Pith. sign in

REVIEW 4 major objections 8 minor 90 references

Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces the suitability filter, a statistical test that detects when a classifier's accuracy on unlabeled user data falls more than a chosen margin below its accuracy on labeled test data.

desk verdict A useful deployment-monitoring heuristic with a real proof bug: the headline FPR guarantee has a sign error, and the experiments deliberately bypass the corrected margin. read the letter →

arxiv 2505.22356 v1 pith:24SCSPPC submitted 2025-05-28 cs.LG cs.AIcs.CYstat.ML

Angéline Pouget , Mohammad Yaghini , Stephan Rabanser , Nicolas Papernot This is my paper
classification cs.LGcs.AIcs.CYstat.ML
keywords suitabilityfiltercovariateshiftnon-inferioritytestingunsupervisedaccuracyestimationmodeldeploymentpredictioncorrectnesscalibrationWILDSbenchmark
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

The paper proposes the suitability filter, a statistical procedure that tells a model user whether a classifier's accuracy on their unlabeled data has dropped more than a chosen margin below its accuracy on the labeled test set. The aim is to make deployment-time monitoring possible in settings where ground-truth labels for user data are unavailable, delayed, or impossible to obtain. The filter estimates a per-sample probability of correct prediction from model-output signals, averages these estimates over test and user data, and runs a one-sided non-inferiority test to decide SUITABLE or INCONCLUSIVE. If the estimates are δ-calibrated on both distributions, the paper proves the false positive rate of the end-to-end decision is bounded by the test's significance level, after a margin adjustment that absorbs the calibration errors. On roughly 29,000 experiments over three real-world shifted datasets, the filter flags accuracy deterioration beyond 3% without labels.

What carries the argument

The engine is a per-sample prediction correctness estimator C, a logistic regressor trained on a labeled holdout set from the source distribution that maps a vector of suitability signals — maximum softmax confidence, softmax entropy, top-two logit difference, logit energy, and similar model-output features — to a probability pc(x) that the classifier is correct on x. These probabilities are averaged over the test and user datasets, and the two means are compared through a one-sided Welch's t-test in a non-inferiority setup with null hypothesis µtarget < µsource − m. The statistical guarantee rests on the δ-calibration condition, which says the mean of pc tracks true accuracy up to an error δ, and on the adjusted margin m' = m + δsource − δtarget.

What would settle it

Take any deployment where a small labeled sample from the target is available, compute the actual accuracy drop and the mean of pc on unlabeled target data, and check whether the false positive rate of the filter exceeds the chosen α once the accuracy drop is larger than the margin. Concretely, on a WILDS fold with known accuracy drop (for instance a geographically shifted FMoW fold), one can measure δtarget = mean(pc) − Acc(M, Du); if |δtarget| is large relative to m and the filter still returns SUITABLE at rate above α, the bounded-FPR claim fails. A direct test: estimate δsource and δtarget from labeled target samples, apply the corrected margin m', and verify empirically that the false positive rate stays at or below α.

Watch

Extended reading notes

Core claim

The central claim is that the suitability filter reliably detects performance deterioration due to covariate shift without access to labels on the target data. Formally, the paper defines suitability as Acc(M, Du) >= Acc(M, Dtest) - m and constructs a filter that outputs SUITABLE only when a non-inferiority test rejects the null hypothesis that the mean estimated correctness on user data is below the mean on test data by more than the margin. The end-to-end false positive rate is bounded by α provided the prediction-correctness estimator is δ-calibrated on both source and target, with the margin corrected to m' = m + δsource − δtarget. In experiments on FMoW-WILDS, RxRx1-WILDS, and CivilComments-WILDS, including temporal, geographical, and subpopulation shifts, the filter detects deterioration beyond roughly 3% with near-perfect accuracy and keeps false positive rates low at α = 0.05.

Load-bearing premise

The correctness estimator, trained only on labeled source data, must remain calibrated on the target distribution so that the average of its probability estimates tracks the classifier's true accuracy on user data; the paper itself notes that without assumptions about source and target, such calibration is impossible to guarantee.

Editorial extensions

If this is right

  • Deployment teams can monitor accuracy degradation on unlabeled user data without waiting for ground truth labels, provided calibration holds.
  • Model providers can write auditable service-level agreements: the filter's SUITABLE decision certifies, up to significance level α, that accuracy on user data is within margin m of test accuracy.
  • The user can pick the margin m and significance α, making the trade-off between accepting mildly degraded models and rejecting acceptable ones explicit.
  • The framework extends beyond detection: substituting an equivalence test gives a tool for checking that performance on different groups or domains is statistically comparable within a margin.
  • Repeated application with multiple-testing or sequential-testing corrections turns the one-shot decision into a continuous monitoring procedure.

