pith. sign in

arxiv: 2606.29403 · v1 · pith:NWTXOFGLnew · submitted 2026-06-28 · 📊 stat.ML · cs.AI· cs.LG

Self-Organized Conformal Prediction: Reducing Regional Coverage Gaps with Unsupervised Group Discovery

Louis Berthier , Ahmed Shokry , Maxime Moreaud , Guillaume Ramelet , Aymeric Dieuleveut This is my paper

Pith reviewed 2026-06-30 02:15 UTC · model grok-4.3

classification 📊 stat.ML cs.AIcs.LG
keywords conformal predictionself-organizing mapregional coveragegroup discoverylocal calibrationprediction setsunsupervised learning
0
0 comments X

The pith

Self-Organized Conformal Prediction discovers input groups with a self-organizing map to reduce regional coverage gaps while preserving validity.

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

Conformal prediction guarantees marginal coverage but can mask undercoverage in some input regions when calibration pools all data together. The paper introduces Self-Organized Conformal Prediction that learns groups directly from input geometry using a self-organizing map. At test time it retrieves a local calibration buffer from the best-matching cell or a fixed neighborhood in the map. This keeps the original predictor and score unchanged and applies to both regression and classification. On eight benchmarks the method reduces the weighted regional coverage gap on seven datasets while increasing average prediction-set size by only 6.2 percent.

Core claim

SOCP discovers input-space groups with a Self-Organizing Map and at test time draws a local calibration buffer from the query's best-matching unit cell or a fixed grid neighborhood. The same retrieval rule applies across tasks and data types. It gives exact validity for BMU-cell retrieval and fixed retrieved-set validity for neighborhood buffers, with central-cell validity holding up to a Kolmogorov-Smirnov bias term. On eight regression and classification benchmarks it reduces the weighted regional coverage gap on seven datasets (mean paired change -7.1 percent) for a mean prediction-set size increase of 6.2 percent.

What carries the argument

Self-Organizing Map that partitions the input space into cells used to retrieve local calibration buffers at test time.

If this is right

  • Exact validity holds when calibration examples are drawn only from the best-matching unit cell.
  • Neighborhood buffers deliver fixed retrieved-set validity.
  • A split-routed extension recovers fixed retrieved-set validity conditional on the routing split.
  • The weighted regional coverage gap decreases on seven of eight benchmarks with only a 6.2 percent mean increase in prediction-set size.
  • The method works without supervised partitions or predictor retraining on both tabular features and image embeddings.

Where Pith is reading between the lines

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

  • Local calibration buffers could support more reliable use in safety-critical settings where certain input regions carry higher decision risk.
  • Because groups are learned unsupervised the approach could extend to data streams where labeled subgroups are unavailable.
  • Alternative unsupervised partitioning methods might produce comparable reductions in regional gaps if they also align with coverage heterogeneity.

Load-bearing premise

The self-organizing map discovers cells whose local data distributions align with regions of differing coverage behavior and the Kolmogorov-Smirnov bias term remains small enough that central-cell validity for neighborhood retrieval is practically useful.

What would settle it

Running the eight benchmarks and finding that the weighted regional coverage gap fails to decrease on most datasets or that observed coverage deviates from the stated validity guarantees by more than the KS bias term.

Figures

Figures reproduced from arXiv: 2606.29403 by Ahmed Shokry, Aymeric Dieuleveut, Guillaume Ramelet, Louis Berthier, Maxime Moreaud.

