acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Driven Unsupervised Detection Of Reliability-Relevant Distribution Drift

Abstract

Unsupervised concept drift detection commonly relies on identifying changes in the distribution of inputs to a trained machine learning model. Yet not every distribution shift should trigger an alarm: a deployment distribution may differ from recent observations, while remaining covered by prediction-relevant variations encountered during training. Drift thus should be triggered when incoming data introduce novel predictive behavior that was insufficiently supported by the training data. We show that this distinction is naturally captured by epistemic uncertainty (EU). Focusing on Bayesian neural networks (NNs) obtained through a Laplace approximation, we de-rive upper and lower bounds on EU in a local prediction-relevant geometry induced jointly by the trained model and its training data. The bounds reveal that EU along training-covered directions is attenuated according to the information accumulated during training, whereas components orthogonal to the training-supported subspace retain substantially larger EU. Guided by this characterization, we develop an unsupervised drift detector that monitors EU rather than input-distribution change directly, and can be applied post hoc to conventionally trained NNs. Experiments across non-stationary learning settings show that the proposed approach suppresses alarms on training-covered shifts while reliably detecting novel ones.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Epistemic uncertainty, Unsupervised drift detection, Concept drift, Model reliability, Reliability-aware drift detection, Reliability-relevant distribution shift, Uncertainty quantification, Bayesian neural networks, Model monitoring, Laplace approximation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WeSJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes to dect drift using model unceratinty. It uses a linearized Laplace approximation to estimate uncertainty and derives a theoretical analysis connecting EU to training data. The approach allows drift detection without requiring labels or access to the training data during deployment. While the framework is interesting, the main contribution is very limited and based on preexisting ideas. The paper spent lot of time trying to justify the method mathematically but much of it follows directly from standard matrix decomposition. The overall novelty of the contribution therefore appears limited.

Critical strengths

The topic is important and relevant.

  • The proposed method uses EU to monitor distribution shift without requiring labels during deployment. This is relevant when training data is not available.
  • the decomposition into training-supported and unsupported parameter directions is interesting interpretation of how training data coverage affects EU.
Critical weaknesses

The theoretical framing offer an incremental contribution but I have not seen a fundamentally new method or a surprising theoretical result. The potential contribution is its explicit theoretical interpretation of uncertainty through prediction-relevant training coverage but the mathematical results are relatively straightforward.

  • Bayesian uncertainty formulation is well established
  • Laplace framework is also established etc.
  • The theoretical analysis is formulated using the full Laplace posterior covariance, which scales quadratically with the number of model parameters. So its unclear how the propose method scals on medium or large models. The validation experiments is done using relatively small models.
  • How is S the number of Monte Carlo samples drawn from the Bayesian posterior to estimate epistemic uncertainty (EU) choosen en pratice and what is its relationship with the batch size?
Questions for authors

-see above

2 comments
AuthorsOpenReview ↗

Response to Reviewer WeSJ

We address the concerns regarding novelty, scalability, and uncertainty estimation below.

W1 – Novelty and contribution. We agree that Bayesian uncertainty, the Laplace approximation, and the associated matrix decompositions are established tools, and we do not claim these individually as contributions. As emphasized in Section 2, our objective is not to develop another uncertainty estimator. Rather, our contribution lies in connecting these tools to a distinct problem: detecting distribution shifts that expose insufficient prediction-relevant training coverage, rather than simply detecting changes in the input distribution.

This connection has three aspects. First, we formulate a model-aware monitoring objective in which shifts between different but sufficiently covered operating conditions should not necessarily trigger alarms. Second, Theorem 4.2 characterizes how training curvature attenuates EU along supported prediction-relevant directions, while insufficiently constrained directions retain higher uncertainty. This provides a theoretical interpretation of when EU reflects limited training coverage, beyond treating uncertainty as a heuristic novelty score. Third, we translate this characterization into a practical label-free sequential monitoring framework.

The distinction has meaningful experimental consequences. For example, in the ambiguous label-prior experiment of Appendix E.7, predictive entropy increases and triggers false alarms despite unchanged prediction error, whereas EU remains stable. Across the original experiments, the advantage is particularly pronounced in audio classification, with AUROC of 0.98 versus 0.59 for predictive entropy on ESC-50 (CNN), and 0.98 versus 0.80 on UrbanSound8K (CNN-wide). These results illustrate the practical significance of isolating the uncertainty associated with limited model knowledge.

