Pith. sign in

REVIEW 4 major objections 6 minor 67 references

An Interactive Framework for Finding the Optimal Trade-off in Differential Privacy

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that in differentially private learning the privacy-accuracy Pareto front is S-shaped, and that modeling it directly—together with asking users to pick a point on a hypothetical trade-off curve—finds the preferred privacy b

desk verdict Clever interactive MOO for choosing epsilon in DP, but the privacy accounting in the HPO loop is wrong and the S-curve derivation has an error; worth a serious referee though. read the letter →

arxiv 2509.04290 v1 pith:HALUYCKX submitted 2025-09-04 cs.LG

Yaohong Yang , Aki Rehn , Sammie Katt , Antti Honkela , Samuel Kaski This is my paper · ORCID
classification cs.LG
keywords differentialprivacybudgetselectionmulti-objectiveoptimizationParetofrontpreferencelearningBayesianS-shapedtrade-offcurveknowledgegradient
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

Choosing a privacy budget in differential privacy is usually a guess: either a conventional value or a costly search over the entire Pareto front. This paper tries to make the choice principled by showing that the front—best achievable accuracy versus privacy level—is S-shaped, and by exploiting that shape directly. Their framework fixes a privacy level, runs hyperparameter optimization to produce one point on the front, and fits a Bayesian sigmoid or Gompertz surrogate to a handful of such points. To learn the decision-maker's preference, it asks them to pick their preferred point on a hypothetical trade-off curve rather than answer pairwise comparisons. If the S-shape claim holds, practitioners reach their optimal privacy-accuracy point with far fewer expensive hyperparameter runs and user queries than generic interactive multi-objective baselines.

What carries the argument

The load-bearing object is the parametric S-shaped surrogate h(p; beta), either a sigmoid or a Gompertz curve mapping privacy level p to the best achievable accuracy. Two facts make it work: the c-constraint property of DP—maximizing accuracy at a fixed privacy level produces a Pareto-optimal point by construction—and a derivation for output-perturbed logistic regression showing that accuracy follows a Gompertz curve, a member of the S-shaped family. On top of this, the curve-selection query presents the fitted front to the user and asks for the preferred point, modeled by a Boltzmann-rational choice likelihood over a Chebyshev utility function; knowledge-gradient acquisition then selects bo

What would settle it

Fit the same sigmoid/Gompertz surrogate to a differentially private model family not covered here, such as DP-SGD training of large language models or private matrix factorization, over a dense grid of privacy levels. If the residuals are systematically structured—a front with two inflection points, a step, or a plateau followed by a cliff—the S-shape claim fails. A cheaper check: simulate output-perturbed logistic regression with non-uniform feature distributions and verify whether accuracy really follows the claimed Gompertz form with a constant rate.

Watch

Extended reading notes

Core claim

The paper's central claim is that in differentially private learning, the Pareto front—the curve giving the best achievable accuracy for each privacy level—has an S-shaped (sigmoid/Gompertz) form, and that this structure can be exploited directly. Fixing a privacy level and running hyperparameter optimization yields a point on the front by construction, so the whole curve can be fit as a Bayesian parametric surrogate after only a few such points. The paper replaces pairwise preference questions with a richer interaction: the user picks their preferred point on a hypothetical trade-off curve, interpreted through a Boltzmann-rational choice model over a Chebyshev utility function. A knowledge-

Load-bearing premise

The load-bearing premise is that the privacy-accuracy Pareto front is S-shaped for every differentially private model family, not just logistic regression with output perturbation; the paper states this as a 'posit' in Section 4.1, and if any common family has a differently shaped front, the sigmoid/Gompertz surrogate is misspecified and the claimed savings may not hold.

Editorial extensions

If this is right

  • Choosing a privacy budget becomes a data-driven process: a practitioner states a plausible range, answers a few curve queries, and receives a preferred privacy-accuracy point instead of relying on convention.
  • Because each hyperparameter run at one privacy level generates a Pareto-optimal point, the front is learned with far fewer expensive evaluations than generic two-surrogate multi-objective optimization.
  • Richer feedback per query—picking a point on a curve rather than a pairwise comparison—reduces the number of user interactions needed to converge.
  • The S-shaped surrogate stays useful even when the observed range covers only part of the curve, such as only the steep middle or only the lower asymptote, so users may define any application-relevant range of privacy levels.
  • Chebyshev utility is necessary in this setup because linear utility functions collapse to the endpoints of an S-shaped front and cannot express balanced trade-offs.