Figure 1
Figure 1. Figure 1: Overview of the SOCP pipeline. A pre-trained SOM partitions the input space into [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Weighted coverage gap against mean prediction size across the benchmark suite. Each [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Retrieved SO-SCP buffer-size distributions across datasets. The dashed line in each [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Strategy-level metrics across the benchmark suite. Each panel summarizes the ten seeds for [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Calibration and test hit maps on the learned SOM grids. The top row reports calibration [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: SOM distance maps for one example seed per dataset. Each cell is shaded by the average [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-cell empirical coverage for one example seed per dataset. Rows are datasets and [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-cell pass/fail coverage diagnostic for one example seed per dataset. Green cells meet [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Per-cell prediction error for one example seed per dataset. Regression cells report mean [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: reports the 90th percentile of the realized test-set score in each SOM cell, organized by score family. SCP, LCP, and SO-SCP share the absolute-residual or softmax score, while CQR and SO-CQR share the CQR score, so the per-cell aggregation depends only on the family. SCP score Bike Sharing Bio California Housing CIFAR-10 Concrete Covertype MNIST MPG CQR score - - - 50 100 −100 0 2.5 5.0 7.5 0.0 2.5 1 2 0… view at source ↗
Figure 11
Figure 11. Figure 11: Per-cell average conformal threshold for each method, on one example seed per dataset. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Per-cell average prediction-output size for each method, on one example seed per dataset. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Plug-in empirical KS bridge bias εbk(A) from calibration scores, for one example seed per dataset. SO-CQR panels are blank for classification datasets. High values mark cells where the bridge bound in Lemma A.2 pays a large term, so cell-conditional coverage may sit further below the retrieved-set guarantee. SO-SCP Bike Sharing Bio California Housing CIFAR-10 Concrete Covertype MNIST MPG SO-CQR - - - 0.0 … view at source ↗
Figure 14
Figure 14. Figure 14: Mixture diagnostic dbmix(k, A) from calibration scores, for one example seed per dataset. SO-CQR panels are blank for classification datasets. The diagnostic is the direct KS distance between the central-cell empirical score CDF and the pooled retrieved-set CDF. MAE / top-1 error Bike Sharing Bio California Housing CIFAR-10 Concrete Covertype MNIST MPG Score p90 Width / set size KS epsilon KS mix 25 50 50… view at source ↗
Figure 15
Figure 15. Figure 15: Compact SO-SCP diagnostic grid for one example seed per dataset. Rows show prediction [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
read the original abstract

Conformal prediction guarantees marginal coverage, but pooled calibration averages over heterogeneous regions and can mask regional undercoverage in safety-critical subgroups. We introduce Self-Organized Conformal Prediction (SOCP), a calibration scheme that discovers input-space groups with a Self-Organizing Map (SOM) and, at test time, draws a local calibration buffer from the query's best-matching unit (BMU) cell or a fixed grid neighborhood. The same retrieval rule applies to regression and classification tasks across tabular features and image embeddings, leaving the predictor and nonconformity score untouched. SOCP gives exact validity for BMU-cell retrieval and fixed retrieved-set validity for neighborhood buffers; central-cell validity for neighborhood retrieval holds up to a Kolmogorov-Smirnov (KS) bias term. A split-routed extension recovers fixed retrieved-set validity conditional on the routing split. On eight regression and classification benchmarks, SO-SCP reduces the weighted regional coverage gap on $7/8$ datasets (mean paired change $-7.1\%$) for a mean prediction-set size increase of $6.2\%$, with negligible overhead on the largest six datasets; SO-CQR yields smaller gains, since quantile regression already absorbs much of the heterogeneity. By learning groups directly from the input geometry, SOCP provides group-local calibration with exact fixed-group guarantees and approximate central-cell guarantees, without supervised partitions or predictor retraining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The paper introduces Self-Organized Conformal Prediction (SOCP), which uses a Self-Organizing Map (SOM) to discover input-space groups from geometry alone and retrieves local calibration sets from the query's best-matching unit (BMU) cell or a fixed grid neighborhood. It claims exact validity for BMU-cell retrieval, fixed retrieved-set validity for neighborhood buffers, and central-cell validity for neighborhoods up to a Kolmogorov-Smirnov (KS) bias term (with a split-routed variant recovering conditional fixed-set validity). On eight regression and classification benchmarks the method reduces the weighted regional coverage gap on 7/8 datasets (mean paired change -7.1%) at a mean prediction-set size increase of 6.2%, without altering the underlying predictor or nonconformity score.