W2 – Scalability. The full Laplace covariance is used for the theoretical characterization, but is not required by our practical implementation. As detailed in Appendix D, we use a post-hoc last-layer Laplace approximation with diagonal generalized Gauss–Newton curvature (Daxberger et al., 2021). Consequently, posterior storage and the associated diagonal operations scale linearly, rather than quadratically, with the number of last-layer parameters. The backbone remains deterministic, and posterior sampling requires only evaluating the final layer on previously computed features.

Our evaluation already includes a ViT-B/16 backbone with approximately 86 million parameters, demonstrating applicability beyond small networks. Furthermore, the framework is not restricted to Laplace approximations: our experiments also use MC-dropout (Gal and Ghahramani, 2016), which does not require storing a posterior covariance and exhibits similar qualitative behavior.

Q – Number of posterior samples and batch size. The posterior sample count SS and deployment batch size BB are independent parameters serving different purposes. Specifically, SS controls the Monte Carlo approximation of per-sample EU in Eq. (17), whereas BB determines how many incoming observations are averaged to obtain the monitoring statistic in Eq. (18). Increasing SS improves the accuracy of uncertainty estimation, while increasing BB reduces the variability of the batch-level monitoring statistic due to sampling deployment inputs.

In our CIFAR-100-C experiments, for example, we use S=20S=20 posterior samples, independently of the deployment batch size. For last-layer Laplace, the backbone is evaluated only once per input, while additional posterior samples require computations only in the final layer. Thus, moderate posterior sampling can be implemented efficiently without repeatedly evaluating the entire network.

References:

Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace redux-effortless Bayesian deep learning. Advances in Neural Information Processing Systems, 34, 2021.

Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning, pp. 1050–1059. PMLR, 2016.

Thanks for your reply and the further explanations. I also note the practical relevance of the method and the experimental results which show some advantages over predictive entropy. I still think the main contribution is very much based on existing ideas and that the theoretical analysis is not particularly novel for ICLR. I'll therefore stick with my original scores.

Reviewer_qUS8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper explores the idea of unsupervised detection of drifts that affect the reliability of the model. The idea is the monitoring of the uncertainty instead of just the input distribution. This allows to identify the chances that truly affect the model performance.

Critical strengths

The paper is well motivated and explores an interesting research direction. The formulation based on the uncertainty is conceptually solid and the geometrical analysis is well described. Overall the paper is well written.

Critical weaknesses

The main concern is the gap between theory and empirical evaluation. The theory provides a local bound about the uncertainty but it does not demonstrate that this truly increases as a consequence of the shift. This is a point to be clarified. The second concern refers to the experimental section. The evaluation is carried out on synthetic streams of corrupted data and the baselines are limited. Why not considering other solutions?

Questions for authors

See previous comments.

1 comment
AuthorsOpenReview ↗

Response to Reviewer qUS8

We address the two concerns below, supported by additional experiments included in the revised appendix.

W1 – Connection between theory and drift detection. Our analysis characterizes how EU depends on the prediction-relevant information accumulated during training. In particular, Theorem 4.2 shows that the local EU contribution along well-constrained training directions is attenuated by the training curvature, whereas directions orthogonal to the training-supported subspace retain their prior uncertainty. We agree that this alone does not guarantee an increase in aggregate EU following every novel shift, since contributions from supported directions may change simultaneously. As clarified in our response to Reviewer M927, we derive an explicit sufficient condition for increasing expected EU that accounts for this compensation. Thus, the theory identifies when novel predictive variations produce an increased uncertainty signal, rather than guaranteeing detection under arbitrary distribution changes.

The original experiments illustrate this mechanism: EU remains relatively stable under training-covered shifts and increases when previously unseen conditions are introduced. Particularly informative is the ambiguous label-prior experiment in Appendix E.7, where predictive entropy increases despite unchanged prediction error, while EU remains stable. This demonstrates why separating EU from intrinsic predictive uncertainty is useful for the proposed monitoring objective.