Reading between the lines

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

  • The same recipe—generating Pareto-optimal points by constrained optimization and fitting a parametric monotone front—could transfer to other two-objective problems where one objective is a hard budget constraint such as latency versus accuracy or cost versus quality; the S-shape claim is a testable template rather than a DP-specific fact.
  • The theoretical derivation actually demonstrates the S-shape only for output-perturbed logistic regression, and its integration step drops the claimed dependence on the input point, so whether the shape persists for other noise mechanisms such as gradient or objective perturbation remains an open empirical question that the paper's experiments only partially address.
  • The framework assumes real decision-makers follow the Boltzmann-rational model with a fixed temperature; a natural test is whether the advantage shrinks when the temperature is inferred online or when users are shown visually similar curves.
  • The paper itself notes that accuracy noise is likely heteroscedastic, with higher variance under strong privacy; modeling the variance as a function of the privacy level is a direct extension that could make the surrogate's uncertainty and the knowledge-gradient acquisition more honest.
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

4 major / 6 minor

Summary. The manuscript proposes an interactive multi-objective optimization framework for selecting a privacy level in differentially private machine learning. The authors claim (i) that the privacy-accuracy Pareto front h(p) is generically S-shaped, with a theoretical derivation for output-perturbed logistic regression; (ii) that the front can therefore be learned efficiently with a Bayesian sigmoid/Gompertz surrogate from few HPO runs; and (iii) that asking the decision-maker to pick a point on a hypothetical trade-off curve yields more information per query than pairwise comparisons. The framework alternates KG-based selection of privacy levels to evaluate and curves to present, and is evaluated on DP logistic regression and DP transfer learning across six datasets against six baselines. The reported experiments show lower preference-inference error and regret for the proposed method.

Significance. If correct, the paper would provide a practical answer to an open question—how to choose a privacy budget—and would reduce both HPO cost and user burden. The curve-based elicitation idea is a natural and promising departure from pairwise queries, and the ablation studies (Fig. 8) support its sample efficiency. However, the two load-bearing pillars—the theoretical S-shape derivation and the privacy accounting of the HPO loop—are not currently sound, and the empirical fits are in-sample. The claimed convergence and regret advantages are therefore not yet established, though the overall approach may be salvageable with re-framing and additional experiments.

major comments (4)
  1. [Section 6.1, Eq. (2)] The statement 'we do not account for the cumulative privacy cost of the iterative search' is invalid under differential privacy. DP is a property of the algorithm mapping the sensitive dataset to the released output. Here the released output is the final model, and its hyperparameters θ are selected by a non-private BO routine using a validation split of D. The mechanism is M(D)=A_{θ(D)}(D); neighboring datasets can induce different θ, and the final-model distributions need not satisfy the claimed (ε,δ)-DP. The trusted-curator threat model does not remove this requirement. Consequently Eq. (2) does not describe accuracy at privacy level p, and the experimental Pareto fronts (Figs. 4, 6) and regret comparisons are not at the reported ε. The manuscript needs either a fully private HPO/selection procedure (e.g., Liu & Talwar; Papernot & Steinke) with an explicit privacy budget, or a princip
  2. [Section 4.1, Eqs. (11)-(14)] The derivation of the S-shape is incorrect. In Eq. (13), h(ε,x)=1-0.5 exp(-Cε) with C=|ξ|/S_f does not depend on x; the negative-class calculation in Eq. (12) gives the same expression. The integral in Eq. (14) is therefore an identity, not an expectation that averages x-dependent terms. The Gompertz form appears only after the change of variables ε=exp(ε~), applied to a curve that is monotone exponential in ε. Moreover C depends on the unperturbed regression coefficient ξ, so the formula is not a parameter-free characterization. This does not support the 'key insight' that Pareto fronts are S-shaped 'in general'. The authors should either present a correct derivation (e.g., involving an x-dependent margin or a nontrivial averaging) or, as Section 4.1 later concedes, present the S-shape as an empirical posit.
  3. [Section 4.1, Figures 2-4 and 6] The empirical support for the S-shape is in-sample curve fitting. The same observations used to fit L,k,b,c are used to compute R²; this shows the functional form is flexible enough to interpolate the data, not that it predicts the front at unobserved privacy levels. The paper's central efficiency claim—modeling the entire Pareto front after observing only a few points—requires out-of-sample evaluation. I request a cross-validation (fit on a subset of privacy levels, evaluate on held-out privacy levels) and, ideally, a comparison with a non-S-shaped alternative (e.g., a monotone GP) on the same data. The Adult/Dutch fits (R²≈0.92 and 0.89) also show only a partial S-shape, so the evidence from these datasets is weak.
  4. [Section 6.4.1, Figure 8] The ablation that 'modeling the Pareto front requires clearly fewer samples' assumes the true front is S-shaped, since the regime is the same contested assumption. The comparison with GP-based objective modeling therefore does not isolate the benefit of the S-shaped surrogate family; it conflates prior knowledge of the front's shape with the choice of modeling the front directly. A misspecification analysis (e.g., data generated from a non-S-shaped front, or a real front known to deviate from S-shape) would be needed to support the claimed robustness and efficiency gains.
minor comments (6)
  1. [Section 4.1] Notation: 'E[h(ε)]' is used for an expectation that equals h(ε); please use distinct symbols for the random variable and the dummy variable in Eq. (14), and define ε~ more carefully.
  2. [Figures 5 and 7] Captions contain a duplicated word: 'smaller errors and and regrets'. There are also typos in §6.5 ('suppoer') and §5 ('Pareo front').
  3. [Figures 4 and 6] Since Section 6.1 reports averages over 30 runs, the empirical accuracy points in the fitting plots would benefit from error bars or confidence intervals.
  4. [Section 6.1] The sentence 'The only privacy guarantee relevant to the threat model is that of the single, final model chosen for deployment' should be removed or justified, as it is a restatement of the privacy-accounting problem rather than an argument.
  5. [Appendix C] The importance-sampling scheme uses the prior as the proposal and likelihood values as weights. This can degenerate as observations accumulate; please comment on effective sample size or resampling steps.
  6. [Algorithm 1] The pseudo-code uses 'if Perform Preference Elicitation:' without a defined condition; either specify the interleaving rule or replace with a for-loop over the alternating schedule.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the S-curve surrogate is a modeling assumption supported by a (flawed but non-circular) derivation and standard closed-loop Bayesian optimization; the privacy-accounting gap is a correctness issue, not a circular reduction.