Reading between the lines

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

  • A natural next step the paper leaves implicit: use the empirical gap between mean pc and observed accuracy on small labeled target samples to detect calibration drift itself, turning the margin adjustment into an online monitoring signal.
  • The framework's reliance on calibrated correctness estimates suggests that its success on a given deployment will track how well the chosen signals separate correct from incorrect predictions on that target; signals like energy and logit max already carry most of the weight in the reported ablations.
  • One testable extension is to couple the filter with disagreement-based signals from a small ensemble, which may preserve calibration under shifts that corrupt a single model's confidence ordering.
  • In adversarial settings the filter's assumption of representative user data fails by construction; a deployment concerned with manipulation would need to pair the filter with data-integrity checks rather than rely on the statistical guarantee alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper proposes the suitability filter, a framework to decide whether a classifier M can be deployed on an unlabeled user dataset Du without the accuracy on Du falling more than a margin m below the accuracy on a labeled test set Dtest. A logistic-regression estimator C, trained on a labeled source holdout Dsf, maps twelve hand-selected suitability signals (confidence, logit, and energy statistics) to a per-sample prediction correctness probability pc(x). The filter compares the means of pc over Dtest and Du via a one-sided Welch's t-test of non-inferiority with null hypothesis H0: mu_target < mu_source - m; rejecting H0 yields SUITABLE, failing to reject yields INCONCLUSIVE. The authors derive a false-positive-rate bound (Corollary 4.4) under a delta-calibration condition on both the source and target distributions, propose a margin-adjustment rule m' = m + Delta_test - Delta_u (Eq. 12) using empirical accuracy-estimation errors, and evaluate the method on FMoW-WILDS, RxRx1-WILDS, and CivilComments-WILDS across roughly 29k experiments for FMoW alone, reporting ROC/PR AUC, decision accuracy, and FPR at alpha = 0.05.

Significance. If the framework and its formal guarantee were fully sound, this would be a useful contribution: the deployment problem (label-free accuracy monitoring under covariate shift) is genuine; the non-inferiority framing with user-specified margins and significance levels is principled and modular; and the empirical evaluation is extensive, covering three WILDS benchmarks, multiple architectures, and ablations over signals, sample sizes, margins, and classifiers, with code released. The observed FPRs in Table 1 (0.000-0.031 at alpha = 0.05) and near-1 ROC AUCs, plus the monotone sensitivity curve in Figure 4, give credible evidence that the implemented procedure detects accuracy degradation in the tested settings. However, the headline formal guarantee is load-bearing and currently compromised: the proof of Lemma 4.3 contains a sign error that propagates into the margin adjustment of Corollary 4.4; the printed test statistic and p-value recipe in Appendix A.1.1 do not define a valid directional non-inferiority rule as written; and Section 5.1 deliberately sets m' = m, so the experiments never exercise the margin-adjusted procedure whose guarantee is claimed.