W2 – Experimental settings and baselines. Our original evaluation uses controlled streams to distinguish training-covered and novel operating conditions with known ground truth, following established distribution-shift evaluation practices (Hendrycks and Dietterich, 2019; Ovadia et al., 2019). However, the experiments cover diverse sources of non-stationarity, including waveform degradations in real audio recordings and propagation changes generated by the physics-based QuaDRiGa simulator.

To further address the concern, we conducted additional experiments involving naturally occurring novel classes and subtypes, without artificial corruptions, on CIFAR-100 and UrbanSound8K, as seen in Appendix E.10. We also added comparisons with aleatoric uncertainty (AU), alongside EU and total predictive uncertainty. The following table reports AUROC for detecting novel data relative to familiar data:

SettingEUAUTotal
CIFAR-100: unseen subtypes0.991.001.00
CIFAR-100: 25% unseen superclasses0.830.860.86
UrbanSound8K: 25% unseen classes, MC-dropout0.890.960.96

All three signals detect novelty in these settings, demonstrating that EU is informative beyond synthetic corruption streams, although it does not always outperform the alternatives.

Importantly, the advantage of EU becomes more apparent when distinguishing novel inputs from difficult but familiar ones. For example, when comparing previously unseen UrbanSound8K classes against faint recordings of familiar classes, we obtain:

Unseen-class fractionEUAUTotal
25%0.330.070.10
50%0.750.390.44
100%1.000.850.95

These MC-dropout results show that EU is more selective to missing training support, although small fractions of novel inputs remain challenging to distinguish from difficult familiar observations.

Finally, our original evaluation compares different monitoring signals under the same sequential detectors (ADWIN, Page–Hinkley, and KSWIN), isolating the contribution of the proposed statistic. The additional AU comparison further strengthens this evaluation, while comparisons with other complete drift-detection frameworks remain a useful extension.

Reviewer_M927OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies unsupervised distribution drift detection for deployed neural networks, focusing on distinguishing training-covered shifts from novel shifts that introduce prediction-relevant variations insufficiently supported by the training data. The authors propose using epistemic uncertainty (EU), rather than changes in the input distribution or total predictive uncertainty, as the monitoring signal.

The theoretical analysis considers a linearized neural network with a Laplace-approximated parameter posterior. It derives a local approximation of EU and establishes a geometric decomposition and spectral bounds relating EU to parameter directions supported by the training data and their orthogonal complement. Based on this characterization, the proposed method estimates EU for incoming unlabeled samples, aggregates it at the batch level, and applies standard sequential change detectors.

The method is evaluated across image classification, audio classification, and wireless communication tasks, using Laplace approximations and MC-dropout. The reported results suggest that EU can distinguish training-covered from held-out conditions, with notable improvements over predictive entropy in several audio settings.

Critical strengths
  1. An interesting research perspective. The paper explores the distinction between training-covered and novel distribution shifts, offering a model-aware perspective on unsupervised drift detection.

  2. Encouraging empirical results. The experiments cover diverse datasets, architectures, and modalities, with notable improvements over predictive entropy in several audio settings.

Critical weaknesses
  1. Limited explanatory power of EU and the theory-to-detection gap. Theorem 4.2 characterizes how training-supported and novel directions contribute to local EU, but these results do not establish that the two types of distribution shifts lead to distinguishable EU statistics. Even when novel directions emerge, their positive contribution may be offset by changes in training-supported directions, leaving the overall EU unchanged. Moreover, the spectral bounds can be rather loose when the positive eigenvalues of MDM_D vary substantially. Although Eqs. (15)–(16) provide bounds on expected local EU under training-covered and novel conditions, respectively, they do not guarantee sufficient separation between the corresponding deployment distributions. Thus, the theoretical analysis does not adequately establish when EU can distinguish these two types of shifts, leaving a significant gap between the geometric characterization in Section 4.2 and the detection objective in Section 4.3.