full rationale

The paper's derivation chain is not circular. The theoretical motivation in Section 4.1 derives a pointwise accuracy formula (Eqs. 11-13) from the Laplace mechanism of output-perturbed logistic regression, then reparameterizes via p = -log ε to obtain a Gompertz form. That is a legitimate (if partly erroneous) mathematical consequence, not an equation that reduces to its own input by construction. The expectation step in Eq. (14) is indeed an identity because h(ε,x) has no x-dependence, and this is a real derivation defect, but it does not make the S-shape claim circular: the Gompertz form comes from the Laplace tail plus the explicit p = -log ε coordinate choice, both stated in the paper. The surrogate family is then fitted to empirical front data and used inside an active-learning loop; the reported regret is evaluated against the true front at the chosen privacy levels, so this is standard Bayesian optimization rather than a fitted parameter being renamed as a prediction. There is no load-bearing self-citation chain: the cited DP-FiLM work is used only as an experimental architecture, and the S-shape claim is supported by in-paper derivation and empirical fits. Separately, the paper has serious non-circular weaknesses: Section 6.1 asserts that non-private BO-based HPO need not be counted in the privacy budget because intermediate results are confidential, which is a privacy-accounting/correctness gap rather than a circular step; and Section 4.1 posits the general S-shape without proof for arbitrary DP models. These concerns affect validity but not circularity.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central premises the framework pulls from outside the paper are: the S-shape of the Pareto front, the Chebyshev utility representation, the Boltzmann-rational user model, and the Gaussian observation model. The S-curve parameters and noise variance are fitted per dataset; T is hand-chosen. No new physical entities or mediators are introduced.

free parameters (3)
  • S-curve parameters L, k, b, c (sigmoid or Gompertz) = not reported per dataset; inferred via Eq. (17) or fitted with curve_fit
    These four parameters define the modeled Pareto front and are the core fitted quantities; the central method depends on them.
  • Observation noise variance sigma^2 = inferred with prior Gamma(0.5, 0.1)
    Gaussian likelihood for observed accuracy (Eq. 17); homoscedastic assumption is acknowledged in Section 7.
  • Temperature T in the Boltzmann-rational choice model = 0.2 (fixed by hand)
    T controls user rationality in Eq. (21); chosen as 0.2 with a sensitivity analysis in Appendix A showing robustness.
assumptions (6)
  • domain assumption The Pareto front coincides with h(p), the best accuracy achievable at privacy level p via constrained optimization (Eq. 2).
    c-constraint solutions are only guaranteed weakly Pareto-optimal, and this assumes the inner HPO reaches the global optimum for every p (Sections 2.2 and 4.1).
  • ad hoc to paper The privacy-accuracy Pareto front is S-shaped for all DP models.
    Section 4.1: 'we posit that this trend is not specific to this definition but is a fundamental characteristic of the privacy-accuracy trade-off.' This is the load-bearing modeling assumption.
  • domain assumption The decision-maker's utility is representable by Chebyshev scalarization (Eq. 30) with unknown weights w.
    Used throughout preference learning and visualization; if the true utility is not Chebyshev-representable, inferred weights do not correspond to actual preferences.
  • domain assumption User choices follow the Boltzmann-rational model with known temperature (Eq. 21).
    The likelihood for the chosen trade-off point is modeled with this choice rule; experiments simulate users with the same rule.
  • domain assumption Observed accuracy is Gaussian noise around the S-curve.
    Eq. (17); homoscedastic Gaussian likelihood, with the limitation explicitly noted in Section 7.
  • standard math Standard Bayesian inference and Monte Carlo importance sampling are valid for posterior and KG estimation.
    Appendix C outlines importance-sampling approximations; these are standard tools.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Interactive Framework for Finding the Optimal Trade-off in Differential Privacy." pith.science (2026). https://pith.science/paper/HALUYCKX

