Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection

T0 review · 3 major / 6 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read A label-free balance criterion picks the right calibration for first-shot anomalous sound detection, lifting a frozen system from below baseline to rank four on DCASE 2025.

desk verdict Solid, honest empirical methods paper: a cheap post-hoc calibration+selection layer that actually moves DCASE 2025 numbers, with multi-year bounds that correctly shrink the claim. read the letter →

arxiv 2607.04526 v1 pith:VLOPOEVI submitted 2026-07-05 cs.SD cs.AI

Grach Mkrtchian This is my paper
classification cs.SDcs.AI
keywords anomaloussounddetectiondomaingeneralizationscorecalibrationmodelselectionfirst-shotASDquantileDCASETask2
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

First-shot anomalous sound detection must catch faults on machine types never seen in training, using one fixed threshold, even though the test clip may come from a data-rich source domain or a 10-clip target domain. Organizers have long noted that systems that do well on source domains do poorly on target domains, and that development-set scores do not predict evaluation-set scores. This paper treats both failures as post-hoc calibration and selection problems. It adds a per-domain quantile map that can be blended toward a pooled map by a single strength knob, and ranks candidate knobs by how well they equalize the two domains’ normal-score distributions under cross-validation on training normals alone. On the 2025 challenge the criterion predicts evaluation rank at Spearman 0.91 while development score is uninformative, and guarded selection raises the official score from 55.83 to 61.05—fourth of 35 teams—using only the same training-free pool. Replication shows the same selection trap every year, but the criterion’s predictive power is year-dependent.

What carries the argument

DACO: a post-hoc per-domain quantile calibration layer whose maps are shrunk toward a pooled map by prior strength m (tracing a source/target balance frontier), together with a Kolmogorov–Smirnov domain-balance criterion that measures how closely the two domains’ calibrated normal scores coincide under repeated half-holdout of the training bank.

What would settle it

Release of the DCASE 2026 per-clip evaluation ground truth: the already-frozen configuration chosen by the criterion on the 2026 development set either does or does not outrank the development-score pick and the fixed full-equalization default under the official evaluator.

Watch

Extended reading notes

Core claim

On DCASE 2025 Task 2 a label-free cross-validated domain-balance criterion, computed solely from training normals, rank-predicts official evaluation harmonic-mean performance across a 45-configuration grid at Spearman ρ = +0.91 (family-block bootstrap 95 % CI [+0.83, +0.95]), while development score is essentially uncorrelated (+0.06). Guarded selection by that criterion raises evaluation Ω from 55.83 to 61.05 with an identical training-free system pool, retrospectively fourth of 35 teams.

Load-bearing premise

That making the two domains’ normal scores look the same on training data is a reliable proxy for good harmonic-mean performance on completely unseen machine types and challenge years.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes DACO, a training-free post-hoc layer for first-shot unsupervised anomalous sound detection under the DCASE Task 2 domain-generalization protocol. Over frozen audio embeddings and a kNN backend it adds (i) per-domain quantile calibration of anomaly scores, shrunk toward a pooled map by prior strength m (Eq. 2), which traces a controllable source/target balance frontier, and (ii) a label-free cross-validated Kolmogorov–Smirnov domain-balance criterion C (Eq. 3) computed only on training normals, paired with a coarse development-set viability veto (Algorithm 1). On DCASE 2025 the criterion rank-predicts official evaluation Ω across a 45-configuration grid at Spearman ρ_s = +0.91 (family-block bootstrap 95% CI [+0.83, +0.95]) while development Ω is uninformative (ρ_s = +0.06); guarded selection raises evaluation Ω from 55.83 to 59.34 on that grid and to 61.05 on an extended grid including local-density normalization (retrospectively fourth of 35 teams). Multi-year replication on 2023/2024 bounds the claim: development score is uninformative every year and degenerate configurations recur (and are vetoed), but under family-clustered uncertainty the criterion’s predictive evidence survives only in 2025; a fixed full-equalization default matches or beats criterion selection in both replication years. A DCASE 2026 configuration is pre-registered before per-clip evaluation ground truth is released; all headline numbers are reproduced by the official evaluator.

