiv preprint arXiv:1602.04287 , 2016

    2016 arXiv

  2. [10]

    Max-information, differential privacy, and post-selection hypothesis testing,

    R. Rogers, A. Roth, A. Smith, and O. Thakkar, “Max-information, differential privacy, and post-selection hypothesis testing,” in 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS) . IEEE, 2016, pp. 487–494

    2016

  3. [11]

    Typicality-based stability and privacy,

    R. Bassily and Y . Freund, “Typicality-based stability and privacy,” arXiv preprint arXiv:1604.03336 , 2016

    2016 arXiv

  4. [12]

    Generalization for adaptively-chosen estimators via stable median,

    V . Feldman and T. Steinke, “Generalization for adaptively-chosen estimators via stable median,” arXiv preprint arXiv:1706.05069, 2017

    2017 arXiv

  5. [13]

    Calibrating noise to variance in adaptive data analysis,

    ——, “Calibrating noise to variance in adaptive data analysis,” in Conference On Learning Theory , 2018, pp. 535–544

    2018

  6. [14]

    A new approach to adaptive data analysis and learning via maximal leakage,

    A. R. Esposito, M. Gastpar, and I. Issa, “A new approach to adaptive data analysis and learning via maximal leakage,” arXiv preprint arXiv:1903.01777, 2019

    1903 arXiv

  7. [15]

    Sampling without compromising accuracy in adaptive data analysis,

    B. Fish, L. Reyzin, and B. I. Rubinstein, “Sampling without compromising accuracy in adaptive data analysis,” in Algorithmic Learning Theory. PMLR, 2020, pp. 297–318

    2020

  8. [16]

    A bounded-noise mechanism for differential privacy,

    Y . Dagan and G. Kur, “A bounded-noise mechanism for differential privacy,” in Conference on Learning Theory . PMLR, 2022, pp. 625–661

    2022

  9. [17]

    Subsampling suffices for adaptive data analysis,

    G. Blanc, “Subsampling suffices for adaptive data analysis,” in Proceedings of the 55th Annual ACM Symposium on Theory of Computing , 2023, pp. 999–1012

    2023

  10. [18]

    Interactive fingerprinting codes and the hardness of preventing false discovery,

    T. Steinke and J. Ullman, “Interactive fingerprinting codes and the hardness of preventing false discovery,” in Conference on Learning Theory, 2015, pp. 1588–1628

    2015

  11. [19]

    Make up your mind: The price of online queries in differential privacy,

    M. Bun, T. Steinke, and J. Ullman, “Make up your mind: The price of online queries in differential privacy,” in Proceedings of the twenty-eighth annual ACM-SIAM symposium on discrete algorithms . SIAM, 2017, pp. 1306–1325

    2017

  12. [20]

    The limits of post-selection generalization,

    J. Ullman, A. Smith, K. Nissim, U. Stemmer, and T. Steinke, “The limits of post-selection generalization,” in Advances in Neural Information Processing Systems , 2018, pp. 6400–6409

    2018

  13. [21]

    A meta-analysis of overfitting in machine learning,

    R. Roelofs, V . Shankar, B. Recht, S. Fridovich-Keil, M. Hardt, J. Miller, and L. Schmidt, “A meta-analysis of overfitting in machine learning,” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  14. [22]

    Natural analysts in adaptive data analysis,

    T. Zrnic and M. Hardt, “Natural analysts in adaptive data analysis,” arXiv preprint arXiv:1901.11143 , 2019

    1901 arXiv

  15. [23]

    Model similarity mitigates test set overuse,

    H. Mania, J. Miller, L. Schmidt, M. Hardt, and B. Recht, “Model similarity mitigates test set overuse,” in Advances in Neural Information Processing Systems, 2019, pp. 9993–10 002

    2019

  16. [24]

    The advantages of multiple classes for reducing overfitting from test set reuse,

    V . Feldman, R. Frostig, and M. Hardt, “The advantages of multiple classes for reducing overfitting from test set reuse,” in Proceedings of the 36th International Conference on Machine Learning , 2019, pp. 1892–1900

    2019

  17. [25]

    Bayesian adaptive data analysis guarantees from subgaussianity,

    S. Elder, “Bayesian adaptive data analysis guarantees from subgaussianity,” arXiv preprint arXiv:1611.00065 , 2016. 22

    2016 arXiv

  18. [26]

    Challenges in bayesian adaptive data analysis,

    ——, “Challenges in bayesian adaptive data analysis,” arXiv preprint arXiv:1604.02492 , 2016

    2016 arXiv

  19. [27]

    Over-the-air federated adaptive data analysis: Preserving accuracy via opportunistic differential privacy,