@misc{pith2026250904290,
  author       = {Pith},
  title        = {Pith review of: An Interactive Framework for Finding the Optimal Trade-off in Differential Privacy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HALUYCKX}},
  note         = {Machine review of arXiv:2509.04290}
}
read the original abstract

Differential privacy (DP) is the standard for privacy-preserving analysis, and introduces a fundamental trade-off between privacy guarantees and model performance. Selecting the optimal balance is a critical challenge that can be framed as a multi-objective optimization (MOO) problem where one first discovers the set of optimal trade-offs (the Pareto front) and then learns a decision-maker's preference over them. While a rich body of work on interactive MOO exists, the standard approach -- modeling the objective functions with generic surrogates and learning preferences from simple pairwise feedback -- is inefficient for DP because it fails to leverage the problem's unique structure: a point on the Pareto front can be generated directly by maximizing accuracy for a fixed privacy level. Motivated by this property, we first derive the shape of the trade-off theoretically, which allows us to model the Pareto front directly and efficiently. To address inefficiency in preference learning, we replace pairwise comparisons with a more informative interaction. In particular, we present the user with hypothetical trade-off curves and ask them to pick their preferred trade-off. Our experiments on differentially private logistic regression and deep transfer learning across six real-world datasets show that our method converges to the optimal privacy-accuracy trade-off with significantly less computational cost and user interaction than baselines.

Figures

Figures reproduced from arXiv: 2509.04290 by the authors.

Figure 1
Figure 1. An overview of our interactive framework for finding the optimal privacy-utility trade-off. (Top Panel) Unlike standard pairwise comparisons (a), our interaction (b) elicits richer feedback by asking the user to select their ideal point on a hypothetical curve. This leads to faster convergence in learning user preferences. (Bottom Panel) Our method alternates between updating the user preference learning model based… view at source ↗
Figure 2
Figure 2. The theoretical (Gompertz) and empirically fitted (sig￾moid) privacy-accuracy tradeoff for differentially private logistic regression. The sigmoid curve presents a good fit to theoretical accuracy. approximate the Pareto front. Our theoretical derivation for differentially private logistic regression proves that its Pareto front follows a Gompertz function. The Gompertz curve is a member of the broader family of sig… view at source ↗
Figure 3
Figure 3. The Gompertz and sigmoid curves fitting to privacy￾accuracy tradeoff for differentially private transfer learning on CIFAR100 with δ = 10−5 . Both curves effectively capture the trend of the observed trade-off data. upper bound at a lower privacy level (e.g., ε → ∞), where the model’s accuracy approaches its non-private counterpart. The transition between these bounds inherently follows a pattern of diminishing retu… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Sigmoid and Gompertz functions fitted to the empirical privacy-accuracy tradeoff for DP logistic regression on the Adult and Dutch datasets with δ = 10−5 . Both curves effectively capture the trend of the observed trade-off data [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 6
Figure 6. Figure 6: Sigmoid and Gompertz functions fitted to the empirical privacy-accuracy tradeoff for DP deep transfer learning models with δ = 10−5 . Both curves effectively capture the trend of the observed trade-off data. PRV accountant (Gopi et al., 2021) that ships with Opacus. We…
Figure 7
Figure 7. Figure 7: Results of DP deep transfer learning results. The top row shows the preference inference errors and the bottom row shows the regrets. Our methods (Sigmoid and Gompertz) have smaller errors and and regrets than baselines. 0 5 10 15 20 25 30 35 40 Number of queries 0.0 0…
Figure 8
Figure 8. Figure 8: Results of ablation studies. left: ablation study of differ￾ent interaction mechanisms, where our method has the smallest inference errors. Right: ablation study of different methods to find the optimal trade-off, where ours converges to a smaller regret. tive, while p…
Figure 10
Figure 10. Figure 10: Visualization of linear utility function in (Priyanshu et al., 2022) for varying preference weights. The red solid curve presents the privacy-accuracy trade-off. Each star point marks the optimal privacy level that maximizes the utility for its correspond￾ing weight, …
Figure 11
Figure 11. Figure 11: Visualization of Chebyshev utility function for varying preference weights. The red solid curve presents the privacy￾accuracy trade-off. Each star point marks the optimal privacy level that maximizes the utility for its corresponding weight, which can successfully ide…
Figure 12
Figure 12. Figure 12: Preference inference errors for different T in the user model. There is no significant difference between different values. B. Baselines Implementation When employing GP to model surrogates of privacy and accuracy in baselines, we use the same Matern kernel as in (Ave…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 59 canonical work pages

  1. [1]