Significance. If the reported bounds hold, the work supplies a practical, training-free answer to two organizer-documented failure modes—source/target AUC trade-off and unreliable dev→eval model selection—that have persisted across DCASE Task 2 editions. Explicit strengths that raise the contribution above a typical challenge-system paper include: end-to-end official-evaluator verification of every headline number, family-block bootstrap and machine-level jackknife uncertainty, partial-Spearman deconfounding of m and family membership, honest three-year replication that narrows rather than inflates the claim, a frozen prospective 2026 test, and a full public artifact release (code, per-clip scores, checkpoint pins, bootstrap scripts). Even with year-dependent transfer, the demonstration that an operating-point balance signal available from training normals alone can outperform labeled development selection when the configuration pool contains discoverable family structure is useful to the ASD community, and the fixed m=0 full-equalization default is itself a clean practical takeaway.

major comments (3)
  1. [§VI-C, Tables I–III, abstract] Section VI-C (partial Spearman analysis) and Tables I–III: the abstract and lead claims attribute the lift from 55.83 to 61.05 primarily to “criterion-based selection.” On the 45-configuration grid the criterion pick is conf. soft m=0 (59.34); the further jump to 61.05 on the extended grid is the admission of the LDN-ratio family (Table III: 61.04–61.05), which is strong on evaluation but weaker on development and which the criterion merely surfaces as a low-C region rather than ranks finely. After partialling log(m+1) and family membership the residual association falls to +0.71 overall and only +0.21 (p=0.27) inside the 30 conformal configurations. The manuscript already reports these numbers; they should be reflected in the abstract and introduction so that the gain is attributed to coarse family-level discovery (plus the veto that keeps degenerate LDN-diff-K=16 out) rather than to fi
  2. [§VI-F, Table IV, abstract/conclusion] Section VI-F / Table IV: under the same family-block bootstrap used for the 2025 CI, only 2025 survives (2023 CI crosses zero; residual association after partialling family and log m is negative). The paper states this bound, yet the abstract still leads with the 2025 ρ_s = +0.91 and the 61.05 rank-4 result. Because a practitioner cannot know a priori whether the target year contains discoverable family structure, the fixed full-equalization default (conf. soft m=0, k=1)—which matches or beats guarded selection in both replication years and was designated before the E7 replication—should be elevated to a co-primary practical recommendation in the abstract and conclusion, not left mainly in the discussion.
  3. [§IV-E, §VI-D, Tables II and IV] Section IV-E and VI-D: the viability veto was introduced after observing that unguarded criterion selection picks the degenerate LDN-diff-K=16 map on EAT (and in both replication years). The disclosure is transparent, but the veto form (bottom-decile vs. δ-threshold) is a post-hoc meta-choice that directly determines the guarded-selection rows of Tables II and IV. Either fix one veto rule a priori and re-report, or present unguarded criterion, fixed default, and guarded selection as three co-equal rules with the residual meta-selection spread (already quantified on EAT as 58.87–59.75) stated as part of the main result rather than only in the veto disclosure paragraph.
minor comments (6)
  1. [§IV-D, §VI-C] Eq. (3) and the finite-sample floor paragraph in §IV-D note that with only 5 held-out target clips the two-sample KS under identical distributions averages ≈0.36; observed C values (0.49–0.72) must be read relative to that floor and differences ≲0.01 are noise. This “coarse region selector” reading is important for interpreting ρ_s = +0.91 and should be cross-referenced when the correlation is first reported in §VI-C.
  2. [Fig. 3] Figure 3: the reversed C axis (“better is rightward”) is helpful but is stated only in the caption; a brief note in the main text would prevent misreading of panel (b).
  3. [§VI-B, Fig. 2] Section VI-B / Fig. 2: the structure–response relationship between train-side domain separability and target-AUC gain is clean on development machines but weak overall (ρ_s = +0.39, p = 0.15). The paper already cautions this; a single sentence in the abstract or introduction that per-machine balance is not guaranteed would further set expectations.
  4. [§III-B, §VI] Notation: the official score is written both as Ω and as “evaluation score”; a single symbol introduced once (Eq. 1) and used consistently would help. Likewise “conf.” for conformal/quantile calibration is overloaded with conformal prediction literature cited in §II-C.
  5. [Appendix Table VI] Table VI (appendix) is dense; highlighting the machines that drive the ToyCar overshoot and the CoffeeGrinder low target AUC would make the per-machine story easier to scan.
  6. [Abstract] Minor prose: abstract sentence “Criterion-based selection raises … to 61.05—retrospectively fourth of 35 teams” packs two grids (45 vs. 51) into one clause; splitting them as in Table I vs. Table II would match the body.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: criterion (eq. 3) and evaluation Ω are independent by construction; gains are empirical on official external ground truth.