major comments (4)
  1. [Lemma 4.3 and Corollary 4.4 (Appendix A.1.2)] Sign inconsistency in the central guarantee. Definition 4.1 states P[M(x)=y | pc(x)=v] = v + eps(v) with the integral of eps(v) fc(v) equal to delta. The proof of Lemma 4.3 in Appendix A.1.2 then correctly derives P[M(x)=y] = E[pc(x)] + delta (Eq. 17), but Eq. (19) concludes E[pc(x)] - P[M(x)=y] = delta; the correct conclusion is -delta. This algebraic slip propagates: Eq. (24) of the proof of Corollary 4.4 (mu_source - Acc_source = delta_source and mu_target - Acc_target = delta_target) inherits the wrong sign, and the margin adjustment m' = m + delta_source - delta_target set in Corollary 4.4 therefore has the opposite sign of what Definition 4.1 implies. Under Definition 4.1 as written, the correct adjustment is m' = m - delta_source + delta_target, or equivalently the sign of eps in Eq. (8) should be flipped. Notably, the practical adjustment in Eq. (12), m' = m + Delta_test - Delta_u, is consistent with the corrected lemma rather than with the printed Lemma 4.3, since the empirical estimator Delta has expectation -delta under Definition 4.1. As published, Corollary 4.4 does not establish the advertised false-positive bound with the stated m', so the theorem, its proof, and Eq. (12) must be brought to one consistent sign convention.
  2. [Appendix A.1.1, Eq. (13)] As printed, the decision rule is not a valid one-sided non-inferiority test. With the margin-adjusted mean defined as mu_hat_Du = mean(pc[Du]) + m, Eq. (13) sets t = (mu_hat_Dtest - mu_hat_Du)/SE; under the stated H0 (mu_target < mu_source - m) this statistic has positive expectation, so the correct rejection region is the left tail. The accompanying description says the two-sided p-value is halved, which yields the tail pointing in the direction of the observed statistic; in the case of primary concern, where the user data is far worse than the test data, the printed statistic is large and positive and p/2 becomes small, so the procedure as written would reject H0 and return SUITABLE precisely when the model is unsuitable. The standard non-inferiority statistic would be t' = (mean(pc[Du]) + m - mean(pc[Dtest]))/SE with an upper-tail rejection region and the one-sided p-value computed in that tail. Because Table 1 reports ROC AUC near one and Figure 4 shows SUITABLE decisions decreasing as the accuracy difference becomes negative, the released implementation evidently performs a correct directional test; the written specification of Eq. (13) and the p-value recipe in Appendix A.1.1 must nevertheless be corrected, since the method is defined by that specification and the guarantee in Corollary 4.4 inherits it.
  3. [Section 5.1, Section 4.4, Remark 4.5] The experimental section does not exercise the guarantee the paper advertises, and the condition under which the guarantee holds is unverifiable in the intended deployment setting. Section 5.1 states that the margin adjustment is not applied (m' = m) because obtaining a reliable approximation for delta_target requires access to a small labeled user dataset, so the false positive rates in Table 1 (e.g., 0.027 +/- 0.033 on FMoW ID, 0.018 +/- 0.017 on FMoW OOD) are not governed by Corollary 4.4; they are raw empirical rates under an uncorrected margin. In addition, Corollary 4.4 requires C to be delta-calibrated on D_target, a condition the paper itself concedes is impossible to achieve without specific assumptions on D_source versus D_target and unverifiable without target labels (Section 4.2, citing David et al. 2010, and Remark 4.5). The practical substitute in Eq. (12), which estimates the delta terms from a small labeled sample D_hat_u, carries sampling variability that Corollary 4.4 does not account for, so the deployed procedure has no quantitative false-positive control either. I recommend (i) restating the abstract and contribution 3 so that bounded false positive rate is explicitly conditional on delta-calibration or on a labeled calibration sample from the deployment distribution; (ii) adding at least one experiment that applies the corrected margin adjustment using labeled D_hat_u simulated from the WILDS folds; and (iii) clarifying that Corollary 4.4 is stated for population-level accuracies, whereas Definition 3.1 and the experimental ground truth use finite-sample empirical accuracies, so the event equivalence in the proof is only asymptotic in the sample sizes.
  4. [Section 5.3 and Table 5] The empirical evaluation contains no external baselines, and the paper's own ablations indicate only modest gains from the full pipeline. Section 5.3 justifies the absence of comparisons on the grounds that the setting is novel, but the decision rule is essentially an unsupervised accuracy estimate (the mean of pc) compared with a threshold derived from Acc(Dtest) through a significance test; natural baselines such as thresholding the difference in mean maximum softmax (or mean energy) between Du and Dtest, or adapting existing unsupervised accuracy estimators (Garg et al. 2022; Deng and Zheng 2021) to this two-sample comparison, should be reported. Table 5 in the appendix compounds the concern: for FMoW ID, the full filter reaches 0.969 ROC AUC while the single signals energy and logit max reach 0.965 each; for FMoW OOD, energy and logit max reach 0.971 versus 0.965 for the full filter. The 12-signal ensemble and the non-inferiority testing machinery thus show no clear improvement over a single-signal mean comparison in these results. Either include external baselines and report the ensemble's gain over its components at matched thresholds, or temper the claim that the specific instantiation rather than the general approach is responsible for the observed reliability.
minor comments (8)
  1. [Appendix A.1.2] The proof of Lemma 4.3 contains the leftover editorial annotation '(CHANGE THIS)' right after the sentence introducing delta-calibration; this placeholder must be removed before publication.
  2. [Appendix A.4.2] The text invokes 'Corollary 4.5,' but no such corollary exists in the manuscript; the intended cross-reference is presumably Corollary 4.4 or Remark 4.5.
  3. [Figure 1 and Definition 3.2] The label 'UNSUITABLE User Data' in Figure 1 does not match the two-valued output set (SUITABLE, INCONCLUSIVE) defined in Definition 3.2; the figure or the definition should be reconciled so that the decision semantics are unambiguous.
  4. [Definition 3.1] The prose says the criterion uses the 'estimated accuracy of M on Du,' whereas the displayed inequality uses oracle ground-truth labels, which are unavailable on Du by assumption; the prose should describe the displayed quantity as the ground-truth target that the filter estimates rather than as an estimate.
  5. [Section 1, contribution 4; Figure 4 caption] The claim of '29k experiments' refers to FMoW-WILDS only (16 x 15 x 14 plus 30 x 15 x 14 experiment settings times 3 seeds = 28,980); the full study including RxRx1-WILDS and CivilComments-WILDS comprises roughly 46k runs, so the count should be scoped precisely in the abstract and contributions.
  6. [Appendix A.2.2] Minor typos: 'disribution shifts' and 'train fro 50 epochs' should read 'distribution shifts' and 'train for 50 epochs.'
  7. [Theorem 4.2] The guarantee is stated under an exact normality and independence assumption on the pc samples, but pc takes values in [0,1] and is typically skewed; the guarantee should be presented as an asymptotic (CLT-based) one, as the Welch test is only approximately valid for such data.
  8. [Eqs. (11)-(12)] Equation (11) defines Delta as a generic empirical mean, while Eq. (12) uses Delta_test and Delta_u; the subscripts and the distinction between the estimator and its realization should be introduced at Eq. (11).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the suitability filter is evaluated against ground-truth suitability labels, and the correctness-estimator calibration premise is an explicit assumption rather than a hidden reuse of the target conclusion.

full rationale

The paper's central claim is that the suitability filter detects accuracy degradation on unlabeled user data, and the derivation chain is not circular. The per-sample correctness estimator C is trained on labeled source data D_sf by minimizing binary cross-entropy against true correctness labels, then applied to unlabeled user data; it is never fitted to the target suitability decision that is being predicted. The theoretical guarantee in Corollary 4.4 is an algebraic consequence of the explicitly stated δ-calibration assumption in Definition 4.1 together with the standard non-inferiority test guarantee in Theorem 4.2: under δ-calibration, the ground-truth suitability condition Acc_target < Acc_source - m is rewritten as μ_target < μ_source - m' with m' = m + δ_source - δ_target. This is a conditional statement under a stated assumption, not a definitional identity that presupposes the filter's correctness. Empirically, the paper defines ground-truth suitability independently using oracle labels (Acc(M, D_u) >= Acc(M, D_test) - m) and reports ROC/PR/FPR against that ground truth, so the decision rule is externally validated rather than being a renamed fit. The reader-identified concerns, including the possible sign inconsistency in Lemma 4.3, the fact that experiments set m' = m rather than using the corrected margin, and the practical replacement of δ_target by an empirical estimate from a small labeled sample, are correctness or statistical-validity issues; they do not make the derivation circular. No load-bearing self-citation or imported uniqueness claim was found, and the cited prior work on non-inferiority testing and dataset inference is used as background rather than as the sole justification for the paper's own guarantees.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central guarantee rests on δ-calibration on both source and target distributions, which is not verified. The experiments effectively assume zero calibration error by setting m'=m. The Welch test assumes normality and independence, and the framework assumes representative non-adversarial user data and covariate-only shift. The correctness estimator is a fitted logistic regression, and the signal set is hand-selected.

