acceptodds
Withdrawn from ICLR 2027

Uncertainty Aware DeepView: Global Visualization of Unsupervised Quantifiable Uncertainty

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While research literature has expanded rapidly regarding explanations in the uncertainty domain, existing methods focus almost exclusively on individual predictions, offering little insight into the distribution of the boundaries that separate confidently modeled regions from unreliable ones. We propose Uncertainty Aware Deepview, the first globally faithful uncertainty visualization technique that is applicable to any type of model, and comes with a robustness estimate. Uncertainty Aware DeepView removes the requirement of an uncertainty output of the model and combines two concepts for the determination of unsupervised quantifiable uncertaintes: An adversarial score estimating the local instability of a sample, and an out-of-distribution score quantifying the distributional anomality of a sample. We demonstrate Uncertainty Aware Deepview's ability to accurately identify out-of-distribution data as well as adversarial examples on different models including Autoencoders, CNNs, Resnets and Vision Transformers on multiple benchmark datasets.

Primary area:
interpretability and explainable AI
Keywords:
Uncertainties, Explainable AI, Verification

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_4jWTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors of this paper propose Uncertainty Aware Deepview, an uncertainty visualisation technique that comes with robustness estimates and they claim is applicable to any type of model. Their approach combines measures for local model instability and OOD detection into an uncertainty score, which is visualised using an extension of the DeepView framework without requiring labels or probabilistic model outputs. Their evaluation consists of multiple benchmarks, across Autoencoders, CNNs, Resnets, and Vision Transformers, measuring Uncertainty Aware Deepview's ability to identify out-of-distribution and adversarial inputs.

Critical strengths
  1. The paper tackles a clear limitation of previous Unsupervised DeepView that is requiring an uncertainty signal or predictive distribution. Combining local instability and outlier scores is a way to construct uncertainty visualisations for models such as autoencoders. The motivation is meaningful.

  2. The authors evaluation spans five datasets, multiple model architectures as well as different instability and outlier proxies. Testing Uncertainty Aware Deepview across clean, OOD, and adversarial inputs shows the possible trade-offs well, providing useful conclusions about how to scoring component behaves across different settings.

Critical weaknesses
  1. The visualisation framework closely follows existing DeepView works where steps 1-3 in S3.2 adopt its distance calculation, embedding and inverse mapping. The main change is to visualise a score combining existing instability and outlier measures. This extension could be useful, but the paper provides limited evidence of its value. Despite claiming the visualisations to be "globally faithful", the experiments measure detection accuracy rather than the visualisation fidelity. Figure 2 is the only example visualisation and it uses a classifier, despite extending visualisation to models without probabilistic outputs being a central motivation. The paper does not specify the output dissimilarity used to construct the embedding for autoencoders, leaving an important part of its proposed extension unclear.

  2. The experiments don't establish the benefit of combining the two score paths and important evaluation details and therefore missing. Table 1 and Appendix A compare different pairs of instability and outlier proxies, but do not evaluate either component alone. Therefore, it is unclear whether the combined score improves on existing detectors or largely inherits their performance. The value of λ in Eq1, the final detection threshold, and several proxy normalisation settings are also unspecified, making the reported accuracies difficult to interpret and reproduce.

Questions for authors
  1. (Relating to W1): Could the authors calify how the visualisation is constructed for autoencoders, particularly which output dissimilarity is used in the embedding? Could the authors also provide an autoencoder visulaisation and evidence supporting the claim of global faithfulness? There appears to be room in the main paper to explain and demonstrate this central contribution more fully.

  2. (Relating to W2): How were λ, the proxy normalisation parameters, and the final detection threshold selected? How does the combined score compare with each component alone under the same calibration procedure?

1 comment
AuthorsOpenReview ↗

Adressing Feedback

We thank the reviewer for their constructive feedback.

We agree that UAD-View deliberately keeps DeepView’s embedding and inverse mapping, so that the resulting visualizations remain comparable with previous DeepView variants. We would like to clarify where the extension goes beyond visualizing a new score, and we have revised the paper to address each concern.

