Pith. sign in

REVIEW 2 major objections 6 minor 35 references

A computable prediction region for multi-task kernel regression always covers the full-conformal set and is tight when tasks are related.

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

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

T0 review · grok-4.5

2026-07-12 09:21 UTC pith:NNETWEP5

load-bearing objection Solid, usable extension of stability-based approximate full conformal to multi-task RKHS regression; coverage by construction, volume bound informative under relatedness. the 2 major comments →

arxiv 2607.00645 v2 pith:NNETWEP5 submitted 2026-07-01 math.ST stat.MLstat.TH

Approximate full-conformal multi-task regression with reproducing kernels

Davidson Lova Razafindrakoto (SAMM) , Alain Celisse (SAMM) , J\'er\^ome Lacaille This is my paper
classification math.ST stat.MLstat.TH MSC 62G0862H1268T05
keywords multi-task regressionreproducing kernel Hilbert spacefull conformal predictionalgorithmic stabilityMahalanobis non-conformityellipsoidal prediction regioninter-task covariance
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.

Full conformal prediction gives a data-dependent region that contains the unknown multi-output vector at a prescribed confidence level, but computing it requires training a new predictor for every possible output and is therefore intractable. This paper works in a reproducing kernel Hilbert space of vector-valued functions and builds an explicit outer approximation that always contains the full conformal region, so the coverage guarantee is preserved. The construction uses algorithmic stability of the ridge-regularized multi-task predictor and comes in two versions: one when the inter-task covariance matrix is known (StableCP) and one when it is estimated from the same data (G-EllipsoidCP). When the covariance is known, a high-probability volume bound shows that the extra thickness shrinks as 1/(λn) and improves further when the tasks are strongly related. Synthetic experiments confirm that both approximations produce smaller, less variable regions than ordinary split conformal prediction while retaining empirical coverage.

Core claim

The upper StableCP (respectively G-EllipsoidCP) region always contains the full-conformal multi-task region and therefore satisfies the distribution-free coverage lower bound 1-α. When the inter-task covariance is known, the Lebesgue measure of the symmetric difference (the “thickness”) admits an explicit high-probability upper bound of order O(1/(λn)) whose dimension-dependent factors improve with stronger polynomial decay of the eigenvalues of the covariance.

What carries the argument

Uniform algorithmic stability of the multi-task ridge predictor in an RKHS of vector-valued functions, which yields additive (and, when the covariance is estimated, also multiplicative) corrections to the Mahalanobis non-conformity scores; these corrections turn the full conformal p-value into a closed-form ellipsoidal region that can be computed after a single training run.

Load-bearing premise

The eigenvalues of the inter-task covariance matrix must decay polynomially with an exponent greater than one that does not grow with the number of tasks; if the tasks are only weakly related the volume bound ceases to be informative.

What would settle it

On synthetic data with controlled inter-task correlation, measure the empirical volume of the StableCP region relative to the oracle conformal region; if the ratio fails to shrink like 1/(λn) or grows with ambient dimension when the eigenvalue-decay exponent is large, the thickness claim is false.

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

If this is right

  • Practitioners can obtain multi-task prediction ellipsoids with guaranteed coverage after training only one predictor instead of infinitely many.
  • When tasks share a strong covariance structure the extra volume of the approximation vanishes at the usual stability rate, making full conformal accuracy essentially free.
  • The same stability corrections extend immediately to robust losses (log-cosh, smoothed pinball, pseudo-Huber) that still satisfy the paper’s Lipschitz-convexity assumptions.
  • Estimating the covariance from the training sample itself does not destroy the coverage guarantee and still yields tighter regions than split conformal on the reported experiments.