free parameters (2)
  • Correctness estimator parameters (w, b) = not reported
    The logistic regression C is trained on D_sf using binary cross-entropy on the signal features. The end-to-end suitability decision depends on these fitted weights, and their values are not provided in the paper.
  • Suitability signal set (12 hand-selected signals)
    The specific signals, such as conf max, logit max, energy, and margin loss, are chosen by hand rather than derived. The selection affects the correctness estimator and the reported performance.
assumptions (4)
  • domain assumption The correctness estimator C is δ-calibrated on both D_source and D_target.
    Required for Corollary 4.4 to bound the false positive rate. The paper states this is impossible to verify without labels (Section 4.2, citing David et al. 2010), and experiments set m'=m, effectively assuming δ_source = δ_target = 0.
  • standard math The pc[Dtest] and pc[Du] samples are independent and normally distributed.
    Needed for the Welch t-test guarantee in Theorem 4.2. The pc values are bounded in [0,1] and often near extremes, so normality is an approximation that is not validated.
  • domain assumption User data Du is a representative, non-adversarial sample from D_target.
    Stated in the Limitations section. Adversarial users could provide unrepresentative data to bypass the filter, which would invalidate the suitability decision.
  • domain assumption The deployment shift is covariate shift, not label shift.
    The Limitations section says the method is designed for covariate shifts and does not address label shift, so the validity of the framework is restricted to that setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings." pith.science (2026). https://pith.science/paper/24SCSPPC

@misc{pith2026250522356,
  author       = {Pith},
  title        = {Pith review of: Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24SCSPPC}},
  note         = {Machine review of arXiv:2505.22356}
}
read the original abstract

Deploying machine learning models in safety-critical domains poses a key challenge: ensuring reliable model performance on downstream user data without access to ground truth labels for direct validation. We propose the suitability filter, a novel framework designed to detect performance deterioration by utilizing suitability signals -- model output features that are sensitive to covariate shifts and indicative of potential prediction errors. The suitability filter evaluates whether classifier accuracy on unlabeled user data shows significant degradation compared to the accuracy measured on the labeled test dataset. Specifically, it ensures that this degradation does not exceed a pre-specified margin, which represents the maximum acceptable drop in accuracy. To achieve reliable performance evaluation, we aggregate suitability signals for both test and user data and compare these empirical distributions using statistical hypothesis testing, thus providing insights into decision uncertainty. Our modular method adapts to various models and domains. Empirical evaluations across different classification tasks demonstrate that the suitability filter reliably detects performance deviations due to covariate shift. This enables proactive mitigation of potential failures in high-stakes applications.

Figures

Figures reproduced from arXiv: 2505.22356 by the authors.

Figure 1
Figure 1. A model M is suitable for use on Du if its accuracy does not fall below the accuracy on Dtest by more than a pre￾defined margin m. The suitability filter calculates per-sample pre￾diction correctness probabilities for both test and user datasets and compares the two distributions through statistical non-inferiority testing. The dashed vertical lines represent the mean values of the distributions corresponding to the… view at source ↗
Figure 2
Figure 2. Schematic overview of the suitability filter. The suitability filter assesses whether model performance on a user sample Du deviates from its performance on the test dataset Dtest. This is achieved by combining different suitability signals {s1, . . . , ss} to estimate per-sample prediction correctness and comparing the distribution of these estimates between the two datasets using a statistical test. a diverse set … view at source ↗
Figure 3
Figure 3. Margin adjustment under accuracy estimation error. In each panel, the solid gray line is the perfect-calibration diagonal, the dashed black/gray lines mark the original margin m and its corrected value m′ , respectively. The blue/orange arrows indicate the estimation errors on the test set (∆test) and user data (∆u), respectively. In the left panel, the user data Du is deemed suitable; in the right panel it is deeme… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Sensitivity of suitability decisions to accuracy dif￾ferences between user and test data on FMoW-WILDS. The plot, summarizing results from nearly 29k individual experiments, shows the percentage of SUITABLE decisions for α = 0.05 and m = 0 across various accuracy diffe…
Figure 5
Figure 5. Figure 5: Suitability filtering on different OOD folds of FMoW-WILDS with and without additional calibration on Du. We choose a non-inferiority margin of m = 0.05 for this experiment. Accuracy Estimation Error. In Section 4.4, we propose using the empirical accuracy estimation e…
Figure 6
Figure 6. Figure 6: Distribution of the empirical accuracy estimation error ∆ for both the user and the test data across 6300 experiments on FMoW-WILDS. The suitability decisions depicted here have been made for a choice of m = 0 without margin adjustment due to miscalibration and at a si…
Figure 7
Figure 7. Figure 7: Relationship between performance deterioration for model M and the empirical accuracy estimation error ∆ for the user data across 6300 experiments on FMoW-WILDS. The suitability decisions depicted here have been made for a choice of m = 0 without margin adjustment due …
Figure 8
Figure 8. Figure 8: SHAP analysis for the prediction correctness estimator on FMoW-WILDS [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