    B., Mironov, I., Talwar, K., and Zhang, L

    Abadi, M., Chu, A., Goodfellow, I., McMahon, H. B., Mironov, I., Talwar, K., and Zhang, L. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, pp.\ 308--318, New York, NY, USA, 2016. ACM, ACM

    work page 2016

  2. [2]

    Optuna: A Next -generation Hyperparameter Optimization Framework

    Akiba, T., Sano, S., Yanase, T., Ohta, T., and Koyama, M. Optuna: A Next -generation Hyperparameter Optimization Framework . In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , KDD '19, pp.\ 2623--2631, New York, NY, USA, July 2019. Association for Computing Machinery. ISBN 978-1-4503-6201-6. doi:10.1145/3...

    arXiv 2019

  3. [3]

    and Frazier, P

    Astudillo, R. and Frazier, P. Multi-attribute Bayesian optimization with interactive preference learning . In Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics, volume 108 of Proceedings of Machine Learning Research, pp.\ 4496--4507, Palermo, Sicily, Italy, 2020. PMLR

    work page 2020

  4. [4]

    J., Bakshy, E., and Frazier, P

    Astudillo, R., Lin, Z. J., Bakshy, E., and Frazier, P. qEUBO: A decision-theoretic acquisition function for preferential Bayesian optimization . In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics, volume 206 of Proceedings of Machine Learning Research, pp.\ 1093--1114, Valencia, Spain, 2023. PMLR

    work page 2023

  5. [5]

    Automatic Discovery of Privacy--Utility Pareto Fronts

    Avent, B., Gonz \'a lez, J., Diethe, T., Paleyes, A., and Balle, B. Automatic Discovery of Privacy--Utility Pareto Fronts . Proceedings on Privacy Enhancing Technologies, 2020 0 (4): 0 441--461, 2020

    work page 2020

  6. [6]

    and Kohavi, R

    Becker, B. and Kohavi, R. Adult . UCI Machine Learning Repository, 1996. DOI : https://doi.org/10.24432/C5XW20

    doi:10.24432/c5xw20 1996

  7. [7]

    Belakaria, S., Deshwal, A., and Doppa, J. R. Max-value entropy search for multi-objective Bayesian optimization . Advances in neural information processing systems, 32, 2019

    work page 2019

  8. [8]

    Multiobjective optimization: Interactive and evolutionary approaches , volume 5252

    Branke, J. Multiobjective optimization: Interactive and evolutionary approaches , volume 5252. Springer Science & Business Media, Berlin, Heidelberg, Germany, 2008

    work page 2008

Show all 67 references
  1. [9]

    and Steinke, T

    Bun, M. and Steinke, T. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of cryptography conference, pp.\ 635--658, Berlin, Heidelberg, 2016. Springer, Springer Berlin Heidelberg

    2016

  2. [10]

    and Monteleoni, C

    Chaudhuri, K. and Monteleoni, C. Privacy-preserving logistic regression. Advances in neural information processing systems, 21, 2008

    2008

  3. [11]

    and Ghahramani, Z

    Chu, W. and Ghahramani, Z. Preference learning with Gaussian processes . In Proceedings of the 22nd international conference on Machine learning, pp.\ 137--144, Vienna, Austria, 2005. PMLR

    2005

  4. [12]

    I need a better description

    Cummings, R., Kaptchuk, G., and Redmiles, E. M. " I need a better description": An Investigation Into User Expectations For Differential Privacy . In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pp.\ 3037--3052, New York, NY, USA, 2021. ACM

    2021

  5. [13]

    Dankar, F. K. and El Emam, K. The application of differential privacy to health data. In Proceedings of the 2012 Joint EDBT/ICDT Workshops, pp.\ 158--166, New York, NY, USA, 2012. ACM

    2012

  6. [14]

    Hypervolume knowledge gradient: a lookahead approach for multi-objective bayesian optimization with partial information

    Daulton, S., Balandat, M., and Bakshy, E. Hypervolume knowledge gradient: a lookahead approach for multi-objective bayesian optimization with partial information. In International Conference on Machine Learning, pp.\ 7167--7204. PMLR, 2023

    2023

  7. [15]

    L., and Balle, B

    De, S., Berrada, L., Hayes, J., Smith, S. L., and Balle, B. Unlocking High - Accuracy Differentially Private Image Classification through Scale , June 2022. URL http://arxiv.org/abs/2204.13650. arXiv:2204.13650 [cs, stat]

    2022 arXiv

  8. [16]

    Preference learning of latent decision utilities with a human-like model of preferential choice

    De Peuter, S., Zhu, S., Guo, Y., Howes, A., and Kaski, S. Preference learning of latent decision utilities with a human-like model of preferential choice. Advances in Neural Information Processing Systems, 37: 0 123608--123636, 2024

    2024

  9. [17]

    Dong, J., Roth, A., and Su, W. J. Gaussian differential privacy. Journal of the Royal Statistical Society Series B: Statistical Methodology, 84 0 (1): 0 3--37, 2022

    2022

  10. [18]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on...

