Pith. sign in

REVIEW 3 major objections 3 minor 46 references

A piecewise-linear solution path makes kernel-based conditional conformal prediction fast enough for routine use, at the same cost as a single kernel quantile fit.

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 · deepseek-v4-flash

2026-08-04 14:38 UTC pith:3VOIU3YT

load-bearing objection A solid algorithmic contribution with credible speedups, but the finite-sample theory does not cover the actually tuned procedure and the coverage claims are overstated. the 3 major comments →

arxiv 2509.24100 v2 pith:3VOIU3YT submitted 2025-09-28 stat.ME cs.LG

SpeedCP: Fast Kernel-based Conditional Conformal Prediction

Yating Liu , Yeo Jin Jung , Zixuan Wu , So Won Jeong , Claire Donnat This is my paper
classification stat.ME cs.LG MSC 62G0862G15
keywords conformal predictionconditional coverageRKHSquantile regressionsolution pathlatent embeddingskernel methodscoverage guarantees
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.

The paper tries to show that the costly RKHS-based conditional conformal prediction of Gibbs et al. can be made practical by tracing the full solution path of the underlying regularized quantile regression. The key algorithmic claim is that the regression coefficients evolve piecewise-linearly in the regularization parameter λ and in the imputed test score S, so the entire family of fits is obtained at essentially the cost of one kernel quantile fit. This makes it possible to tune smoothness and construct prediction sets without repeated refitting. The authors further claim that conditioning on learned low-rank embeddings yields reliable conditional coverage in high dimensions, with experiments reporting 30% shorter intervals and a 40x speedup over the earlier method.

Core claim

The central discovery is that the RKHS-based conditional conformal prediction problem—a regularized quantile regression with pinball loss over an RKHS augmented by a linear term—has a solution path that is piecewise-linear in both the regularization parameter λ and the imputed test score S. By tracking only the 'elbow' set of points whose residuals are exactly zero, the algorithm updates the solution only at discrete events, reducing computation dramatically. On the theory side, the paper proves finite-sample coverage guarantees for the randomized version: a localized coverage bound with an explicit gap (Theorem 1) and exact group-conditional coverage at level 1−α when the estimated top-1 la

What carries the argument

The load-bearing object is the pair of solution-path algorithms: the λ-path traces how the regularized RKHS quantile regression coefficients (kernel weights υ and linear coefficients η) change as the smoothness parameter λ varies, and the S-path traces the same coefficients as the imputed test score S varies. Both paths rely on the Karush-Kuhn-Tucker structure of the pinball-loss problem, which partitions points into elbow (zero residual), left, and right sets; between events, υ is affine in λ (and η affine in 1/λ) on the λ-path, and both are affine in S on the S-path. The elbow set is typically much smaller than n, giving the speedup.

Load-bearing premise

The exact group-conditional coverage theorem requires that the estimated top-1 latent topic equals the true top-1 topic for every point almost surely (Assumption/equation 11); if the embedding is imperfect, the stated equality of coverage at level 1−α need not follow, and this condition is not checked in the experiments.

What would settle it

On a synthetic dataset with three overlapping latent topics chosen so that the top-1 margin is small for a nontrivial fraction of points (violating the alignment condition), compute SpeedCP's coverage conditioned on estimated topics at level 1−α=0.9 over many trials; if coverage departs materially from 0.9 in the regions where T̂(X) differs from T(X), the exact guarantee in Theorem 2 is not operating as stated.

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

If this is right

  • If the path algorithm works as claimed, RKHS-based conditional conformal prediction becomes deployable on large datasets, where the previous method was computationally prohibitive.
  • The closed-form hyperparameter selection via λ-path cross-validation removes the need for expensive grid search over regularization.
  • The S-path replaces binary search over imputed scores with a single root computation, a direct corollary of the monotonicity of the test-point coefficient in S.
  • The extension to low-rank latent embeddings makes conditional coverage meaningful in high-dimensional problems like molecular property prediction and topic modeling, provided the embedding preserves the conditioning structure.
  • The theory quantifies how embedding estimation error degrades localized coverage, giving a measurable gap rather than a purely asymptotic statement.