90 extracted references · 57 canonical work pages

  1. [1]

    write newline

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

  2. [2]

    Ensemble of averages: Improving model selection and boosting performance in domain generalization

    Arpit, D., Wang, H., Zhou, Y., and Xiong, C. Ensemble of averages: Improving model selection and boosting performance in domain generalization. Advances in Neural Information Processing Systems, 35: 0 8265--8277, 2022

    2022

  3. [3]

    Baek, C., Jiang, Y., Raghunathan, A., and Kolter, J. Z. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. Advances in Neural Information Processing Systems, 35: 0 19274--19289, 2022

    2022

  4. [4]

    and Hochberg, Y

    Benjamini, Y. and Hochberg, Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57 0 (1): 0 289--300, 1995

    1995

  5. [5]

    Estimating model performance under covariate shift without labels

    Bialek, J., Kuberski, W., Perrakis, N., and Bifet, A. Estimating model performance under covariate shift without labels. 2024

    2024

  6. [6]

    Nuanced metrics for measuring unintended bias with real data for text classification

    Borkan, D., Dixon, L., Sorensen, J., Thain, N., and Vasserman, L. Nuanced metrics for measuring unintended bias with real data for text classification. In Companion proceedings of the 2019 world wide web conference, pp.\ 491--500, 2019

    2019

  7. [7]

    Domain generalization by mutual-information regularization with pre-trained models

    Cha, J., Lee, K., Park, S., and Chun, S. Domain generalization by mutual-information regularization with pre-trained models. In European conference on computer vision, pp.\ 440--457. Springer, 2022

    2022

  8. [8]

    Detecting errors and estimating accuracy on unlabeled data with self-training ensembles

    Chen, J., Liu, F., Avci, B., Wu, X., Liang, Y., and Jha, S. Detecting errors and estimating accuracy on unlabeled data with self-training ensembles. Advances in Neural Information Processing Systems, 34: 0 14980--14992, 2021 a

    2021

Show all 90 references
  1. [9]

    S., Poms, F., Fatahalian, K., and R \'e , C

    Chen, M., Goel, K., Sohoni, N. S., Poms, F., Fatahalian, K., and R \'e , C. Mandoline: Model evaluation under distribution shift. In International conference on machine learning, pp.\ 1617--1629. PMLR, 2021 b

    2021

  2. [10]

    An optimum character recognition system using decision functions

    Chow, C.-K. An optimum character recognition system using decision functions. IRE Transactions on Electronic Computers, 0 (4): 0 247--254, 1957

    1957

  3. [11]

    Functional map of the world

    Christie, G., Fendley, N., Wilson, J., and Mukherjee, R. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 6172--6180, 2018

    2018

  4. [12]

    Estimating generalization under distribution shifts via domain-invariant representations

    Chuang, C.-Y., Torralba, A., and Jegelka, S. Estimating generalization under distribution shifts via domain-invariant representations. In Proceedings of the 37th International Conference on Machine Learning, pp.\ 1984--1994, 2020

    1984

  5. [13]

    F., Williams, R., Reguera, B., Herry, V., and Gonz \'a lez-Gil, S

    Culverhouse, P. F., Williams, R., Reguera, B., Herry, V., and Gonz \'a lez-Gil, S. Do experts make mistakes? a comparison of human and machine indentification of dinoflagellates. Marine ecology progress series, 247: 0 17--25, 2003

    2003

  6. [14]

    B., Lu, T., Luu, T., and P \'a l, D

    David, S. B., Lu, T., Luu, T., and P \'a l, D. Impossibility theorems for domain adaptation. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pp.\ 129--136. JMLR Workshop and Conference Proceedings, 2010

    2010

  7. [15]

    and Zheng, L

    Deng, W. and Zheng, L. Are labels always necessary for classifier accuracy evaluation? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15069--15078, 2021

    2021

  8. [16]

    What does rotation prediction tell us about classifier accuracy under varying testing environments? In International Conference on Machine Learning, pp.\ 2579--2589

    Deng, W., Gould, S., and Zheng, L. What does rotation prediction tell us about classifier accuracy under varying testing environments? In International Conference on Machine Learning, pp.\ 2579--2589. PMLR, 2021

    2021

  9. [17]

    Confidence and dispersity speak: Characterizing prediction matrix for unsupervised accuracy estimation

    Deng, W., Suh, Y., Gould, S., and Zheng, L. Confidence and dispersity speak: Characterizing prediction matrix for unsupervised accuracy estimation. In International Conference on Machine Learning, pp.\ 7658--7674. PMLR, 2023

    2023

  10. [18]

    Unsupervised supervised learning i: Estimating classification and regression errors without labels

    Donmez, P., Lebanon, G., and Balasubramanian, K. Unsupervised supervised learning i: Estimating classification and regression errors without labels. Journal of Machine Learning Research, 11 0 (4), 2010

    2010

  11. [19]

    El-Yaniv, R. et al. On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11 0 (5), 2010

    2010

  12. [20]

    and Gall \'e , M

    Elsahar, H. and Gall \'e , M. To annotate or not? predicting performance drop under domain shift. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP),...