Novelty beyond DeepView: DeepView’s discriminative distance is not model-agnostic. It is defined through the Fisher information of the class posterior p(y|x), so steps 1–3 cannot be applied to models without a predictive distribution, such as autoencoders. The property DeepView needs (stretching the metric where the model’s output changes and contracting it where the output is constant) does not depend on p(y|x). We generalize the metric to the pullback of an arbitrary output dissimilarity δ along the interpolation path, which recovers DeepView’s Fisher distance when δ is defined on class posteriors. This generalization is what makes a discriminative embedding, and therefore a DeepView-type visualization, possible for likelihood-free models. On top of it, the colouring no longer requires a model output: the uncertainty score is computed from model-agnostic proxies on real and inverse-mapped synthetic points.

W1/ Q1: Missing technical details and experiments on visualization fidelity

The computation of the output dissimilarity in the embedding is similar to the original DeepView embedding. We compute the difference between the raw model outputs. For autoencoders this is the raw difference in the reconstruction rather than the difference in class probabilities (e.g., for classifiers with a softmax layer). We will make this more clear in a revised version and also clarify the consequences on a theoretical level. We also plan to include more visualizations (across different models and datasets) and extend the experiments to provide more evidence of the models global faithfulness.

W2/ Q2:

The combination of the two scores is consistently outperforming the individual scores in our experimental evaluation as the individual scores can often only capture the model's instability under small perturbations or the sample's distributional atypicality, but not both. Hence, they usually perform poorly on the opposite portion of the evaluation set. We will include more experimental results in a revised version of the paper, demonstrating the benefit of the linear combination over the individual scores.

For λ\lambda we chose a value of 0.50.5, indicating equal importance of both individual scores in the general setting. The normalization of our scores is set to usually map the 95% quantile among the training data to a value of $0.5% and the mean among the training data to 0.050.05. As a simple default we set the final detection threshold to a value of 0.30.3 (Note that this is slightly higher than scoring an individual score of 0.50.5 in either the instability or out-of-distribution score due to the factor of λ=0.5\lambda = 0.5). We will make the default values of these and all other hyperparameter more clear in a revised version of the paper.

Reviewer_z2RVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper extends the existing DeepView framework to define "Unsupervised Quantifiable Uncertainty" (UQU)—which combines local model instability with out-of-distribution characteristics—and proposes "Uncertainty-Aware DeepView" to visualize this uncertainty globally. A key contribution of this work is its applicability to a wide range of model architectures without relying on probabilistic model outputs or data labels. The proposed method was validated using various benchmark datasets (including MNIST, SVHN, CIFAR-10, and Tiny-ImageNet) and model architectures (such as CNNs, ResNets, and Vision Transformers).

Critical strengths

None

Critical weaknesses
  1. Lack of sufficient novelty in the proposed method

The paper's core methodological contribution lies in defining a new metric by linearly combining instability and outlier scores. However, since combining two distinct metrics via a weighted sum is a relatively intuitive and simple approach, it is difficult to claim high methodological originality based on this alone.

While simplicity is not necessarily a drawback, and demonstrating the effectiveness of a simple method can indeed constitute a significant contribution, this paper lacks sufficient theoretical or empirical justification for why a linear combination is necessary and what advantages it offers over using the individual metrics. Given that a substantial portion of the paper relies on the existing DeepView methodology, the necessity and distinctiveness of the newly proposed uncertainty definition must be clearly demonstrated.

  1. Insufficient justification for the UQU proposed in Definition 1

For UQU, the two uncertainty measures are normalized to the [0, 1] range and then linearly combined using a hyperparameter, lambda. However, the scores are calculated and normalized differently, and they represent distinct concepts. The mere fact that both metrics share the same range does not imply they are on a comparable scale; the final UQU value could vary significantly depending on the choice of normalization functions or the characteristics of the score distributions, yet the paper lacks an analysis of this sensitivity. For instance, while temperature is a key parameter for the energy score, the paper simply states that T=1 is used without providing any further analysis regarding this choice. Therefore, if linking the two metrics via lambda is central to the analysis, guidelines must be provided regarding the criteria and method for determining this lambda; without such a discussion, the problem of estimating uncertainty could be interpreted as merely shifting into a problem of selecting the lambda value.