Where Pith is reading between the lines

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

  • The same outer-approximation idea should transfer to other multi-output settings (multi-label classification, multi-target quantile regression) once a suitable matrix-valued kernel and Lipschitz loss are available.
  • If the eigenvalue-decay condition is replaced by a finite effective-rank assumption the dimension factors in the volume bound would remain controlled even for very high-dimensional outputs.
  • A natural next diagnostic is to track how the thickness behaves when the regularization parameter is chosen by structural risk minimization rather than fixed in advance.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper develops computable approximations to the full-conformal prediction region for multi-task kernel ridge regression in an RKHS of vector-valued functions. Using uniform algorithmic stability, it constructs upper (and lower) approximate non-conformity scores that sandwich the full-conformal p-value, yielding StableCP regions (known inter-task covariance Γ) and G-EllipsoidCP regions (estimated Γ) that contain the full-conformal set and therefore inherit the distribution-free coverage guarantee P(Y_{n+1} ∈ region) ≥ 1-α. For known Γ the regions are explicit Γ^{-1}-ellipsoids; a high-probability finite-sample upper bound on thickness (Lebesgue measure of the symmetric difference) of order O(1/(λn)) is proved under convex Lipschitz losses, bounded kernels, a source condition and polynomial eigenvalue decay of Γ. Synthetic experiments show the approximations are tighter than split conformal and that the empirical thickness decays consistently with the theory.

Significance. The work fills a genuine gap: full conformal is intractable for multi-task vector-valued predictors, while split conformal wastes data and produces larger regions. The stability-based sandwich construction is clean, the closed-form ellipsoidal expressions are immediately usable, and the thickness analysis (Theorem 35) supplies the first non-asymptotic volume control that explicitly improves with stronger task relatedness (larger γ). Reproducible code is provided. These contributions are of clear interest to the conformal-prediction and multi-task learning communities; the theoretical development is self-contained and the coverage claim rests only on exchangeability plus elementary score inequalities.

major comments (2)
  1. Theorem 35 (and the supporting volume argument in Lemma 33) relies on Assumption 29 (γ-EVDΓ) both to control det(Γ^{1/2}) and to keep the effective rank of Γ from exploding with ambient dimension p. The subsequent limit statement further requires 2t-1<γ whenever the output bound C_Y(p) grows like p^t. The paper correctly flags that the bound becomes non-informative for weak relatedness, yet the main text never quantifies how large γ must be relative to realistic multi-task covariances, nor does it supply a simple diagnostic (e.g., estimated effective rank) that a practitioner could check before trusting the volume guarantee. A short discussion or numerical illustration of the regime in which the prefactors remain useful would strengthen the claim that the approximation is 'tight'.
  2. All empirical comparisons (Figures 1–7) are performed on a single synthetic generator (Braun et al., 2026) with p=d=2. While the coverage and thickness plots are consistent with theory, the claim that StableCP/G-EllipsoidCP 'improve upon the split-conformal prediction' is supported only in this narrow setting. At least one additional synthetic regime (higher p, weaker eigenvalue decay, or non-Gaussian noise) or a small real multi-task data set would make the practical advantage more convincing; without it the empirical section remains illustrative rather than confirmatory.
minor comments (6)
  1. Author name and affiliation block appear corrupted in the source ('Davidson Lova Razafindrakotoda vidson-lov a.razafindrakoto'); please restore the correct spelling and contact information.
  2. Throughout the manuscript many mathematical expressions suffer from missing spaces or concatenated tokens (e.g., 'piqwhen', 'piiqwhen', 'pRλ;D', 'FullCP-region'). A careful typesetting pass is needed for readability.
  3. Definition 11 and Lemma 14 introduce the thickness via the symmetric difference; it would help the reader if the same symbol THK were used consistently for both the known-Γ and estimated-Γ cases (currently THK^Γ vs. THK^{pΓa}).
  4. In Section 4.4 the penalized criterion used to select λ is motivated by an expectation bound (Proposition 36) but the high-probability version is dropped without comment. A one-sentence justification would clarify why the SRM-style proxy is still reliable for the subsequent volume comparison.
  5. Figure captions (especially Figures 2 and 6) report estimated slopes; stating the exact regression model and the number of points used would improve reproducibility.
  6. References to 'Razafindrakoto et al. (2026)' appear both as prior single-task work and as the present multi-task paper; disambiguate the two citations.

Circularity Check

1 steps flagged

No significant circularity: coverage is inherited by elementary sandwiching of non-conformity scores from algorithmic stability; the thickness bound is a standard high-probability volume estimate under explicit assumptions, not a fitted prediction.

specific steps
  1. self citation load bearing [Section 3.1 / Definition 11 and surrounding text]
    "Compared with Razafindrakoto et al. (2026, see Definition 4), the present one is a generalization. ... the present definition allows for more general types of correction ..."