    2021

  11. [19]

    and Rothblum, G

    Dwork, C. and Rothblum, G. N. Concentrated differential privacy. arXiv preprint arXiv:1603.01887, abs/1603.01887, 2016

    2016 arXiv

  12. [20]

    Calibrating noise to sensitivity in private data analysis

    Dwork, C., McSherry, F., Nissim, K., and Smith, A. Calibrating noise to sensitivity in private data analysis . In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, March 4-7, 2006. Proceedings 3, pp.\ 265--284, New York, NY, USA, 2006. Springer

    2006

  13. [21]

    Differential privacy in practice: Expose your epsilons! Journal of Privacy and Confidentiality, 9 0 (2), 2019

    Dwork, C., Kohli, N., and Mulligan, D. Differential privacy in practice: Expose your epsilons! Journal of Privacy and Confidentiality, 9 0 (2), 2019

    2019

  14. [22]

    and Martino, L

    Elvira, V. and Martino, L. Advances in Importance Sampling . IEEE Signal Processing Magazine, 39 0 (3): 0 39--60, 2022

    2022

  15. [23]

    The knowledge-gradient policy for correlated normal beliefs

    Frazier, P., Powell, W., and Dayanik, S. The knowledge-gradient policy for correlated normal beliefs . INFORMS journal on Computing, 21 0 (4): 0 599--613, 2009

    2009

  16. [24]

    u rnkranz, J. and H \

    F \"u rnkranz, J. and H \"u llermeier, E. Preference learning and ranking by pairwise comparison . In Preference learning, pp.\ 65--82. Springer, Berlin, Heidelberg, Germany, 2010

    2010

  17. [25]

    B., Stern, H

    Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. Bayesian data analysis. Chapman and Hall/CRC, 1995

    1995

  18. [26]

    Interactive hyperparameter optimization in multi-objective problems via preference learning

    Giovanelli, J., Tornede, A., Tornede, T., and Lindauer, M. Interactive hyperparameter optimization in multi-objective problems via preference learning . In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 12172--12180, Palo Alto, CA, USA, 2024. AAAI Press

    2024

  19. [27]

    Gonz \'a lez, J., Dai, Z., Damianou, A., and Lawrence, N. D. Preferential Bayesian Optimization . In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp.\ 1282--1291, Sydney, Australia, 2017. PMLR, PMLR

    2017

  20. [28]

    T., and Wutschitz, L

    Gopi, S., Lee, Y. T., and Wutschitz, L. Numerical Composition of Differential Privacy . In Advances in Neural Information Processing Systems , volume 34, pp.\ 11631--11642, Red Hook, New York, USA., 2021. Curran Associates, Inc. URL https://proceedings.neurips.cc/paper/2021/ha...

    2021

  21. [29]

    A review of multi-objective optimization: Methods and its applications

    Gunantara, N. A review of multi-objective optimization: Methods and its applications . Cogent Engineering, 5 0 (1): 0 1502242, 2018

    2018

  22. [30]

    J., Milli, S., and Dragan, A

    Jeon, H. J., Milli, S., and Dragan, A. Reward-rational (implicit) choice: A unifying formalism for reward learning . Advances in Neural Information Processing Systems, 33: 0 4415--4426, 2020

    2020

  23. [31]

    B., et al

    Kinga, D., Adam, J. B., et al. A method for stochastic optimization . In International conference on learning representations (ICLR), volume 5. California;, 2015

    2015

  24. [32]

    and Kulkarni, T

    Koskela, A. and Kulkarni, T. D. Practical differentially private hyperparameter tuning with subsampling. Advances in Neural Information Processing Systems, 36: 0 28201--28225, 2023

    2023

  25. [33]

    Learning multiple layers of features from tiny images

    Krizhevsky, A., Hinton, G., et al. Learning multiple layers of features from tiny images . Master's thesis, University of Toronto, 2009

    2009

  26. [34]

    J., Astudillo, R., Frazier, P., and Bakshy, E

    Lin, Z. J., Astudillo, R., Frazier, P., and Bakshy, E. Preference exploration for efficient Bayesian optimization with multiple outcomes . In International Conference on Artificial Intelligence and Statistics, pp.\ 4235--4258. PMLR, 2022

    2022

  27. [35]

    and Talwar, K

    Liu, J. and Talwar, K. Private selection from private candidates. In Charikar, M. and Cohen, E. (eds.), Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing, STOC 2019, Phoenix, AZ, USA, June 23-26, 2019 , pp.\ 298--309, New York, NY, USA, 2019. ACM . doi...