full rationale

The paper's load-bearing claims are empirical rank correlations and selection deltas on DCASE evaluation sets. The domain-balance criterion C (eq. 3) is the mean KS distance between calibrated held-out source and target training-normal score distributions; it uses no anomalies, no test clips, and no evaluation labels. Evaluation Ω (eq. 1) is the official harmonic mean of per-machine AUCs and pAUC computed on held-out test clips with post-challenge ground truth that enters only the metric module (explicitly audited and re-scored by the official evaluator). There is no equation equating C to Ω, no parameter fitted to evaluation scores and then re-presented as a prediction, and no self-citation that supplies a uniqueness theorem or ansatz on which the result rests. Shrinkage m (eq. 2) parametrizes a controllable frontier by design, but the paper selects m via C rather than fitting it to targets; residual analyses after partialling log(m+1) and family membership are reported transparently and do not force the headline numbers. Multi-year replications and the pre-registered 2026 forward test further treat transfer as falsifiable rather than definitional. The construction is therefore self-contained against external benchmarks; any weakness is statistical (year-dependence of the proxy) rather than circular.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central transfer claim rests on empirical protocol assumptions of DCASE first-shot ASD, on treating KS balance of normal scores as a proxy for harmonic-mean evaluation performance, and on a small set of design knobs (m, veto form, CV splits, assignment temperature) rather than on new physical entities. Free parameters are mostly grid-searched or fixed a priori; the load-bearing modeling choices are domain assumptions about exchangeability of LOO scores and usefulness of operating-point balance across machine years.

free parameters (5)
  • prior strength m in quantile shrinkage
    Continuous/grid hyperparameter that interpolates raw ranking and full per-domain equalization (eq. 2); selected by the criterion rather than fixed from theory. Grid {0,10,30,100,300}.
  • viability veto form (bottom-decile or δ-threshold)
    Coarse filter on development Ω introduced after observing degenerate criterion-only picks; δ∈[3,8] and bottom-decile are design choices that change selections on some backbones (VI-D).
  • number of CV splits S and seed
    S=10 with fixed seed used for criterion; authors report stability for S∈{5,10,20} and seed variation ±0.12 Ω, but the statistic remains a free design choice.
  • assignment temperature T
    Set to median of pooled LOO calibration scores; controls soft latent-domain weights and is not derived from a uniqueness theorem.
  • kNN neighborhood size k
    Base scorer hyperparameter grid {1,2,4}; interacts with calibration family in selection.
assumptions (5)
  • domain assumption Leave-one-out kNN distances on training normals are exchangeable enough with test-time scores for quantile maps to approximately equalize per-domain false-positive rates.
    Section IV-C explicitly disclaims conformal guarantees (dependence, interpolation, soft assignment, session shift) and relies on empirical FPR-imbalance reduction (VI-B).
  • ad hoc to paper Kolmogorov–Smirnov distance between held-out calibrated source and target normal scores measures the domain-balanced operating point rewarded by the official harmonic-mean metric.
    Eq. (3) and section IV-D define the selection criterion; finite-sample floor with 5 target clips is acknowledged, so the axiom is that coarse ranking of this statistic still transfers.
  • domain assumption Development-set labels are reliable for coarse viability veto of broken configurations but unreliable for fine ranking of viable ones.
    Algorithm 1 and section IV-E; motivated by organizer non-transfer observation and by observed 8–16 point dev deficits of degenerate maps.
  • domain assumption Soft (or hard) latent-domain weights from embedding-space proximity to source/target banks are accurate enough that mixing per-domain maps improves single-threshold performance.
    Stage 1 (IV-B); paper shows structure–response on some machines and weak overall correlation of separability with target-AUC gain (ρ=+0.39, Fig. 2).
  • standard math Standard two-sample KS and Spearman rank correlation with family-block bootstrap are appropriate uncertainty tools for configuration grids that cluster into method families.
    Used throughout VI-C and VI-F; family-block treatment is what demotes 2023 evidence.
invented entities (2)
  • DACO domain-aware calibration layer (shrinkage quantile maps + soft domain weights) independent evidence
    purpose: Map raw anomaly scores onto a controllable source/target balance frontier without retraining embeddings.
    Composite of known ingredients (group-conditional quantiles, m-estimate shrinkage, proximity weights) packaged as a selectable post-hoc layer for ASD; independent evidence is empirical DCASE performance, not a new physical object.
  • Label-free CV domain-balance criterion C independent evidence
    purpose: Rank scoring configurations from training normals only so model selection does not rely on non-transferring development Ω.
    Defined in eq. (3); claimed novel for first-shot single-threshold ASD. Falsifiable via evaluation rank correlation and the frozen 2026 test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection." pith.science (2026). https://pith.science/paper/VLOPOEVI