  2. An unclear criterion for sufficient training coverage. The paper defines novel shifts through prediction-relevant variations insufficiently supported during training, but does not specify what constitutes sufficient support. In Section 4.2, SD=range⁡(MD)S_D=\operatorname{range}(M_D) includes all directions with nonzero training curvature, regardless of how strongly they are constrained. For example, a direction with an arbitrarily small positive eigenvalue belongs to SDS_D, despite receiving very little information from training. Thus, membership in the training-supported subspace does not necessarily imply sufficient training coverage.

Questions for authors

Theoretical justification for drift discrimination. Can the authors clarify under what additional conditions the geometric bounds in Theorem 4.2 and Eqs. (15)–(16) imply distinguishable EU statistics for training-covered and novel deployment distributions? In particular, how can the analysis account for possible compensation between the contributions of training-supported and novel directions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer M927

We thank the reviewer for the constructive comments and for recognizing the motivation and empirical results of our work. We address the concerns below, including additional experiments in the revised appendix.

W1/Q – Theoretical justification for drift discrimination. Our theoretical analysis aims to characterize how prediction-relevant training coverage governs EU, rather than guarantee that every novel distribution shift necessarily produces a detectable increase. We agree that the latter requires additional conditions, since a decrease in uncertainty along training-supported directions may compensate for the contribution of novel directions.

Note that Eq. (13) provides an exact decomposition of the local EU approximation, beyond the spectral bounds of Theorem 4.2. Let γj>0\gamma_j > 0 denote the positive eigenvalues of the training curvature matrix, with corresponding eigenvectors vjv_j, and define ej(x)=∥Bxvj∥2e_j(x) = \lVert B_x v_j \rVert^2. Then

U~(x)=12∑jej(x)λ+γj+ν(x)2λ.\widetilde{U}(x) = \frac{1}{2}\sum_j \frac{e_j(x)}{\lambda + \gamma_j} + \frac{\nu(x)}{2\lambda}.

Consequently, for a training-covered distribution PCP_C and a novel distribution PNP_N, the expected local EU is guaranteed to increase whenever

EN[ν]λ>∑jEC[ej]−EN[ej]λ+γj.\frac{\mathbb{E}_N[\nu]}{\lambda} > \sum_j \frac{\mathbb{E}_C[e_j] - \mathbb{E}_N[e_j]}{\lambda + \gamma_j}.

This condition explicitly accounts for the possible compensation raised by the reviewer. It also uses the individual curvature eigenvalues, avoiding the potentially loose spectral bounds based on their extrema. The condition applies to local EU; transferring it to the exact EU requires the approximation error to be sufficiently small. We add this clarification to Section 4.2.

W2 – Sufficient training coverage. We agree that membership in the training-supported subspace alone does not imply sufficient coverage. Our analysis provides a graded notion of coverage, determined by the accumulated curvature relative to the prior precision. As the expression above shows, directions with γj≫λ\gamma_j \gg \lambda contribute little EU, whereas weakly constrained directions with γj≪λ\gamma_j \ll \lambda retain substantial uncertainty, even if they formally belong to the supported subspace.

To further illustrate this behavior, we conducted an additional experiment in Appendix E.10 on CIFAR-100, varying the number of training examples from previously held-out subtypes while predicting their superclass. We report classification accuracy and the AUROC of EU for distinguishing the held-out subtypes from familiar ones:

Training examples per held-out subtypeAccuracyEU AUROC
00.490.99
500.600.97
1250.680.89
2500.760.77
4500.800.50

As training coverage increases, accuracy approaches that of familiar subtypes (0.81), while EU discrimination decreases toward chance. Thus, coverage is not determined by a binary subspace-membership criterion, but progressively improves as training provides additional information. These findings empirically support the graded interpretation of coverage underlying our analysis.

Thank you very much for the thorough response, helpful clarifications, and additional experiments. While I agree that the theory need not guarantee detectable EU shifts under every novel distribution, the analysis would still benefit from identifying meaningful structural conditions under which EU separates novel inputs from training-covered or difficult-but-familiar ones. The added condition is mathematically sound, yet it follows fairly directly from the existing decomposition and offers limited insight into when such separation holds or why EU outperforms total predictive uncertainty. Although the new empirical results are very encouraging, my primary concern regarding the explanatory depth of the theoretical results remains. I will therefore maintain my original assessment, though I appreciate the authors' efforts.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.