Questions for authors

None

2 comments

W1: Lack of sufficient novelty and justification. We agree with the reviewer that linearly combining two scores is simple, and we do not claim it as the main methodological contribution. Our contribution is the globally faithful uncertainty visualization framework that extents existing work in two directions. First, it shows how uncertainty is structured across a whole dataset and its decision landscape, while existing work mostly focuses on a single prediction. Second, existing visualization methods often rely on a probabilistic model output (softmax or likelihood), while our framework removes this restriction. The linear combination is chosen deliberately as the simplest way of combining the two scores such that every proxy can be exchanged, the framework can be easily extended (i.e., by including a third score), and the result stays interpretable.

The combination of the two scores is consistently outperforming the individual scores in our experimental evaluation as the individual scores can often only capture the model's instability under small perturbations or the sample's distributional atypicality, but not both. Hence, they usually perform poorly on the opposite portion of the evaluation set. We will include more experimental results in a revised version of the paper, demonstrating the benefit of the linear combination over the individual scores.

W2: Insufficient justification for the UQU. We agree with the reviewer that sharing the range [0, 1] alone does not make two scores comparable. This is why all scores are normalized in the same/ a similar way using a sigmiod function σ(α⋅(x−t))\sigma(\alpha \cdot(x-t)), where xx represents the value computed by the chosen proxy. The slope α\alpha and the threshold tt are calibrated such that the 95% quantile of values in the training data (we consider them as mostly in-distribution and adversarial free) is maped to 0.50.5 while the mean is mapped to 0.050.05. Both the quantile and the target value for the mean are hyperparameters of our framework and can be adoped. For the energy score, the temperature of T=1 was chosen as it was reported to work the best by the authors of the original paper. In a revised version of the paper, we will make this normalization process and the choice of hyperparameter setting more clear.

3: Lack of comparison: As elaborated in 1, this is a misunderstanding. We compared our related work with other visualizations, our actual contribution. We do not propose a novel uncertainty score, hence we did not look at related work in that area.

Reviewer_icyqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a novel uncertainty score that integrates a model-dependent instability score and a dataset-dependent outlier score through a regularization-based formulation. The instability score is estimated using three approaches: uniform sampling, local intrinsic dimensionality, and maximal verified adversarial robustness. Similarly, the outlier score is computed using four methods: k-nearest neighbors (kNN), HDBSCAN, confidence scores, and energy scores. To evaluate the proposed framework, the authors construct combinations of inlier and outlier samples from four datasets and consider four different baseline models. They subsequently compare the detection accuracy across different combinations of instability and outlier scores to assess the effectiveness of the proposed uncertainty estimation framework.

Critical strengths

The paper is well-motivated and well-written. The problem is clearly stated and the methodology and results are presented with sufficient details.

Critical weaknesses

The major weaknesses of the paper are summarized as follows:

1. Limited methodological novelty and unclear distinction from established uncertainty estimation frameworks: The primary contribution of this paper, which combines instability and outlier scores, appears to revisit concepts closely related to model and data uncertainty, commonly discussed in the context of epistemic and aleatoric uncertainty. In particular, the uniform sampling approach, which aims to capture model disagreement, resembles established uncertainty estimation techniques in Bayesian neural networks, where epistemic uncertainty is commonly quantified through predictive disagreement across model samples. Similarly, adversarial robustness-based scores are closely related to existing perturbation-based approaches for assessing model sensitivity and uncertainty. The paper does not sufficiently clarify how the proposed instability score differs conceptually or methodologically from these established uncertainty estimation techniques. Furthermore, although the outlier score is described as dataset-dependent, some of its implementations rely on model-derived quantities, such as classifier confidence and energy scores. This raises a fundamental question about what the proposed score actually measures: Does it quantify the outlier status of an input xx with respect to the dataset DD, or does it reflect the model's lack of knowledge about xx, as captured by a classifier ff trained on DD? Without a clear theoretical distinction, the separation between model-dependent instability and dataset-dependent outlier scores remains insufficiently justified.