    2019

  28. [36]

    and Bu, Z

    Liu, R. and Bu, Z. Towards hyperparameter-free optimization with differential privacy . In The Thirteenth International Conference on Learning Representations, Singapore, 2025. ICLR

    2025

  29. [37]

    Effective implementation of the -constraint method in multi-objective mathematical programming problems

    Mavrotas, G. Effective implementation of the -constraint method in multi-objective mathematical programming problems . Applied mathematics and computation, 213 0 (2): 0 455--465, 2009

    2009

  30. [38]

    Visualizing Privacy-Utility Trade-Offs in Differentially Private Data Releases

    Nanayakkara, P., Bater, J., He, X., Hullman, J., and Rogers, J. Visualizing Privacy-Utility Trade-Offs in Differentially Private Data Releases . Proceedings on Privacy Enhancing Technologies, 2022 0 (3): 0 521--541, 2022

    2022

  31. [39]

    A., Cummings, R., Kaptchuk, G., and Redmiles, E

    Nanayakkara, P., Smart, M. A., Cummings, R., Kaptchuk, G., and Redmiles, E. M. What are the chances? explaining the epsilon parameter in differential privacy . In 32nd USENIX Security Symposium (USENIX Security 23), pp.\ 1613--1630, Berkeley, CA, USA, 2023. USENIX Association

    2023

  32. [40]

    Measure-observe-remeasure: An interactive paradigm for differentially-private exploratory analysis

    Nanayakkara, P., Kim, H., Wu, Y., Sarvghad, A., Mahyar, N., Miklau, G., and Hullman, J. Measure-observe-remeasure: An interactive paradigm for differentially-private exploratory analysis . In 2024 IEEE Symposium on Security and Privacy (SP), pp.\ 1047--1064, Piscataway, NJ, US...

    2024

  33. [41]

    Y., et al

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A. Y., et al. Reading digits in natural images with unsupervised feature learning . In NIPS workshop on deep learning and unsupervised feature learning, volume 2011, pp.\ 4. Granada, 2011

    2011

  34. [42]

    Multi-Objective Bayesian Optimization with Active Preference Learning

    Ozaki, R., Ishikawa, K., Kanzaki, Y., Takeno, S., Takeuchi, I., and Karasuyama, M. Multi-Objective Bayesian Optimization with Active Preference Learning . In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 14490--14498, Palo Alto, CA, USA, 2024. ...

    2024

  35. [43]

    A New Linear Scaling Rule for Private Adaptive Hyperparameter Optimization

    Panda, A., Tang, X., Mahloujifar, S., Sehwag, V., and Mittal, P. A New Linear Scaling Rule for Private Adaptive Hyperparameter Optimization . In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp.\ 3...

    2024

  36. [44]

    and Steinke, T

    Papernot, N. and Steinke, T. Hyperparameter tuning with renyi differential privacy. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. URL https://openreview.net/forum?id=-70L8lpp9DF

    2022

  37. [45]

    and Khorram, E

    Pirouz, B. and Khorram, E. A computational approach based on the -constraint method in multi-objective optimization problems. Adv. Appl. Stat, 49 0 (6): 0 453--483, 2016

    2016

  38. [46]

    Efficient hyperparameter optimization for differentially private deep learning

    Priyanshu, A., Naidu, R., Mireshghallah, F., and Malekzadeh, M. Efficient hyperparameter optimization for differentially private deep learning . In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pp.\ 2455--2468, New York, NY, USA, 2022. ACM

    2022

  39. [47]

    and Agarwal, S

    Rajkumar, A. and Agarwal, S. A Differentially Private Stochastic Gradient Descent Algorithm for Multiparty Classification . In AISTATS , pp.\ 933--941, La Palma, Canary Islands, January 2012

    2012

  40. [48]

    B., Noy, A., and Zelnik, L

    Ridnik, T., Baruch, E. B., Noy, A., and Zelnik, L. ImageNet-21K Pretraining for the Masses . In Vanschoren, J. and Yeung, S.-K. (eds.), The Thirty-fifth Conference on Neural Information Processing Systems, NeurIPS, Track on Datasets and Benchmarks, 2021

    2021

  41. [49]

    and Monro, S

    Robbins, H. and Monro, S. A stochastic approximation method. The annals of mathematical statistics, 22: 0 400--407, 1951

    1951

  42. [50]

    TAN Without a Burn : Scaling Laws of DP-SGD

    Sander, T., Stock, P., and Sablayrolles, A. TAN Without a Burn : Scaling Laws of DP-SGD . In The Fortieth International Conference on Machine Learning , ICML 2023 , pp.\ 29937--29949. PMLR, 2023

    2023

  43. [51]

    B., Kirsch, A., Farquhar, S., Gal, Y., Foster, A., and Rainforth, T

    Smith, F. B., Kirsch, A., Farquhar, S., Gal, Y., Foster, A., and Rainforth, T. Prediction-oriented bayesian active learning . In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics, volume 206 of Proceedings of Machine Learning Research, ...