Where Pith is reading between the lines

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

  • The piecewise-linear path structure is not specific to conformal prediction: the same λ/S tracing could be applied to any regularized RKHS quantile regression, potentially benefiting other resampling or sensitivity analyses that need many refits.
  • The exact group-conditional guarantee in Theorem 2 depends on the alignment condition T̂(X)=T(X) almost surely; if that condition fails—for instance when topics overlap so the top-1 margin is small—coverage may degrade, and the paper's experiments do not verify the condition, so a natural test is to measure coverage conditioned on estimated topics under violated alignment.
  • The claim that RLCP reduces to SplitCP in high dimensions (because locality weights become uninformative) suggests a general phenomenon about kernel/weighted conformal methods in high-dimensional spaces, and SpeedCP's low-rank conditioning might serve as a remedy; this is an implicit lesson beyond the paper.

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

3 major / 3 minor

Summary. The paper proposes SpeedCP, a fast algorithm for RKHS-based conditional conformal prediction. It replaces the repeated full RKHS quantile regressions of Gibbs et al. (2023) with piecewise-linear solution paths in the regularization parameter λ and the imputed score S, and it extends the framework to low-rank latent embeddings with a data-driven kernel. The paper claims finite-sample conditional coverage guarantees (Theorems 1 and 2, via Lemma 4), a 30% reduction in interval length, and a 40-fold speedup over CondCP. Experiments cover synthetic admixture data, molecular graphs, ArXiv abstracts, and brain-tumor MRI.

Significance. The algorithmic contribution is substantial: the path-following formulation in Appendices B and C is detailed and appears to deliver the claimed computational gains; the experimental study is extensive and compares against a broad set of competitors; the code is publicly available. If the theoretical guarantees held for the deployed algorithm, the paper would be a strong methodological advance. However, the core theoretical result (Lemma 4) is not justified for the actual algorithm because the hyperparameters are tuned on the calibration set only, breaking the exchangeability/symmetry step in its proof. This undercuts the finite-sample coverage claims in Theorems 1 and 2. The algorithmic findings remain interesting, but the theoretical claims need substantial repair.

major comments (3)
  1. [Appendix C.1, Lemma 4] The proof of Lemma 4 relies on the step: 'By the exchangeability of the data and the symmetry of ˆg_{S_{n+1}}' to assert that -E[f(X_{n+1})ˆυ_{S_{n+1},n+1}] equals -2E[λ⟨ˆg_{S_{n+1},ψ*}, f_ψ*⟩]. This symmetry requires the fitted coefficients to be invariant under permutations of all n+1 points. However, Algorithm 1 (Step 3) selects (γ̂, λ̂) by k-fold cross-validation on D_calib before the S-path; λ̂ is therefore a function of the calibration data but not of the test point. The fitted function is not symmetric in the n+1 positions, so the exchangeability argument fails. Remark 5 addresses only the monotonicity of S↦ˆυ_S, not this dependence on calibration outcomes. Since Theorem 1 (Eq. 10) and Theorem 2 (Eq. 12) both invoke Lemma 4, their conclusions are not established for the tuned algorithm. The authors should either prove a version of Lemma 4 that accounts for data-dependent hyperpara
  2. [Section 2.2, Theorem 2 and Eq. (11)] The exact group-conditional coverage P(Y_{n+1}∈Ĉ*_rand | ˆT(X_{n+1})=k)=1−α holds only under the alignment condition ˆT(X)=T(X) almost surely (Eq. 11). This condition is not verified in any experiment. In the synthetic admixture setting the true mixture proportions π(X) are known, so the alignment accuracy of the estimated topic assignment could be reported, but it is not. Lemma 9 in Appendix C provides only a sufficient margin condition; the paper does not check whether that margin holds in the experiments (e.g., the ArXiv topic model). Without (11), Theorem 2 does not apply to the reported 'topic-conditional coverage' results. At minimum, the experiments should include a diagnostic for alignment, or the claims should be weakened to approximate coverage under a misspecification error.
  3. [Section 2.2, Theorem 1 and Eq. (10)] Equation (10) is not a finite-sample coverage guarantee. The residual term after 1−α is -E[Σ_{i∈[n+1]} ˆυ_{S_rand,i} ψ*_W(W′, ˆπ(X_i))] / E[ψ*_W(W′, ˆπ(X))], which depends on the fitted kernel coefficients. These coefficients can be positive or negative, and the term is not bounded in magnitude. The paper labels this a 'coverage gap' and says it is 'explicitly measurable,' but without a bound on the residual, the statement is an exact decomposition rather than a guarantee. The abstract and Section 1 claim 'finite-sample guarantees'; this theorem does not provide one. The authors should either supply a bound on the residual (in terms of n, λ, and the function class) or explicitly state that Theorem 1 is only an approximation and not a finite-sample guarantee.
minor comments (3)
  1. [Section 2.1, Eq. (7)] The pinball loss is denoted 'lα' in Eq. (7) but 'ℓα' elsewhere; please unify notation.
  2. [Section 2.1.2] The notation 'U∼U nif(−α,1−α)' appears to be a typo; should be 'U∼Unif(−α,1−α)'.
  3. [Appendix B.4.3, Eq. (56)] The expression 'δ^T (S_{1:n} − ... )' mixes a row vector and a column vector without explicit transpose in the surrounding derivation; please check the typesetting and clarify the algebra.