    The approximation scheme is presented as a generalization of the authors' own prior single-task construction. This is ordinary self-citation for context and is not load-bearing: the multi-task sandwich (Lemma 14), stability bound (Lemma 18), and coverage inheritance (Theorem 12) are proved from first principles (strong convexity + Lipschitz) without invoking the prior paper's theorems as black-box uniqueness or existence results.

full rationale

The derivation chain is self-contained and non-circular. FullCP is the classical exchangeability construction of Vovk et al. (Theorem 10). The upper-approximate region of Definition 11 is defined so that its p-value dominates the full conformal p-value whenever the approximate scores sandwich the true scores (Eq. 10); Theorem 12 and Lemma 14 then give containment and coverage inheritance by elementary indicator inequalities, with no free parameters fitted to the test point. The sandwich itself follows from the uniform stability bound of Lemma 18 (strong convexity of the ridge objective plus Lipschitz loss), which is re-derived in the multi-task RKHS setting and does not rely on the authors' prior single-task paper for its validity. Explicit ellipsoid expressions (Propositions 23, 39) and the thickness upper bound (Lemma 33, Theorem 35) are volume calculations under the listed assumptions (including the polynomial eigenvalue decay (γ-EVDΓ)); they quantify tightness but do not redefine the region or the coverage claim. Regularization parameters are chosen by a penalized empirical-risk proxy on an independent copy D1 (Proposition 36), independent of the conformal scores. Self-citations to Razafindrakoto et al. (2026) merely locate the single-task precursor; the multi-task arguments, stability constants, and volume factors are re-proved. No step reduces a claimed prediction or first-principles result to its own inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 1 invented entities

The central coverage claim rests on classical exchangeability plus the new stability sandwich; the volume claim additionally needs the listed analytic assumptions on loss, kernel and covariance spectrum. No free parameters are fitted to the final thickness bound; λ and a are chosen by a separate SRM-style criterion.

free parameters (2)
  • regularization λ
    Chosen by minimizing a penalized empirical risk on an independent hold-out set; the order of magnitude affects the thickness rate but is not fitted to the coverage or volume claims themselves.
  • covariance ridge a
    Fixed at 10^{-6} in experiments; appears in the estimated-covariance non-conformity scores.
axioms (5)
  • domain assumption Exchangeability of the n+1 pairs (X_i,Y_i)
    Standard conformal assumption (Vovk et al.); invoked for Theorem 10 and all coverage statements.
  • domain assumption Loss ℓ is convex, lower-semicontinuous, bounded below and ρ_p-Lipschitz in its second argument
    Assumptions 2-4,16; needed for existence of the ridge minimizer and for the uniform stability bound of Lemma 18.
  • domain assumption Matrix-valued kernel satisfies the operator-norm bound (κ_Γ-BdKΓ) linking K(x,x) to Γ
    Assumption 24; used to convert H-norm stability into Mahalanobis-score stability.
  • ad hoc to paper Eigenvalues of Γ decay as µ_ℓ ≤ C_Γ ℓ^{-γ} with γ>1 independent of p
    Assumption 29 (γ-EVDΓ); essential for the p-dependent factors in Theorem 35 to remain controlled.
  • domain assumption Source condition: the minimal-norm risk minimizer f_H has finite H-norm
    Assumption 28; classical for rates of kernel ridge regression, used to bound the population predictor.
invented entities (1)
  • StableCP / G-EllipsoidCP regions no independent evidence
    purpose: Explicit ellipsoidal outer approximations to the full-conformal set obtained by additive (and multiplicative) stability corrections.
    Defined in Definitions 21 and 30 / Proposition 39; they are algorithmic constructions, not new physical or mathematical objects requiring independent evidence.

pith-pipeline@v1.1.0-grok45 · 59963 in / 2915 out tokens · 31887 ms · 2026-07-12T09:21:04.110657+00:00 · methodology

0 comments
read the original abstract

Multi-task regression aims at jointly solving multiple regression problems, called tasks. Compared to solving each task separately, better performances can be achieved as long as the tasks are sufficiently related. Full-conformal prediction is a framework that formulates a data-dependent prediction-region containing the unknown output-vector at any prescribed confidence level. However, explicit computation of this prediction-region is intractable in general since it requires training infinitely many predictors. The present work focuses on multi-task regression in a Reproducing Kernel Hilbert Space (RKHS) of vector-valued functions. This computational issue is addressed by designing an approximating predictionregion containing the full-conformal one. This construction is carried out in two scenarios: piq when the inter-task covariance-matrix is known, and piiq when this matrix is estimated. In terms of volume, the tightness of this approximation is assessed theoretically by means of an upper-bound in the first scenario. It is also empirically proved to improve upon the split-conformal prediction on synthetic data in both scenarios.