    2023

  44. [52]

    Song, S., Chaudhuri, K., and Sarwate, A. D. Stochastic gradient descent with differentially private updates. In 2013 IEEE Global Conference on Signal and Information Processing , pp.\ 245--248. IEEE, December 2013. ISBN 978-1-4799-0248-4. doi:10.1109/GlobalSIP.2013.6736861

    2013

  45. [53]

    Multi-objective Bayesian optimization using Pareto-frontier entropy

    Suzuki, S., Takeno, S., Tamura, T., Shitara, K., and Karasuyama, M. Multi-objective Bayesian optimization using Pareto-frontier entropy . In International conference on machine learning, pp.\ 9279--9288. PMLR, 2020

    2020

  46. [54]

    Privacy Loss in Apple's Implementation of Differential Privacy on MacOS 10.12

    Tang, J., Korolova, A., Bai, X., Wang, X., and Wang, X. Privacy Loss in Apple's Implementation of Differential Privacy on MacOS 10.12 . In Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pp.\ 183--199, New York, NY, USA, 2017. ACM

    2017

  47. [55]

    F., Paverd, A., Tople, S., Zanella-Beguelin, S., Turner, R

    Tobaben, M., Shysheya, A., Bronskill, J. F., Paverd, A., Tople, S., Zanella-Beguelin, S., Turner, R. E., and Honkela, A. On the efficacy of Differentially Private Few -shot Image Classification . Transactions on Machine Learning Research, 2023: 0 49, August 2023. ISSN 2835-885...

    2023

  48. [56]

    and Branke, J

    Ungredda, J. and Branke, J. When to elicit preferences in multi-objective Bayesian optimization . In Proceedings of the Companion Conference on Genetic and Evolutionary Computation, pp.\ 1997--2003, New York, NY, USA, 2023. ACM

    1997

  49. [57]

    Single interaction multi-objective Bayesian optimization

    Ungredda, J., Branke, J., Marchi, M., and Montrone, T. Single interaction multi-objective Bayesian optimization . In Parallel Problem Solving from Nature -- PPSN XVII, volume 13398 of Lecture Notes in Computer Science, pp.\ 132--145, Cham, Switzerland, 2022. Springer, Springer...

    2022

  50. [58]

    Integrating administrative registers and household surveys

    Van der Laan, P. Integrating administrative registers and household surveys. Netherlands Official Statistics, 15 0 (2): 0 7--15, 2000

    2000

  51. [59]

    S., Linmans, J., Winkens, J., Cohen, T., and Welling, M

    Veeling, B. S., Linmans, J., Winkens, J., Cohen, T., and Welling, M. Rotation equivariant CNNs for digital pathology . In Medical Image Computing and Computer Assisted Intervention -- MICCAI 2018, volume 11071 of Lecture Notes in Computer Science, pp.\ 210--218, Cham, Switzerl...

    2018

  52. [60]

    PyTorch Image Models

    Wightman, R. PyTorch Image Models . https://github.com/rwightman/pytorch-image-models, 2019

    2019

  53. [61]

    and Rasmussen, C

    Williams, C. and Rasmussen, C. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995

    1995

  54. [62]

    Winsor, C. P. The gompertz curve as a growth curve. Proceedings of the national academy of sciences, 18 0 (1): 0 1--8, 1932

    1932

  55. [63]

    A., Oliva, A., and Torralba, A

    Xiao, J., Hays, J., Ehinger, K. A., Oliva, A., and Torralba, A. Sun database: Large-scale scene recognition from abbey to zoo . In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.\ 3485--3492, Piscataway, NJ, USA, 2010. IEEE, IEEE

    2010

  56. [64]

    Relatively Rational: Learning Utilities and Rationalities Jointly from Pairwise Preferences

    Yamagata, T., Oberkofler, T., Kaufmann, T., Bengs, V., H \"u llermeier, E., and Santos-Rodriguez, R. Relatively Rational: Learning Utilities and Rationalities Jointly from Pairwise Preferences . In ICML 2024 Workshop on Models of Human Feedback for AI Alignment, volume 2024, V...

    2024

  57. [65]

    Multi-objective bayesian global optimization using expected hypervolume improvement gradient

    Yang, K., Emmerich, M., Deutz, A., and B \"a ck, T. Multi-objective bayesian global optimization using expected hypervolume improvement gradient. Swarm and evolutionary computation, 44: 0 945--956, 2019

    2019

  58. [66]

    Opacus: U ser-Friendly Differential Privacy Library in PyTorch

    Yousefpour, A., Shilov, I., Sablayrolles, A., Testuggine, D., Prasad, K., Malek, M., Nguyen, J., Ghosh, S., Bharadwaj, A., Zhao, J., Cormode, G., and Mironov, I. Opacus: U ser-Friendly Differential Privacy Library in PyTorch . arXiv preprint arXiv:2109.12298, abs/2109.12298, 2021

    2021 arXiv

  59. [67]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.