Circularity Check

0 steps flagged

No significant circularity: the central derivation is imported from an external theorem, predictions are evaluated on held-out data, and no equation reduces to its own input by construction.

full rationale

The claimed derivation chain is not circular. The method fits a regularized RKHS quantile regression on calibration scores plus an imputed test score, traces lambda- and S-paths, and evaluates prediction sets on held-out test points; hyperparameters (gamma, lambda) are selected on a separate calibration split before the S-path. The coverage results (Theorems 1 and 2) are derived from Lemma 4, which is explicitly adapted from Theorem 3 of Gibbs et al. (2023) via the same exchangeability/duality argument; this is an external foundation, not a self-citation. The only circularity-adjacent feature is that the function class F* is defined using the fitted regularization parameter \hat{lambda} (Sections 2.1.1 and Lemma 4), but this is a modeling choice: the coverage statements are not equivalent to the fitted values by construction, and the theoretical claims include explicit coverage-gap terms (e.g., eq. 10) rather than asserting nominal coverage tautologically. Theorem 2's exact group-conditional statement relies on the alignment condition \hat{T}(X)=T(X) a.s. (eq. 11), which is an unverified assumption and a limitation, not a circular reduction. A separate concern—that tuning \hat{lambda} by cross-validation on the calibration set may break the exchangeability/symmetry step imported in the proof of Lemma 4—is a correctness risk, not a circularity; no equation in the paper reduces to its own input. The conclusion explicitly lists limitations, but none assert a circular step.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

The central machinery relies on standard conformal exchangeability plus several domain assumptions specific to the latent-space setup. The most fragile are the alignment condition (11) and the data-dependent definition of the function class via lambda_hat.

free parameters (4)
  • kernel bandwidth gamma = selected by cross-validation over a grid
    Controls the locality of the kernel psi* in the latent space; chosen by CV on the calibration set, so it is fit to the data.
  • regularization parameter lambda = selected by cross-validation along the lambda-path on the calibration set
    Controls smoothness of the quantile fit; the lambda-path computes the entire path, and the optimal lambda_hat is chosen by k-fold CV; lambda_hat also enters the definition of the function class F* in Lemma 4.
  • latent dimension K = 3 (molecules), 5 (ArXiv), 8 (MRI after PCA)
    Chosen per dataset by the practitioner; not data-adaptive, but central to the latent-space extension.
  • number of CV folds k = not specified precisely, described as k-fold
    Used for hyperparameter selection; user-chosen.
axioms (7)
  • domain assumption The n+1 data points (X_i, S_i) are exchangeable (Assumption 1)
    Standard in conformal prediction; stated in Assumption 1 and used throughout the proofs.
  • domain assumption The low-rank embedding pi_hat is fitted symmetrically with respect to the n+1 inputs (Assumption 2)
    Ensures conformal validity under permutation; used in Lemma 4 and Theorem 1.
  • ad hoc to paper The estimated top-1 topic assignment equals the true one almost surely: T_hat(X) = T(X) a.s. (equation 11)
    Required for Theorem 2's exact group-conditional coverage; very strong and not verified in experiments.
  • domain assumption There exists a density kernel psi*_W on the latent space with psi*_W(pi_hat(x1), pi_hat(x2)) = psi*(x1,x2) (Theorem 1)
    Needed to state the localized coverage gap in equation 10.
  • domain assumption Population strong convexity of the regularized loss (Assumption 3) and moment bounds (Assumption 4) hold
    Used for the coverage-gap estimation in Proposition 4.
  • standard math Columns of Phi* are linearly independent
    Used to derive the closed-form solution and projection arguments in the lambda- and S-path proofs (Appendix B).
  • standard math Positive definite kernel psi* and representer theorem apply
    Foundation of the RKHS representation used throughout.