@misc{pith2026260704526,
  author       = {Pith},
  title        = {Pith review of: Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLOPOEVI}},
  note         = {Machine review of arXiv:2607.04526}
}
read the original abstract

First-shot anomalous sound detection in DCASE Challenge Task 2 must flag anomalies of unseen machine types with a single threshold, without knowing whether a test clip comes from the data-rich source domain (990 normal training clips) or the data-scarce target domain (10). Two organizer-reported problems remain open: source- and target-domain AUC are negatively correlated across systems, and development-set performance does not predict evaluation-set performance. We address both with a training-free post-hoc layer over frozen audio embeddings: (i) per-domain quantile calibration shrunk toward a pooled map by a prior strength m, tracing a source/target balance frontier, and (ii) a label-free cross-validated domain-balance criterion that ranks candidate configurations from training normals only, paired with a coarse development-labeled viability veto. On DCASE 2025, the criterion rank-predicts the official evaluation score across a 45-configuration grid (Spearman rho = +0.91; family-block bootstrap 95% CI [+0.83, +0.95]) while development score is uninformative (+0.06). Criterion-based selection raises the evaluation score from 55.83 to 59.34 (jackknife CI [2.2, 4.8]) and, on an extended grid, to 61.05 -- retrospectively fourth of 35 teams. Replicating on DCASE 2023 and 2024 bounds the claim: development score is uninformative in all three years and degenerate configurations recur (vetoed every time), but under family-clustered uncertainty the criterion's predictive evidence survives only in 2025; in both replication years a fixed full-equalization default matches or beats criterion-based selection. A DCASE 2026 forward test is frozen before the 2026 evaluation ground truth is released; all headline numbers are reproduced by the official evaluator.

Figures

Figures reproduced from arXiv: 2607.04526 by the authors.

Figure 1
Figure 1. The source/target balance frontier: mean source vs. target AUC over [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Per-machine target-AUC gain of calibration ( [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. What predicts evaluation-set performance? Each point is one of 45 configurations (BEATs backbone); colors denote the five method groups (the soft [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 1 linked inside Pith

  1. [1]

    Description and discussion on DCASE 2025 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    T. Nishida, N. Harada, D. Niizumi, D. Albertini, R. Sannino, S. Pradolini, F. Augusti, K. Imoto, K. Dohi, H. Purohit, T. Endo, and Y . Kawaguchi, “Description and discussion on DCASE 2025 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,” inProc. Detection and Classification of Acous- tic Scenes and Even...

    2025

  2. [2]

    First- shot anomaly sound detection for machine condition monitoring: A domain generalization baseline,

    N. Harada, D. Niizumi, Y . Ohishi, D. Takeuchi, and M. Yasuda, “First- shot anomaly sound detection for machine condition monitoring: A domain generalization baseline,” inProc. European Signal Processing Conference (EUSIPCO), 2023, pp. 191–195

    2023

  3. [3]

    Local density-based anomaly score normalization for domain generalization,

    K. Wilkinghoff, H. Yang, J. Ebbers, F. G. Germain, G. Wichern, and J. Le Roux, “Local density-based anomaly score normalization for domain generalization,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 4642–4652, 2025

    2025

  4. [4]

    Sub-cluster AdaCos: Learning representations for anomalous sound detection,

    K. Wilkinghoff, “Sub-cluster AdaCos: Learning representations for anomalous sound detection,” inProc. International Joint Conference on Neural Networks (IJCNN), 2021, pp. 1–8

    2021

  5. [5]

    Self-supervised learning for anomalous sound detection,

    ——, “Self-supervised learning for anomalous sound detection,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 276–280

    2024

  6. [6]