    2019

  13. [21]

    and Davidson, I

    Fan, W. and Davidson, I. Reverse testing: an efficient framework to select amongst classifiers under sample selection bias. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pp.\ 147--156, 2006

    2006

  14. [22]

    Selective prediction-set models with coverage rate guarantees

    Feng, J., Sondhi, A., Perry, J., and Simon, N. Selective prediction-set models with coverage rate guarantees. Biometrics, 79 0 (2): 0 811--825, 2023

    2023

  15. [23]

    and Ghahramani, Z

    Gal, Y. and Ghahramani, Z. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pp.\ 1050--1059. PMLR, 2016

    2016

  16. [24]

    Selective classification via one-sided prediction

    Gangrade, A., Kag, A., and Saligrama, V. Selective classification via one-sided prediction. In International Conference on Artificial Intelligence and Statistics, pp.\ 2179--2187. PMLR, 2021

    2021

  17. [25]

    C., Neyshabur, B., and Sedghi, H

    Garg, S., Balakrishnan, S., Lipton, Z. C., Neyshabur, B., and Sedghi, H. Leveraging unlabeled data to predict out-of-distribution performance. arXiv preprint arXiv:2201.04234, 2022

    2022 arXiv

  18. [26]

    and El-Yaniv, R

    Geifman, Y. and El-Yaniv, R. Selective classification for deep neural networks. Advances in neural information processing systems, 30, 2017

    2017

  19. [27]

    and El-Yaniv, R

    Geifman, Y. and El-Yaniv, R. Selectivenet: A deep neural network with an integrated reject option. In International conference on machine learning, pp.\ 2151--2159. PMLR, 2019

    2019

  20. [28]

    Bias-reduced uncertainty estimation for deep neural classifiers

    Geifman, Y., Uziel, G., and El-Yaniv, R. Bias-reduced uncertainty estimation for deep neural classifiers. In International Conference on Learning Representations, 2019

    2019

  21. [29]

    Ginsberg, T., Liang, Z., and Krishnan, R. G. A learning based hypothesis test for harmful covariate shift. arXiv preprint arXiv:2212.02742, 2022

    2022 arXiv

  22. [30]

    Ginsberg, T., Liang, Z., and Krishnan, R. G. A learning based hypothesis test for harmful covariate shift. In The Eleventh International Conference on Learning Representations, 2023

    2023

  23. [31]

    and Yuan, X

    Guan, L. and Yuan, X. Instance segmentation model evaluation and rapid deployment for autonomous driving using domain differences. IEEE Transactions on Intelligent Transportation Systems, 24 0 (4): 0 4050--4059, 2023

    2023

  24. [32]

    Predicting with confidence on unseen distributions

    Guillory, D., Shankar, V., Ebrahimi, S., Darrell, T., and Schmidt, L. Predicting with confidence on unseen distributions. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 1134--1144, 2021

    2021

  25. [33]

    and Lopez-Paz, D

    Gulrajani, I. and Lopez-Paz, D. In search of lost domain generalization. In International Conference on Learning Representations, 2021

    2021

  26. [34]

    Guo, C., Pleiss, G., Sun, Y., and Weinberger, K. Q. On calibration of modern neural networks. In International conference on machine learning, pp.\ 1321--1330. PMLR, 2017

    2017

  27. [35]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  28. [36]

    Multicalibration: Calibration for the (computationally-identifiable) masses

    H \'e bert-Johnson, U., Kim, M., Reingold, O., and Rothblum, G. Multicalibration: Calibration for the (computationally-identifiable) masses. In International Conference on Machine Learning, pp.\ 1939--1948. PMLR, 2018

    1939

  29. [37]

    and Dietterich, T

    Hendrycks, D. and Dietterich, T. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018

    2018

  30. [38]

    and Gimpel, K

    Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136, 2016

    2016 arXiv

  31. [39]

    Aries: Efficient testing of deep neural networks via labeling-free accuracy estimation

    Hu, Q., Guo, Y., Xie, X., Cordy, M., Papadakis, M., Ma, L., and Le Traon, Y. Aries: Efficient testing of deep neural networks via labeling-free accuracy estimation. In 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE), pp.\ 1776--1787. IEEE, 2023

    2023

  32. [40]

    Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger, K. Q. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 4700--4708, 2017

    2017

  33. [41]

    Self-adaptive training: beyond empirical risk minimization

    Huang, L., Zhang, C., and Zhang, H. Self-adaptive training: beyond empirical risk minimization. Advances in neural information processing systems, 33: 0 19365--19376, 2020

    2020

  34. [42]

    Estimating the accuracies of multiple classifiers without labeled data

    Jaffe, A., Nadler, B., and Kluger, Y. Estimating the accuracies of multiple classifiers without labeled data. In Artificial Intelligence and Statistics, pp.\ 407--415. PMLR, 2015

    2015

  35. [43]

    Jiang, Y., Nagarajan, V., Baek, C., and Kolter, J. Z. Assessing generalization of sgd via disagreement. arXiv preprint arXiv:2106.13799, 2021

    2021 arXiv

  36. [44]

    Kingma, D. P. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  37. [45]