Figures

Figures reproduced from arXiv: 2607.00645 by Alain Celisse (SAMM), Davidson Lova Razafindrakoto (SAMM), J\'er\^ome Lacaille.

Figure 3
Figure 3. Figure 3: Evolution of the penalized criterion for [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: Evolution of the empirical coverage-probability of the upper StableCP-region, for λ P ␣ 10´3 , 10´2 , 10´1 ( and for α within a logarithmic grid. The results are displayed in [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Evolution of the computable empirical upper-bound V ´ CrΓ,up λ;α pXn`1q zCrΓ,lo λ;α pXn`1q ¯ for the thickness THKΓ λ;α pXn`1q across 20 repetitions for α “ 0.1 for fixed λ “ 10´2 . On average, the empirical upper-bound V ´ CrΓ,up λ;α pXn`1q zCrΓ,lo λ;α pXn`1q ¯ gets smaller as the training sample size increases. The estimated exponent is close to the theoretical one which is 1 for a fixed λ (see Theorem 3… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison between the StableCP-regions (in red) and the SplitCP-regions (in green) in terms of volume relative to that of the OracleCP-regions (see Ap￾pendix F.2) for α “ 0.1 with |D| “ n “ 500 and |Dtrain| “ 375 [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evolution of the empirical coverage-probability of the upper G-EllipsoidCP￾region, for λ P ␣ 10´3 , 10´2 , 10´1 ( and for α within a grid. The red-line represents the desired control-level 1´α over which the theoretical coverage￾probability is guaranteed to remain (see Corollary 38). The empirical coverage-probabilities are mostly above the red-line, which illustrates that the prediction-regions does empir… view at source ↗
Figure 6
Figure 6. Figure 6: Evolution of the computable empirical upper-bound V ´ CrΓpa,up λ;α pXn`1q zCrΓpa,lo λ;α pXn`1q ¯ for the thickness THKΓpa λ;α pXn`1q across 20 repetitions for α “ 0.1 and λ “ 10´2 . As in [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison between the G-EllipsoidCP-regions (in red) and the SplitCP￾regions (in green) in terms of volume relative to that of the OracleCP-regions (see Appendix F.2) for α “ 0.1 with |D| “ n “ 500 and |Dtrain| “ 375. According to [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Evolution of the computable empirical upper-bound V ´ CrΓ,up λ;α pXn`1q zCrΓ,lo λ;α pXn`1q ¯ for the thickness THKΓ λ;α pXn`1q across 20 repetitions for α “ 0.1, and for λ9 ? 1 n . Since λ9 ? 1 n then, 1 λn “ O ´ ? 1 n ¯ , that is to say, by Theorem 35, the theoretical rate should be O ´ ? 1 n ¯ . On average, V ´ CrΓ,up λ;α pXn`1q zCrΓ,lo λ;α pXn`1q ¯ gets smaller as the training-sample size grows larger. … view at source ↗
Figure 9
Figure 9. Figure 9: Evolution of the computable empirical upper-bound V ´ CrΓpa,up λ;α pXn`1q zCrΓpa,lo λ;α pXn`1q ¯ for the thickness THKΓpa λ;α pXn`1q across 20 repetitions for α “ 0.1 and for λ9 ? 1 n . On average, V ´ CrΓpa,up λ;α pXn`1q zCrΓpa,lo λ;α pXn`1q ¯ gets smaller as the training-sample size n gets larger. Additionally, the estimated rate of improvement (represented by the dashed red-line) is about the same as th… view at source ↗

Review history (2 revisions) →

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

35 extracted references · 1 linked inside Pith

  1. [1]

    URL https://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise.laplacian\_kernel.html

    Laplacian\_kernel. URL https://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise.laplacian\_kernel.html

  2. [2]