pith-pipeline@v1.3.0-alltime-deepseek · 37834 in / 15888 out tokens · 410421 ms · 2026-08-04T14:38:13.980152+00:00 · methodology

0 comments
read the original abstract

Conformal prediction provides distribution-free prediction sets with finite-sample conditional guarantees. We build upon the RKHS-based framework of Gibbs et al. (2023), which leverages families of covariate shifts to provide approximate conditional conformal prediction intervals, an approach with strong theoretical promise, but with prohibitive computational cost. To bridge this gap, we develop a stable and efficient algorithm that computes the full solution path of the regularized RKHS conformal optimization problem, at essentially the same cost as a single kernel quantile fit. Our path-tracing framework simultaneously tunes hyperparameters, providing smoothness control and data-adaptive calibration. To extend the method to high-dimensional settings, we further integrate our approach with low-rank latent embeddings that capture conditional validity in a data-driven latent space. Empirically, our method provides reliable conditional coverage across a variety of modern black-box predictors, improving the interval length of Gibbs et al. (2023) by 30%, while achieving a 40-fold speedup.

Figures

Figures reproduced from arXiv: 2509.24100 by Claire Donnat, So Won Jeong, Yating Liu, Yeo Jin Jung, Zixuan Wu.

Figure 1
Figure 1. Figure 1: Mean coverage on fine-gridded partitions on the latent space (a 2D simplex). The results are aggregated over [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Conditional miscoverage and prediction set size for each fixed partition on the latent space. SpeedCP [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coverage on fixed partitions of the PC space for each molecule dataset. We use PCA on the last layer [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Averaged calibration and test density over 50 random generations of data. We use kmeans followed by [PITH_FULL_IMAGE:figures/full_fig_p032_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Voronoi tessellation of the PC space. We plot PC representation of graph embeddings where each color [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prediction set size on fixed partitions of the PC space for each molecule dataset. We use PCA on the last [PITH_FULL_IMAGE:figures/full_fig_p034_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Latent topics of ArXiv abstracts identified by probabilistic latent semantic indexing (pLSI), a topic modeling [PITH_FULL_IMAGE:figures/full_fig_p035_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of the most likely topic over all abstracts with [PITH_FULL_IMAGE:figures/full_fig_p035_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Evaluation of the CNN classifier on the Brain Tumor MRI dataset. [PITH_FULL_IMAGE:figures/full_fig_p036_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Predicted-label conditional coverage on the Brain Tumor MRI test set under the PCA-based model. Cali [PITH_FULL_IMAGE:figures/full_fig_p036_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Predicted-label conditional coverage on the Brain Tumor MRI test set by calibrating with the intercept only [PITH_FULL_IMAGE:figures/full_fig_p037_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Predicted-label conditional coverage on the Brain Tumor MRI test set under the PCA-based model. Cali [PITH_FULL_IMAGE:figures/full_fig_p037_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Predicted-label conditional coverage on the Brain Tumor MRI test set by calibrating with the intercept only [PITH_FULL_IMAGE:figures/full_fig_p038_13.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

46 extracted references · 11 linked inside Pith

  1. [1]

    Aitchison

    J. Aitchison. The statistical analysis of compositional data. Journal of the Royal Statistical Society: Series B (Methodological), 44 0 (2): 0 139--160, 1982

    1982

  2. [2]

    Angelopoulos, S

    A. Angelopoulos, S. Bates, J. Malik, and M. I. Jordan. Uncertainty sets for image classifiers using conformal prediction. arXiv preprint arXiv:2009.14193, 2020

    Pith/arXiv arXiv 2009

  3. [3]