Significance. If the stated validity properties hold and the observed gap reductions can be attributed to SOM cells aligning with coverage heterogeneity (rather than input geometry alone), the approach would provide a practical, unsupervised route to group-local calibration that preserves exact or approximate guarantees while remaining applicable across tabular and embedding-based tasks. The negligible overhead on large datasets and the fact that gains are smaller when quantile regression already absorbs heterogeneity are also positive features. The lack of a bound on the KS term or empirical distribution of the statistic across benchmarks, however, limits the strength of the approximate-central-cell claim.

major comments (3)
  1. [Abstract] Abstract (guarantees paragraph): the central claim of practically useful approximate central-cell validity for neighborhood retrieval rests on the KS bias term remaining small relative to the reported 7.1% gap reduction, yet the manuscript supplies neither a theoretical bound on this term nor the empirical distribution of the KS statistic across the eight benchmarks; without these the attribution of the empirical improvement to the claimed validity properties cannot be assessed.
  2. [Abstract] Abstract and method description: the procedure relies on the unsupervised SOM discovering cells whose local nonconformity distributions differ meaningfully from the global pool and from each other; no experiment or diagnostic is reported that tests whether the discovered partitions align with coverage heterogeneity rather than merely reflecting input-space geometry, which is load-bearing for interpreting the 7/8-dataset improvement as evidence for the method's validity properties.
  3. [Abstract] Abstract (empirical results): the reported mean paired change of -7.1% and 6.2% size increase are presented without accompanying per-dataset tables, standard errors, or ablation on the free parameters (SOM grid size, neighborhood radius), making it impossible to judge robustness or to separate the contribution of the validity guarantees from post-hoc tuning.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below, indicating the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract (guarantees paragraph): the central claim of practically useful approximate central-cell validity for neighborhood retrieval rests on the KS bias term remaining small relative to the reported 7.1% gap reduction, yet the manuscript supplies neither a theoretical bound on this term nor the empirical distribution of the KS statistic across the eight benchmarks; without these the attribution of the empirical improvement to the claimed validity properties cannot be assessed.

    Authors: We agree that the empirical distribution of the KS statistic is needed to evaluate the practical size of the bias term. Deriving a general theoretical bound without strong distributional assumptions is not feasible while preserving the method's broad applicability. In the revision we will add the per-benchmark KS values (and their relation to the observed gap reductions) so readers can directly assess the term. revision: yes

  2. Referee: [Abstract] Abstract and method description: the procedure relies on the unsupervised SOM discovering cells whose local nonconformity distributions differ meaningfully from the global pool and from each other; no experiment or diagnostic is reported that tests whether the discovered partitions align with coverage heterogeneity rather than merely reflecting input-space geometry, which is load-bearing for interpreting the 7/8-dataset improvement as evidence for the method's validity properties.

    Authors: This observation is correct and highlights an important interpretive gap. While the consistent performance gains provide indirect support, a direct diagnostic is warranted. We will include in the revision an analysis that quantifies how the discovered cells differ in nonconformity-score distributions from the global pool (e.g., average cell-to-global KS distances) and will report the resulting regional coverage gaps within cells. revision: yes

  3. Referee: [Abstract] Abstract (empirical results): the reported mean paired change of -7.1% and 6.2% size increase are presented without accompanying per-dataset tables, standard errors, or ablation on the free parameters (SOM grid size, neighborhood radius), making it impossible to judge robustness or to separate the contribution of the validity guarantees from post-hoc tuning.