    AdaProj: Adaptively scaled angular margin subspace projections for anomalous sound detection with auxiliary classification tasks,

    ——, “AdaProj: Adaptively scaled angular margin subspace projections for anomalous sound detection with auxiliary classification tasks,” in Proc. Detection and Classification of Acoustic Scenes and Events (DCASE) Workshop, 2024, pp. 186–190

    2024

  7. [7]

    Handling domain shifts for anomalous sound detection: A review of DCASE-related work,

    K. Wilkinghoff, T. Fujimura, K. Imoto, J. Le Roux, Z.-H. Tan, and T. Toda, “Handling domain shifts for anomalous sound detection: A review of DCASE-related work,” inProc. Detection and Classification of Acoustic Scenes and Events (DCASE) Workshop, 2025, pp. 20–24

    2025

  8. [8]

    BEATs: Audio pre-training with acoustic tokenizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “BEATs: Audio pre-training with acoustic tokenizers,” in Proc. International Conference on Machine Learning (ICML), 2023, pp. 5178–5193

    2023

Show all 35 references
  1. [9]

    EAT: Self- supervised pre-training with efficient audio transformer,

    W. Chen, Y . Liang, Z. Ma, Z. Zheng, and X. Chen, “EAT: Self- supervised pre-training with efficient audio transformer,” inProc. In- ternational Joint Conference on Artificial Intelligence (IJCAI), 2024, pp. 3807–3815

    2024

  2. [10]

    PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

    2020

  3. [11]

    Deep generic representations for domain-generalized anomalous sound detection,

    P. Saengthong and T. Shinozaki, “Deep generic representations for domain-generalized anomalous sound detection,” inProc. IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

    2025

  4. [12]

    Scoring backends matter more than pooling: A systematic study of training-free anomalous sound detection under domain shift,

    J. Zhou and M. Wang, “Scoring backends matter more than pooling: A systematic study of training-free anomalous sound detection under domain shift,” 2026

    2026

  5. [13]

    Keeping the balance: Anomaly score calculation for domain generalization,

    K. Wilkinghoff, H. Yang, J. Ebbers, F. G. Germain, G. Wichern, and J. Le Roux, “Keeping the balance: Anomaly score calculation for domain generalization,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  6. [14]

    Mind the gap: Detecting cluster exits for robust local density-based score normalization in anomalous sound detection,

    K. Wilkinghoff, G. Wichern, J. Le Roux, and Z.-H. Tan, “Mind the gap: Detecting cluster exits for robust local density-based score normalization in anomalous sound detection,”arXiv preprint arXiv:2602.18777, 2026

    2026 arXiv

  7. [15]

    V ovk, A

    V . V ovk, A. Gammerman, and G. Shafer,Algorithmic Learning in a Random World. Springer, 2005

    2005

  8. [16]

    Inductive conformal anomaly detec- tion for sequential detection of anomalous sub-trajectories,

    R. Laxhammar and G. Falkman, “Inductive conformal anomaly detec- tion for sequential detection of anomalous sub-trajectories,”Annals of Mathematics and Artificial Intelligence, vol. 74, pp. 67–94, 2015

    2015

  9. [17]

    Conformal prediction: A gentle introduction,

    A. N. Angelopoulos and S. Bates, “Conformal prediction: A gentle introduction,”Foundations and Trends in Machine Learning, vol. 16, no. 4, pp. 494–591, 2023

    2023

  10. [18]

    Class-conditional conformal prediction with many classes,

    T. Ding, A. N. Angelopoulos, S. Bates, M. I. Jordan, and R. J. Tibshirani, “Class-conditional conformal prediction with many classes,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  11. [19]

    Conformal prediction with conditional guarantees,

    I. Gibbs, J. J. Cherian, and E. J. Candès, “Conformal prediction with conditional guarantees,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 87, no. 4, pp. 1100–1126, 2025

    2025

  12. [20]

    Adaptive conformal anomaly detection with time series foundation models for signal monitoring,

    N. Martinez Gil, F. O’Donncha, W. M. Gifford, N. Zhou, D. C. Patel, and R. Vaculin, “Adaptive conformal anomaly detection with time series foundation models for signal monitoring,” inProc. International Conference on Learning Representations (ICLR), 2026

    2026

  13. [21]