    A. N. Angelopoulos, R. F. Barber, and S. Bates. Theoretical foundations of conformal prediction. arXiv preprint arXiv:2411.11824, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    M. C. U. Ara \'u jo, T. C. B. Saldanha, R. K. H. Galvao, T. Yoneyama, H. C. Chame, and V. Visani. The successive projections algorithm for variable selection in spectroscopic multicomponent analysis. Chemometrics and Intelligent Laboratory Systems, 57 0 (2): 0 65--73, 2001

    2001

  5. [5]

    Arora, R

    S. Arora, R. Ge, and A. Moitra. Learning topic models--going beyond svd. In 2012 IEEE 53rd Annual Symposium on Foundations of Computer Science, pages 1--10. IEEE, 2012

    2012

  6. [6]

    R. F. Barber, E. J. Candes, A. Ramdas, and R. J. Tibshirani. The limits of distribution-free conditional predictive inference. Information and Inference: A Journal of the IMA, 10 0 (2): 0 455--482, 2021

    2021

  7. [7]

    D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3 0 (Jan): 0 993--1022, 2003

    2003

  8. [8]

    Boucheron, O

    S. Boucheron, O. Bousquet, and G. Lugosi. Theory of classification: A survey of some recent advances. ESAIM: probability and statistics, 9: 0 323--375, 2005

    2005

  9. [9]

    C. B. Clement, M. Bierbaum, K. P. O'Keeffe, and A. A. Alemi. On the use of arxiv as a dataset, 2019

    2019

  10. [10]

    Donoho and V

    D. Donoho and V. Stodden. When does non-negative matrix factorization give a correct decomposition into parts? Advances in Neural Information Processing Systems, 16, 2003

    2003

  11. [11]

    Farquhar, J

    S. Farquhar, J. Kossen, L. Kuhn, and Y. Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630 0 (8017): 0 625--630, 2024

    2024

  12. [12]

    J. H. Friedman, T. Hastie, and R. Tibshirani. Regularization paths for generalized linear models via coordinate descent. Journal of statistical software, 33 0 (1): 0 1--22, 2010

    2010

  13. [13]

    Gibbs, J

    I. Gibbs, J. J. Cherian, and E. J. Cand \`e s. Conformal prediction with conditional guarantees. arXiv preprint arXiv:2305.12616, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Gillis and S

    N. Gillis and S. A. Vavasis. Fast and robust recursive algorithmsfor separable nonnegative matrix factorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36 0 (4): 0 698--714, 2013

    2013

  15. [15]

    L. Guan. Localized conformal prediction: A generalized inference framework for conformal prediction. Biometrika, 110 0 (1): 0 33--50, 2023

    2023

  16. [16]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman. The elements of statistical learning: data mining, inference, and prediction. Springer, 2009

    2009

  17. [17]

    T. Hofmann. Probabilistic latent semantic indexing. In Proceedings of the 22nd annual international ACM SIGIR Conference on Research and Development in Information Retrieval, pages 50--57, 1999

    1999

  18. [18]

    Hore and R

    R. Hore and R. F. Barber. Conformal prediction with local weights: randomization enables local guarantees. arXiv preprint arXiv:2310.07850, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Jang and E

    J. Jang and E. Cand \`e s. Tight distribution-free confidence intervals for local quantile regression. arXiv preprint arXiv:2307.08594, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Javadi and A

    H. Javadi and A. Montanari. Nonnegative matrix factorization via archetypal analysis. Journal of the American Statistical Association, 115 0 (530): 0 896--907, 2020

    2020

  21. [21]

    Johnstone and E

    C. Johnstone and E. Ndiaye. Exact and approximate conformal inference in multiple dimensions. arXiv preprint arXiv:2210.17405, 2022

    Pith/arXiv arXiv 2022

  22. [22]

    C. Jung, G. Noarov, R. Ramalingam, and A. Roth. Batch multivalid conformal prediction. arXiv preprint arXiv:2209.15145, 2022

    Pith/arXiv arXiv 2022

  23. [23]

    Kimeldorf and G

    G. Kimeldorf and G. Wahba. Some results on tchebycheffian spline functions. Journal of mathematical analysis and applications, 33 0 (1): 0 82--95, 1971

    1971

  24. [24]

    Klopp, M

    O. Klopp, M. Panov, S. Sigalla, and A. Tsybakov. Assigning topics to documents by successive projections. arXiv preprint arXiv:2107.03684, 2021