    Authors: We concur that the current aggregate reporting limits assessment of robustness. The revised manuscript will contain a supplementary table listing per-dataset metrics together with bootstrap standard errors, and we will add an ablation study varying SOM grid size and neighborhood radius to demonstrate stability of the reported gains. revision: yes

Circularity Check

0 steps flagged

No significant circularity; validity claims follow from standard CP applied to retrieved sets

full rationale

The derivation applies exchangeability-based conformal coverage to calibration subsets retrieved by BMU or neighborhood rules from an unsupervised SOM. Exact BMU-cell validity is the direct consequence of running split conformal prediction on the cell's own calibration points; neighborhood fixed-set validity likewise follows from treating the retrieved buffer as a fixed calibration set. The KS bias term is explicitly introduced as an approximation bound rather than asserted to be zero or fitted. No parameter is tuned to the coverage target and then relabeled as a prediction, no self-citation supplies a uniqueness theorem, and no ansatz is smuggled via prior work. The empirical benchmark results are presented as separate validation of practical effect size, not as the source of the theoretical guarantees. The procedure is therefore self-contained against external conformal-prediction benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The method rests on standard conformal exchangeability plus the modeling assumption that SOM geometry captures coverage heterogeneity; two free hyperparameters control the map and retrieval neighborhood.

free parameters (2)
  • SOM grid size
    Determines number of discovered cells; chosen to balance granularity and sample size per cell.
  • Neighborhood radius for buffer retrieval
    Controls how many adjacent cells contribute to the local calibration set.
axioms (2)
  • domain assumption Data points are exchangeable conditional on the calibration set for marginal coverage
    Required for all conformal validity statements; invoked in the guarantee paragraphs.
  • domain assumption SOM topology reflects regions of heterogeneous nonconformity behavior
    Necessary for local buffers to improve regional coverage; stated implicitly in the group-discovery motivation.

pith-pipeline@v0.9.1-grok · 5801 in / 1292 out tokens · 37420 ms · 2026-06-30T02:15:36.893336+00:00 · methodology

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

40 extracted references · 30 canonical work pages · 7 internal anchors

  1. [1]

    A. N. Angelopoulos and S. Bates. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. Dec. 2022. doi: 10.48550/arXiv.2107.07511

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2107.07511 2022

  2. [2]

    Bairaktari, J

    K. Bairaktari, J. Wu, and Z. S. Wu. Kandinsky Conformal Prediction: Beyond Class- and Covariate-Conditional Coverage. July 2025. doi: 10.48550/arXiv.2502.17264

    work page doi:10.48550/arxiv.2502.17264 2025

  3. [3]

    R. F. Barber, E. J. Candès, A. Ramdas, and R. J. Tibshirani. The limits of distribution-free conditional predictive inference. Apr. 2020. doi: 10.48550/arXiv.1903.04684

    work page doi:10.48550/arxiv.1903.04684 2020

  4. [4]

    Berthier, A

    L. Berthier, A. Shokry, M. Moreaud, G. Ramelet, and E. Moulines. Torchsom: The Reference PyTorch Library for Self-Organizing Maps. Oct. 2025. doi: 10.48550/arXiv.2510.11147

    work page doi:10.48550/arxiv.2510.11147 2025

  5. [5]

    Bhattacharyya and R

    A. Bhattacharyya and R. F. Barber. Group-weighted conformal prediction.Electronic Journal of Statistics, 20(1):1171–1199, Jan. 2026. ISSN 1935-7524, 1935-7524. doi: 10.1214/26-EJS2506

    work page doi:10.1214/26-ejs2506 2026

  6. [6]

    Blackard

    J. Blackard. Covertype. UCI Machine Learning Repository, 1998. doi: 10.24432/C50K5N

    work page doi:10.24432/c50k5n 1998

  7. [7]

    Conditional Coverage Diagnostics for Conformal Prediction

    S. Braun, D. Holzmüller, M. I. Jordan, and F. Bach. Conditional Coverage Diagnostics for Conformal Prediction. Dec. 2025. doi: 10.48550/arXiv.2512.11779