    URL https://docs.scipy.org/doc/scipy/reference/optimize.minimize-newtoncg.html

    Minimize(method=' Newton-CG ') --- SciPy v1.18.0 Manual . URL https://docs.scipy.org/doc/scipy/reference/optimize.minimize-newtoncg.html

  3. [3]

    Optimization in infinite-dimensional Hilbert spaces

    Alen Alexanderian. Optimization in infinite-dimensional Hilbert spaces. North Carolina State University, Raleigh, NC, USA, 2019

    2019

  4. [4]

    \'A lvarez, Lorenzo Rosasco, and Neil D

    Mauricio A. \'A lvarez, Lorenzo Rosasco, and Neil D. Lawrence. Kernels for Vector-Valued Functions : A Review . Foundations and Trends in Machine Learning , 4 0 (3): 0 195--266, 2012. ISSN 1935-8237, 1935-8245. doi:10.1561/2200000036

    doi:10.1561/2200000036 2012

  5. [5]

    Theory of reproducing kernels

    Nachman Aronszajn. Theory of reproducing kernels. Transactions of the American mathematical society, 68 0 (3): 0 337--404, 1950

    1950

  6. [6]

    Stability of multi-task kernel regression algorithms

    Julien Audiffren and Hachem Kadri. Stability of multi-task kernel regression algorithms. In Asian Conference on Machine Learning , pages 1--16. PMLR, 2013

    2013

  7. [7]

    Learning Theory from First Principles

    Francis Bach. Learning Theory from First Principles . 2024

    2024

  8. [8]

    Bartlett, Philip M

    Peter L. Bartlett, Philip M. Long, G \'a bor Lugosi, and Alexander Tsigler. Benign overfitting in linear regression. Proceedings of the National Academy of Sciences, 117 0 (48): 0 30063--30070, December 2020. ISSN 0027-8424, 1091-6490. doi:10.1073/pnas.1907378117

    doi:10.1073/pnas.1907378117 2020

  9. [9]

    Stability and generalization

    Olivier Bousquet and Andr \'e Elisseeff. Stability and generalization. Journal of machine learning research, 2 0 (Mar): 0 499--526, 2002

    2002

  10. [10]

    Jordan, and Francis Bach

    Sacha Braun, Liviu Aolaritei, Michael I. Jordan, and Francis Bach. Minimum volume conformal sets for multivariate regression. arXiv preprint arXiv:2503.19068, 2025

    arXiv 2025

  11. [11]

    Jordan, and Francis Bach

    Sacha Braun, Eug \`e ne Berta, Michael I. Jordan, and Francis Bach. Multivariate Standardized Residuals for Conformal Prediction , May 2026

    2026

  12. [12]

    Micchelli, Massimiliano Pontil, and Yiming Ying

    Andrea Caponnetto, Charles A. Micchelli, Massimiliano Pontil, and Yiming Ying. Universal multi-task kernels. The Journal of Machine Learning Research, 9: 0 1615--1646, 2008

    2008

  13. [13]

    Two deterministic half-quadratic regularization algorithms for computed imaging

    Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st international conference on image processing, volume 2, pages 168--172. IEEE, 1994

    1994

  14. [14]

    A Unified Comparative Study with Generalized Conformity Scores for Multi-Output Conformal Regression , February 2025

    Victor Dheur, Matteo Fontana, Yorick Estievenart, Naomi Desobry, and Souhaib Ben Taieb. A Unified Comparative Study with Generalized Conformity Scores for Multi-Output Conformal Regression , February 2025

    2025

  15. [15]

    Micchelli, Massimiliano Pontil, and John Shawe-Taylor

    Theodoros Evgeniou, Charles A. Micchelli, Massimiliano Pontil, and John Shawe-Taylor . Learning multiple tasks with kernel methods. Journal of machine learning research, 6 0 (4), 2005

    2005

  16. [16]

    Exact and Approximate Conformal Inference for Multi-Output Regression , June 2024

    Chancellor Johnstone and Eugene Ndiaye. Exact and Approximate Conformal Inference for Multi-Output Regression , June 2024

    2024

  17. [17]

    Leave- One-Out Stable Conformal Prediction , April 2025

    Kiljae Lee and Yuan Zhang. Leave- One-Out Stable Conformal Prediction , April 2025

    2025

  18. [18]