    Pith/arXiv arXiv 2021

  25. [25]

    R. Koenker. Quantile regression, volume 38. Cambridge university press, 2005

    2005

  26. [26]

    Y. Li, Y. Liu, and J. Zhu. Quantile regression in reproducing kernel hilbert spaces. Journal of the American Statistical Association, 102 0 (477): 0 255--268, 2007

    2007

  27. [27]

    X.-L. Meng. Posterior predictive p -values. The annals of statistics, 22 0 (3): 0 1142--1160, 1994

    1994

  28. [28]

    Messoudi, S

    S. Messoudi, S. Destercke, and S. Rousseau. Copula-based conformal prediction for multi-target regression. Pattern Recognition, 120: 0 108101, 2021

    2021

  29. [29]

    Papadopoulos, K

    H. Papadopoulos, K. Proedrou, V. Vovk, and A. Gammerman. Inductive confidence machines for regression. In Machine learning: ECML 2002: 13th European conference on machine learning Helsinki, Finland, August 19--23, 2002 proceedings 13, pages 345--356. Springer, 2002

    2002

  30. [30]

    Shahrokhi, D

    H. Shahrokhi, D. R. Roy, Y. Yan, V. Arnaoudova, and J. R. Doppa. Conformal prediction sets for deep generative models via reduction to conformal regression. arXiv preprint arXiv:2503.10512, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    J. Su, J. Luo, H. Wang, and L. Cheng. Api is enough: Conformal prediction for large language models without logit-access.(2024). URL https://arxiv. org/abs/2403, 1216, 2024

    2024

  32. [32]

    J. Sun, Y. Jiang, J. Qiu, P. Nobel, M. J. Kochenderfer, and M. Schwager. Conformal prediction for uncertainty-aware planning with diffusion dynamics model. Advances in Neural Information Processing Systems, 36: 0 80324--80337, 2023

    2023

  33. [33]

    Takeuchi, K

    I. Takeuchi, K. Nomura, and T. Kanamori. The entire solution path of kernel-based nonparametric conditional quantile estimator. In The 2006 IEEE International Joint Conference on Neural Network Proceedings, pages 153--158. IEEE, 2006

    2006

  34. [34]

    K. M. Tan, H. Battey, and W.-X. Zhou. Communication-constrained distributed quantile regression with optimal statistical guarantees. Journal of machine learning research, 23 0 (272): 0 1--61, 2022

    2022

  35. [35]

    Tibshirani

    R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58 0 (1): 0 267--288, 1996

    1996

  36. [36]

    R. J. Tibshirani. The solution path of the generalized lasso. Stanford University, 2011

    2011

  37. [37]

    Udell and A

    M. Udell and A. Townsend. Big data is low rank. SIAM News, 52 0 (9), 2019

    2019

  38. [38]

    V. Vovk. Conditional validity of inductive conformal predictors. In Asian conference on machine learning, pages 475--490. PMLR, 2012

    2012

  39. [39]

    V. Vovk, D. Lindsay, I. Nouretdinov, and A. Gammerman. Mondrian confidence machine. Technical Report, 2003

    2003

  40. [40]

    Wieslander, P

    H. Wieslander, P. J. Harrison, G. Skogberg, S. Jackson, M. Frid \'e n, J. Karlsson, O. Spjuth, and C. W \"a hlby. Deep learning with conformal prediction for hierarchical analysis of large-scale whole-slide tissue images. IEEE journal of biomedical and health informatics, 25 0 (2): 0 371--380, 2020

    2020

  41. [41]

    Z. Wu, B. Ramsundar, E. N. Feinberg, J. Gomes, C. Geniesse, A. S. Pappu, K. Leswing, and V. Pande. Moleculenet: a benchmark for molecular machine learning. Chemical science, 9 0 (2): 0 513--530, 2018

    2018

  42. [42]

    C. Xu, H. Jiang, and Y. Xie. Conformal prediction for multi-dimensional time series by ellipsoidal sets. arXiv preprint arXiv:2403.03850, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Zhang and E

    Y. Zhang and E. J. Cand \`e s. Posterior conformal prediction. arXiv preprint arXiv:2409.19712, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  45. [45]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  46. [46]

    A (ER T EX@Q

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    arXiv 2000