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2512.11779 2025

  8. [8]

    Campbell, S

    T. Campbell, S. Syed, C.-Y . Yang, M. I. Jordan, and T. Broderick. Local Exchangeability. July

  9. [9]

    doi: 10.48550/arXiv.1906.09507

    work page doi:10.48550/arxiv.1906.09507 1906

  10. [10]

    T. Ding, A. Angelopoulos, S. Bates, M. Jordan, and R. J. Tibshirani. Class-Conditional Conformal Prediction with Many Classes.Advances in Neural Information Processing Systems, 36:64555–64576, Dec. 2023

    2023

  11. [11]

    UCI Machine Learning Repository

    H. Fanaee-T. Bike Sharing. UCI Machine Learning Repository, 2013. doi: 10.24432/C5W894

    work page doi:10.24432/c5w894 2013

  12. [12]

    arXiv preprint arXiv:2305.12616 , year=

    I. Gibbs, J. J. Cherian, and E. J. Candès. Conformal Prediction With Conditional Guarantees. Sept. 2024. doi: 10.48550/arXiv.2305.12616

    work page doi:10.48550/arxiv.2305.12616 2024

  13. [13]

    L. Guan. Localized Conformal Prediction: A Generalized Inference Framework for Conformal Prediction. Feb. 2022. doi: 10.48550/arXiv.2106.08460

    work page doi:10.48550/arxiv.2106.08460 2022

  14. [14]

    X. Han, Z. Tang, J. Ghosh, and Q. Liu. Split Localized Conformal Prediction. Feb. 2023. doi: 10.48550/arXiv.2206.13092

    work page doi:10.48550/arxiv.2206.13092 2023

  15. [15]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. Dec

  16. [16]

    doi: 10.48550/arXiv.1512.03385

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1512.03385

  17. [17]

    Concrete Compressive Strength

    I-Cheng Yeh. Concrete Compressive Strength. UCI Machine Learning Repository, 1998. doi: 10.24432/C5PK67. 10

    work page doi:10.24432/c5pk67 1998

  18. [18]

    J. N. Kaur, M. I. Jordan, and A. Alaa. Conformal Prediction Sets with Improved Conditional Coverage using Trust Scores. Feb. 2025. doi: 10.48550/arXiv.2501.10139

    work page doi:10.48550/arxiv.2501.10139 2025

  19. [19]

    Kelley Pace and R

    R. Kelley Pace and R. Barry. Sparse spatial autoregressions.Statistics & Probability Letters, 33 (3):291–297, May 1997. ISSN 0167-7152. doi: 10.1016/S0167-7152(96)00140-X

    work page doi:10.1016/s0167-7152(96)00140-x 1997

  20. [20]

    T. Kohonen. Self-organized formation of topologically correct feature maps.Biological Cybernetics, 43(1):59–69, Jan. 1982. ISSN 1432-0770. doi: 10.1007/BF00337288

    work page doi:10.1007/bf00337288 1982

  21. [21]

    T. Kohonen. The self-organizing map.Proceedings of the IEEE, 78(9):1464–1480, Sept. 1990. ISSN 1558-2256. doi: 10.1109/5.58325

    work page doi:10.1109/5.58325 1990

  22. [22]

    Kohonen.Self-Organizing Maps, volume 30 ofSpringer Series in Information Sciences

    T. Kohonen.Self-Organizing Maps, volume 30 ofSpringer Series in Information Sciences. Springer, Berlin, Heidelberg, 2001. ISBN 978-3-540-67921-9 978-3-642-56927-2. doi: 10. 1007/978-3-642-56927-2

    2001

  23. [23]

    Krizhevsky

    A. Krizhevsky. Learning Multiple Layers of Features from Tiny Images.University of Toronto, May 2012

    2012

  24. [24]

    Lecun, L

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, Nov. 1998. ISSN 1558-2256. doi: 10.1109/5.726791

    work page doi:10.1109/5.726791 1998

  25. [25]