    K., and Kuberski, W

    Kivim \"a ki, J., Bia ek, J., Nurminen, J. K., and Kuberski, W. Confidence-based estimators for predictive performance in model monitoring. arXiv preprint arXiv:2407.08649, 2024

    2024 arXiv

  38. [46]

    W., Sagawa, S., Marklund, H., Xie, S

    Koh, P. W., Sagawa, S., Marklund, H., Xie, S. M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R. L., Gao, I., et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pp.\ 5637--5664. PMLR, 2021

    2021

  39. [47]

    Fairness in credit scoring: Assessment, implementation and profit implications

    Kozodoi, N., Jacob, J., and Lessmann, S. Fairness in credit scoring: Assessment, implementation and profit implications. European Journal of Operational Research, 297 0 (3): 0 1083--1094, 2022

    2022

  40. [48]

    and Inouye, D

    Kulinski, S. and Inouye, D. I. Towards explaining distribution shifts. In International Conference on Machine Learning, pp.\ 17931--17952. PMLR, 2023

    2023

  41. [49]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Lakshminarayanan, B., Pritzel, A., and Blundell, C. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017

    2017

  42. [50]

    L., Romano, J

    Lehmann, E. L., Romano, J. P., and Casella, G. Testing statistical hypotheses, volume 3. Springer, 1986

    1986

  43. [51]

    Learning a data-driven policy network for pre-training automated feature engineering

    Li, L., Wang, H., Zha, L., Huang, Q., Wu, S., Chen, G., and Zhao, J. Learning a data-driven policy network for pre-training automated feature engineering. In The Eleventh International Conference on Learning Representations, 2023

    2023

  44. [52]

    P., Salakhutdinov, R

    Liu, Z., Wang, Z., Liang, P. P., Salakhutdinov, R. R., Morency, L.-P., and Ueda, M. Deep gamblers: Learning to abstain with portfolio theory. Advances in Neural Information Processing Systems, 32, 2019

    2019

  45. [53]

    Fixing weight decay regularization in adam

    Loshchilov, I., Hutter, F., et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017

    2017 arXiv

  46. [54]

    Predicting out-of-distribution error with confidence optimal transport

    Lu, Y., Wang, Z., Zhai, R., Kolouri, S., Campbell, J., and Sycara, K. Predicting out-of-distribution error with confidence optimal transport. arXiv preprint arXiv:2302.05018, 2023

    2023 arXiv

  47. [55]

    Co-validation: Using model disagreement on unlabeled data to validate classification algorithms

    Madani, O., Pennock, D., and Flake, G. Co-validation: Using model disagreement on unlabeled data to validate classification algorithms. Advances in neural information processing systems, 17, 2004

    2004

  48. [56]

    Performance prediction under dataset shift

    Maggio, S., Bouvier, V., and Dreyfus-Schmidt, L. Performance prediction under dataset shift. In 2022 26th International Conference on Pattern Recognition (ICPR), pp.\ 2466--2474. IEEE, 2022

    2022

  49. [57]

    Dataset inference: Ownership resolution in machine learning

    Maini, P., Yaghini, M., and Papernot, N. Dataset inference: Ownership resolution in machine learning. arXiv preprint arXiv:2104.10706, 2021

    2021 arXiv

  50. [58]

    K-means clustering based feature consistency alignment for label-free model evaluation

    Miao, S., Zheng, L., Liu, J., and Jin, H. K-means clustering based feature consistency alignment for label-free model evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3299--3307, 2023

    2023

  51. [59]

    O'Brien, P. C. and Fleming, T. R. A multiple testing procedure for clinical trials. Biometrics, pp.\ 549--556, 1979

    1979

  52. [60]

    Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift

    Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., and Snoek, J. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information processing systems, 32, 2019

    2019

  53. [61]

    Came: Contrastive automated model evaluation

    Peng, R., Duan, Q., Wang, H., Ma, J., Jiang, Y., Tu, Y., Jiang, X., and Zhao, J. Came: Contrastive automated model evaluation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 20121--20132, 2023

    2023

  54. [62]

    Energy-based automated model evaluation

    Peng, R., Zou, H., Wang, H., Zeng, Y., Huang, Z., and Zhao, J. Energy-based automated model evaluation. arXiv preprint arXiv:2401.12689, 2024

    2024 arXiv

  55. [63]

    M., and Horvitz, E

    Platanios, E., Poon, H., Mitchell, T. M., and Horvitz, E. J. Estimating accuracy from unlabeled data: A probabilistic logic approach. Advances in neural information processing systems, 30, 2017

    2017

  56. [64]

    A., Dubey, A., and Mitchell, T

    Platanios, E. A., Dubey, A., and Mitchell, T. Estimating accuracy from unlabeled data: A bayesian approach. In International Conference on Machine Learning, pp.\ 1416--1425. PMLR, 2016

    2016

  57. [65]

    Platt, J. et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10 0 (3): 0 61--74, 1999

    1999

  58. [66]

    Pocock, S. J. Clinical trials: a practical approach. John Wiley & Sons, 2013

    2013

  59. [67]

    and Ramdas, A

    Podkopaev, A. and Ramdas, A. Tracking the risk of a deployed model and detecting harmful distribution shifts. arXiv preprint arXiv:2110.06177, 2021

    2021 arXiv

  60. [68]

    Qui \ n onero-Candela, J., Sugiyama, M., Schwaighofer, A., and Lawrence, N. D. Dataset shift in machine learning. Mit Press, 2022