    Optimal rates for regularized conditional mean embedding learning

    Zhu Li, Dimitri Meunier, Mattes Mollenhauer, and Arthur Gretton. Optimal rates for regularized conditional mean embedding learning. Advances in Neural Information Processing Systems, 35: 0 4433--4445, 2022

    2022

  19. [19]

    Towards optimal sobolev norm rates for the vector-valued regularized least-squares algorithm

    Zhu Li, Dimitri Meunier, Mattes Mollenhauer, and Arthur Gretton. Towards optimal sobolev norm rates for the vector-valued regularized least-squares algorithm. Journal of Machine Learning Research, 25 0 (181): 0 1--51, 2024

    2024

  20. [20]

    A vector-contraction inequality for Rademacher complexities, May 2016

    Andreas Maurer. A vector-contraction inequality for Rademacher complexities, May 2016

    2016

  21. [21]

    Copula-based conformal prediction for multi-target regression

    Soundouss Messoudi, S \'e bastien Destercke, and Sylvain Rousseau. Copula-based conformal prediction for multi-target regression. Pattern Recognition, 120: 0 108101, 2021

    2021

  22. [22]

    Ellipsoidal conformal inference for multi-target regression

    Soundouss Messoudi, S \'e bastien Destercke, and Sylvain Rousseau. Ellipsoidal conformal inference for multi-target regression. In Conformal and Probabilistic Prediction with Applications , pages 294--306. PMLR, 2022

    2022

  23. [23]

    Kernels for Multi --task Learning

    Charles Micchelli and Massimiliano Pontil. Kernels for Multi --task Learning . Advances in neural information processing systems, 17, 2004

    2004

  24. [24]

    Micchelli and Massimiliano Pontil

    Charles A. Micchelli and Massimiliano Pontil. On learning vector-valued functions. Neural computation, 17 0 (1): 0 177--204, 2005

    2005

  25. [25]

    Stable conformal prediction sets

    Eugene Ndiaye. Stable conformal prediction sets. In International Conference on Machine Learning , pages 16462--16479. PMLR, 2022

    2022

  26. [26]

    Inductive Conformal Prediction: Theory and Application to Neural Networks

    Harris Papadopoulos. Inductive Conformal Prediction: Theory and Application to Neural Networks . INTECH Open Access Publisher Rijeka, 2008

    2008

  27. [27]

    Approximate full conformal prediction in an RKHS , January 2026

    Davidson Lova Razafindrakoto, Alain Celisse, and J \'e r \^o me Lacaille. Approximate full conformal prediction in an RKHS , January 2026

    2026

  28. [28]

    Saleh and A

    Resve A. Saleh and A. K. Saleh. Statistical properties of the log-cosh loss function used in machine learning. arXiv preprint arXiv:2208.04564, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Learning theory estimates via integral operators and their approximations

    Steve Smale and Ding-Xuan Zhou. Learning theory estimates via integral operators and their approximations. Constructive approximation, 26 0 (2): 0 153--172, 2007

    2007

  30. [30]

    Multi-task regression using minimal penalties

    Matthieu Solnon, Sylvain Arlot, and Francis Bach. Multi-task regression using minimal penalties. The Journal of Machine Learning Research, 13 0 (1): 0 2773--2812, 2012

    2012

  31. [31]

    Hush, and Clint Scovel

    Ingo Steinwart, Don R. Hush, and Clint Scovel. Optimal Rates for Regularized Least Squares Regression . In COLT , pages 79--93, 2009

    2009

  32. [32]

    Algorithmic Learning in a Random World

    Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic Learning in a Random World . Springer Science & Business Media, 2005

    2005

  33. [33]

    Algorithmic Learning in a Random World

    Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic Learning in a Random World. 2 edition, December 2022. ISBN 978-3-031-06648-1. doi:10.10007/978-3-031-06649-8

    doi:10.10007/978-3-031-06649-8 2022

  34. [34]

    A survey on multi-task learning

    Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34 0 (12): 0 5586--5609, 2021

    2021

  35. [35]

    Gradient descent algorithms for quantile regression with smooth approximation

    Songfeng Zheng. Gradient descent algorithms for quantile regression with smooth approximation. International Journal of Machine Learning and Cybernetics, 2: 0 191--207, 2011

    2011