    A. H. Hadavi, M. M. Mojahedian, and M. R. Aref, “Over-the-air federated adaptive data analysis: Preserving accuracy via opportunistic differential privacy,” 2024. [Online]. Available: https://arxiv.org/abs/2411.15948

    2024 arXiv

  20. [28]

    Guaranteed validity for empirical approaches to adaptive data analysis,

    R. Rogers, A. Roth, A. Smith, N. Srebro, O. D. Thakkar, and B. Woodworth, “Guaranteed validity for empirical approaches to adaptive data analysis,” in International Conference on Artificial Intelligence and Statistics , 2020, pp. 2830–2840

    2020

  21. [29]

    Pu and M

    X. Pu and M. Kay, “The garden of forking paths in visualization: A design space for reliable exploratory visual analytics: Position paper. in 2018 ieee evaluation and beyond-methodological approaches for visualization (beliv),” IEEE Computer Society, Berlin, Germany , pp. 37–45, 2018

    2018

  22. [30]

    Bayesian causal inference: A unifying neuroscience theory,

    L. Shams and U. Beierholm, “Bayesian causal inference: A unifying neuroscience theory,” Neuroscience & Biobehavioral Reviews , vol. 137, p. 104619, 2022

    2022

  23. [31]

    Neural implementations of bayesian inference,

    H. Sohn and D. Narain, “Neural implementations of bayesian inference,” Current Opinion in Neurobiology , vol. 70, pp. 121–129, 2021

    2021

  24. [32]

    Learning invariant features using the transformed indian buffet process,

    J. Austerweil and T. Griffiths, “Learning invariant features using the transformed indian buffet process,” Advances in neural information processing systems, vol. 23, 2010

    2010

  25. [33]

    Bayesian models of conceptual development: Learning as building models of the world,

    T. D. Ullman and J. B. Tenenbaum, “Bayesian models of conceptual development: Learning as building models of the world,” Annual Review of Developmental Psychology , vol. 2, pp. 533–558, 2020

    2020

  26. [34]

    The frequentist theory of bayesian statistics,

    B. Kleijn, “The frequentist theory of bayesian statistics,” 2022

    2022

  27. [35]

    N. L. Hjort, C. Holmes, P. Müller, and S. G. Walker, Bayesian nonparametrics. Cambridge University Press, 2010, vol. 28

    2010

  28. [36]

    Lectures on nonparametric bayesian statistics,

    B. Kleijn, A. van der Vaart, and H. van Zanten, “Lectures on nonparametric bayesian statistics,” Lecture notes, under construction , 2012

    2012

  29. [37]

    Pólya tree posterior distributions on densities,

    I. Castillo, “Pólya tree posterior distributions on densities,” 2017

    2017

  30. [38]

    Prior knowledge elicitation: The past, present, and future,

    P. Mikkola, O. A. Martin, S. Chandramouli, M. Hartmann, O. A. Pla, O. Thomas, H. Pesonen, J. Corander, A. Vehtari, S. Kaski et al., “Prior knowledge elicitation: The past, present, and future,” arXiv preprint arXiv:2112.01380 , 2021

    2021 arXiv

  31. [39]

    Bayesian workflow,

    A. Gelman, A. Vehtari, D. Simpson, C. C. Margossian, B. Carpenter, Y . Yao, L. Kennedy, J. Gabry, P.-C. Bürkner, and M. Modrák, “Bayesian workflow,” 2020

    2020

  32. [40]

    Efficient stabilization of imprecise statistical inference through conditional belief updating,

    J. Drevet, J. Drugowitsch, and V . Wyart, “Efficient stabilization of imprecise statistical inference through conditional belief updating,” Nature Human Behaviour , vol. 6, no. 12, pp. 1691–1704, 2022

    2022

  33. [41]

    Modeling, replicating, and predicting human behavior: A survey,

    A. Fuchs, A. Passarella, and M. Conti, “Modeling, replicating, and predicting human behavior: A survey,” ACM Transactions on Autonomous and Adaptive Systems , 2023

    2023

  34. [42]

    Optional pólya trees: Posterior rates and uncertainty quantification,

    I. Castillo and T. Randrianarisoa, “Optional pólya trees: Posterior rates and uncertainty quantification,” Electronic Journal of Statistics , vol. 16, no. 2, pp. 6267–6312, 2022

    2022

  35. [43]

    Prior distributions on spaces of probability measures,

    T. S. Ferguson, “Prior distributions on spaces of probability measures,” The Annals of Statistics , vol. 2, no. 4, pp. 615–629, 1974

    1974

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.