    Between resolution collapse and variance inflation: Weighted conformal anomaly detection in low-data regimes,

    O. Hennhöfer and C. Preisach, “Between resolution collapse and variance inflation: Weighted conformal anomaly detection in low-data regimes,” 2026

    2026

  14. [22]

    Conformal machine learning for reliable anomaly detection in industrial cyber-physical systems,

    S. Yuan, J. Li, C. Wang, and X. Zhang, “Conformal machine learning for reliable anomaly detection in industrial cyber-physical systems,” Reliability Engineering & System Safety, vol. 274, p. 112417, 2026

    2026

  15. [23]

    Set valued predictions for robust domain generalization,

    R. Tsibulsky, D. Nevo, and U. Shalit, “Set valued predictions for robust domain generalization,” inProc. International Conference on Machine Learning (ICML), 2025

    2025

  16. [24]

    Localized conformal prediction: a generalized inference framework for conformal prediction,

    L. Guan, “Localized conformal prediction: a generalized inference framework for conformal prediction,”Biometrika, vol. 110, no. 1, pp. 33–50, 2023

    2023

  17. [25]

    Conformal prediction with local weights: ran- domization enables robust guarantees,

    R. Hore and R. F. Barber, “Conformal prediction with local weights: ran- domization enables robust guarantees,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 87, no. 2, pp. 549–578, 2025

    2025

  18. [26]

    Estimating probabilities: A crucial task in machine learn- ing,

    B. Cestnik, “Estimating probabilities: A crucial task in machine learn- ing,” inProc. 9th European Conference on Artificial Intelligence (ECAI), 1990, pp. 147–149

    1990

  19. [27]

    Gelman and J

    A. Gelman and J. Hill,Data Analysis Using Regression and Multi- level/Hierarchical Models. Cambridge University Press, 2007

    2007

  20. [28]

    Automatic unsupervised outlier model selection,

    Y . Zhao, R. A. Rossi, and L. Akoglu, “Automatic unsupervised outlier model selection,” inAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  21. [29]

    Unsu- pervised model selection for time-series anomaly detection,

    M. Goswami, C. Challu, L. Callot, L. Minorics, and A. Kan, “Unsu- pervised model selection for time-series anomaly detection,” inProc. International Conference on Learning Representations (ICLR), 2023

    2023

  22. [30]

    Covariate shift adap- tation by importance weighted cross validation,

    M. Sugiyama, M. Krauledat, and K.-R. Müller, “Covariate shift adap- tation by importance weighted cross validation,”Journal of Machine Learning Research, vol. 8, no. 35, pp. 985–1005, 2007

    2007

  23. [31]

    Towards accurate model se- lection in deep unsupervised domain adaptation,

    K. You, X. Wang, M. Long, and M. Jordan, “Towards accurate model se- lection in deep unsupervised domain adaptation,” inProc. International Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 97, 2019, pp. 7124–7133

    2019

  24. [32]

    Tune it the right way: Unsupervised validation of domain adaptation via soft neighborhood density,

    K. Saito, D. Kim, P. Teterwak, S. Sclaroff, T. Darrell, and K. Saenko, “Tune it the right way: Unsupervised validation of domain adaptation via soft neighborhood density,” inProc. IEEE/CVF International Con- ference on Computer Vision (ICCV), 2021, pp. 9164–9173

    2021

  25. [33]

    ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, M. Yasuda, and S. Saito, “ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions,” inProc. Detection and Classification of Acoustic Scenes and Events (DCASE) W...

    2021

  26. [34]

    MIMII DG: Sound dataset for mal- functioning industrial machine investigation and inspection for domain generalization task,

    K. Dohi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, M. Yamamoto, Y . Nikaido, and Y . Kawaguchi, “MIMII DG: Sound dataset for mal- functioning industrial machine investigation and inspection for domain generalization task,” inProc. Detection and Classification of Acoustic Sce...

    2022

  27. [35]

    Description and discussion on DCASE 2026 challenge task 2: Noise-aware unsupervised anomalous sound detection for machine condition monitoring,

    T. Nishida, N. Harada, D. Takeuchi, D. Niizumi, K. Imoto, K. Dohi, H. Purohit, T. Endo, and Y . Kawaguchi, “Description and discussion on DCASE 2026 challenge task 2: Noise-aware unsupervised anomalous sound detection for machine condition monitoring,” 2026

    2026

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.