    Lei and L

    J. Lei and L. Wasserman. Distribution-free Prediction Bands for Non-parametric Regression. Journal of the Royal Statistical Society Series B: Statistical Methodology, 76(1):71–96, Jan

  26. [26]

    doi: 10.1111/rssb.12021

    ISSN 1369-7412. doi: 10.1111/rssb.12021

    work page doi:10.1111/rssb.12021

  27. [27]

    J. Lei, M. G’Sell, A. Rinaldo, R. J. Tibshirani, and L. Wasserman. Distribution-Free Predictive Inference For Regression. Mar. 2017. doi: 10.48550/arXiv.1604.04173

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1604.04173 2017

  28. [28]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. Sept. 2020. doi: 10.48550/arXiv.1802.03426

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1802.03426 2020

  29. [29]

    Papadopoulos, K

    H. Papadopoulos, K. Proedrou, V . V ovk, and A. Gammerman. Inductive Confidence Machines for Regression. InMachine Learning: ECML 2002, pages 345–356, Berlin, Heidelberg, 2002. Springer. ISBN 978-3-540-36755-0. doi: 10.1007/3-540-36755-1_29

    work page doi:10.1007/3-540-36755-1_29 2002

  30. [30]

    Plassier, A

    V . Plassier, A. Fishkov, V . Dheur, M. Guizani, S. B. Taieb, M. Panov, and E. Moulines. Rectifying Conformity Scores for Better Conditional Coverage. Aug. 2025. doi: 10.48550/ arXiv.2502.16336

    work page arXiv 2025

  31. [31]

    R. Quinlan. Auto MPG. UCI Machine Learning Repository, 1993. doi: 10.24432/C5859H

    work page doi:10.24432/c5859h 1993

  32. [32]

    P. Rana, H. Sharma, M. Bhattacharya, and A. Shukla. Physicochemical Properties of Protein Structure. UCI Machine Learning Repository, Oct. 2015. doi: 10.24432/C5QW3H

    work page doi:10.24432/c5qw3h 2015

  33. [33]

    Conformalized Quantile Regression

    Y . Romano, E. Patterson, and E. J. Candès. Conformalized Quantile Regression. May 2019. doi: 10.48550/arXiv.1905.03222

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1905.03222 2019

  34. [34]

    Romano, M

    Y . Romano, M. Sesia, and E. J. Candès. Classification with Valid and Adaptive Coverage. June

  35. [35]

    doi: 10.48550/arXiv.2006.02544

    work page doi:10.48550/arxiv.2006.02544 2006

  36. [36]

    ImageNet Large Scale Visual Recognition Challenge

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. Jan. 2015. doi: 10.48550/arXiv.1409.0575

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1409.0575 2015

  37. [37]

    Shafer and V

    G. Shafer and V . V ovk. A Tutorial on Conformal Prediction.Journal of Machine Learning Research, 9(12):371–421, 2008. ISSN 1533-7928

    2008

  38. [38]

    V . V ovk. Conditional Validity of Inductive Conformal Predictors. InProceedings of the Asian Conference on Machine Learning, pages 475–490. PMLR, Nov. 2012

    2012

  39. [39]

    V ovk, D

    V . V ovk, D. Lindsay, I. Nouretdinov, and A. Gammerman. Mondrian Confidence Machine. Technical report, Royal Holloway University of London, 2003

    2003

  40. [40]

    V ovk, A

    V . V ovk, A. Gammerman, and G. Shafer.Algorithmic Learning in a Random World. Springer, New York, NY , USA, 2005. doi: 10.1007/b106715. 11 A Appendix A.1 Deferred methodological details This appendix collects the proof tools (Mondrian validity, the KS bridge, and tower-property transfer), the prototype-KNN extension, and the split-routed buffer that comp...

    work page doi:10.1007/b106715 2005