    2022

  61. [69]

    Failing loudly: An empirical study of methods for detecting dataset shift

    Rabanser, S., G \"u nnemann, S., and Lipton, Z. Failing loudly: An empirical study of methods for detecting dataset shift. Advances in Neural Information Processing Systems, 32, 2019

    2019

  62. [70]

    Selective classification via neural network training dynamics

    Rabanser, S., Thudi, A., Hamidieh, K., Dziedzic, A., and Papernot, N. Selective classification via neural network training dynamics. arXiv preprint arXiv:2205.13532, 2022

    2022 arXiv

  63. [71]

    Learning to validate the predictions of black box machine learning models on unseen data

    Redyuk, S., Schelter, S., Rukat, T., Markl, V., and Biessmann, F. Learning to validate the predictions of black box machine learning models on unseen data. In Proceedings of the Workshop on Human-In-the-Loop Data Analytics, pp.\ 1--4, 2019

    2019

  64. [72]

    Imagenet large scale visual recognition challenge

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115: 0 211--252, 2015

    2015

  65. [73]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter

    Sanh, V. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019

    1910 arXiv

  66. [74]

    Learning to validate the predictions of black box classifiers on unseen data

    Schelter, S., Rukat, T., and Bie mann, F. Learning to validate the predictions of black box classifiers on unseen data. In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data, pp.\ 1289--1299, 2020

    2020

  67. [75]

    Label-free model evaluation with semi-structured dataset representations

    Sun, X., Hou, Y., Li, H., and Zheng, L. Label-free model evaluation with semi-structured dataset representations. arXiv preprint arXiv:2112.00694, 2021

    2021 arXiv

  68. [76]

    Target specification bias, counterfactual prediction, and algorithmic fairness in healthcare

    Tal, E. Target specification bias, counterfactual prediction, and algorithmic fairness in healthcare. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pp.\ 312--321, 2023

    2023

  69. [77]

    Rxrx1: An image set for cellular morphological variation across many experimental batches

    Taylor, J., Earnshaw, B., Mabey, B., Victors, M., and Yosinski, J. Rxrx1: An image set for cellular morphological variation across many experimental batches. In International Conference on Learning Representations (ICLR), volume 22, pp.\ 23, 2019

    2019

  70. [78]

    V., Pacchioni, F., Bellumore, S., Bosia, M., and De Micco, F

    Travaini, G. V., Pacchioni, F., Bellumore, S., Bosia, M., and De Micco, F. Machine learning and criminal justice: A systematic review of advanced methodology for recidivism risk prediction. International journal of environmental research and public health, 19 0 (17): 0 10594, 2022

    2022

  71. [79]

    A bag-of-prototypes representation for dataset-level applications

    Tu, W., Deng, W., Gedeon, T., and Zheng, L. A bag-of-prototypes representation for dataset-level applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2881--2892, 2023

    2023

  72. [80]

    Predicting neural network accuracy from weights

    Unterthiner, T., Keysers, D., Gelly, S., Bousquet, O., and Tolstikhin, I. Predicting neural network accuracy from weights. arXiv preprint arXiv:2002.11448, 2020

    2002 arXiv

  73. [81]

    and Nowacki, A

    Walker, E. and Nowacki, A. S. Understanding equivalence and noninferiority testing. Journal of general internal medicine, 26: 0 192--196, 2011

    2011

  74. [82]

    Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., and Philip, S. Y. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35 0 (8): 0 8052--8072, 2022

    2022

  75. [83]

    Toward auto-evaluation with confidence-based category relation-aware regression

    Wang, J., Chen, J., and Su, B. Toward auto-evaluation with confidence-based category relation-aware regression. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.\ 1--5. IEEE, 2023

    2023

  76. [84]

    Welch, B. L. The generalization of ‘student's’problem when several different population varlances are involved. Biometrika, 34 0 (1-2): 0 28--35, 1947

    1947

  77. [85]

    Testing statistical hypotheses of equivalence

    Wellek, S. Testing statistical hypotheses of equivalence. Chapman and Hall/CRC, 2002

    2002

  78. [86]

    A fine-grained analysis on distribution shift

    Wiles, O., Gowal, S., Stimberg, F., Alvise-Rebuffi, S., Ktena, I., Dvijotham, K., and Cemgil, T. A fine-grained analysis on distribution shift. arXiv preprint arXiv:2110.11328, 2021

    2021 arXiv

  79. [87]

    On the importance of feature separability in predicting out-of-distribution error

    Xie, R., Wei, H., Feng, L., Cao, Y., and An, B. On the importance of feature separability in predicting out-of-distribution error. Advances in Neural Information Processing Systems, 36, 2024

    2024

  80. [88]

    Predicting out-of-distribution error with the projection norm

    Yu, Y., Yang, Z., Wei, A., Ma, Y., and Steinhardt, J. Predicting out-of-distribution error with the projection norm. In International Conference on Machine Learning, pp.\ 25721--25746. PMLR, 2022

    2022

  81. [89]

    Domain generalization with mixstyle

    Zhou, K., Yang, Y., Qiao, Y., and Xiang, T. Domain generalization with mixstyle. In International Conference on Learning Representations, 2021

    2021

  82. [90]

    Zhou, K., Liu, Z., Qiao, Y., Xiang, T., and Loy, C. C. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (4): 0 4396--4415, 2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.