2. Limited novelty and scalability of the outlier detection components: The proposed framework relies on established outlier detection techniques, including k-nearest neighbors (kNN) and HDBSCAN, which are widely used for identifying anomalous samples. Therefore, these components do not independently constitute methodological novelty. More importantly, the computational scalability of these approaches is not adequately addressed. Depending on their implementation, both kNN-based distance estimation and HDBSCAN clustering can introduce substantial computational and memory overhead as the dataset size increases. However, the experimental evaluation is limited to datasets containing approximately 100 to 1,000 samples. Given that modern large-scale machine learning applications frequently involve millions or even billions of samples, it remains unclear whether the proposed outlier estimation framework can be applied efficiently to realistic, large-scale uncertainty quantification tasks. The authors should provide computational complexity analyses and empirical scalability evaluations to establish the practical applicability of their approach.

3. Lack of comparison with recent uncertainty estimation methods and unclear experimental baselines: A substantial portion of the referenced literature is approximately seven to eight years old, and the experimental evaluation does not include sufficient comparisons with recent uncertainty quantification methods. Consequently, it is difficult to assess whether the proposed framework offers meaningful improvements over contemporary approaches. Furthermore, the experimental design does not clearly distinguish the proposed method from the baseline approaches. In particular, it is unclear which experimental configurations represent the authors' proposed method and which serve as baselines. The authors should explicitly identify their proposed configurations, incorporate relevant state-of-the-art uncertainty estimation baselines, and provide a more systematic comparison to substantiate the claimed contributions.

Questions for authors

I would suggest motivating the proposed framework from the perspective of established concepts of model- and data-related uncertainty. In particular, the conventional decomposition of total predictive uncertainty into epistemic and aleatoric components bears a strong conceptual resemblance to the paper's proposed composite uncertainty score. Establishing a clear connection to this existing theoretical framework would help contextualize the proposed approach, clarify its methodological novelty, and distinguish it from established uncertainty quantification techniques.

1 comment
AuthorsOpenReview ↗

Adressing Novelty Concerns comment 1

W1: Novelty: We thank the reviewer for their feedback, and have to first clarify the contribution of our paper before addressing the points raised and answering their questions.

In our paper, we introduce a global uncertainty visualization framework and not a novel uncertainty score. This visualization framework extents existing work in two directions. First, it shows how uncertainty is structured across a whole dataset and its decision landscape, while existing work mostly focuses on a single prediction. Second, existing visualization methods often rely on a probabilistic model output (softmax or likelihood), thus they are not applicable to likelihood-free models such as autoencoders. We remove this restriction and introduce a framework that is applicable to any network architecture/ model type. This is achieved by decoupling the visualization from the model’s output through a modular scoring interface

Our Unsupervised Quantifiable Uncertainty Score captures the two sources of model uncertainty that have been identified as relevant [C. Newen and E. Müller, "Unsupervised DeepView: Global Explainability of Uncertainties for High Dimensional Data," 2022 IEEE International Conference on Knowledge Graph (ICKG), Orlando, FL, USA, 2022, pp. 196-202, doi: 10.1109/ICKG55886.2022.00032.]: the local instability of the model’s behavior under bounded input perturbations and the distributional atypicality of a sample with respect to the models unlabeled training dataset. However, the framework could, in general, also be adopted to any other uncertainty metric.

We agree with the reviewer that regarding a novel uncertainty metric the paper mostly relies on existing methods. However, the goal of the paper was never to introduce a new metric but a globally faithful uncertainty visualization framework that is applicable to every model without restrictions. As such it builds a harness that allows for using individual proxies as interchangeable plug-ins and will therefore benefit from future advances in outlier and adversarial detection.

In a revised version, we will put more emphasis on the visualization part of our framework.

W2: Limited Scalability. We are aware that some of the individual methods can scale poorly with large dataset sizes. We compensate for this fact by introducing a hyperparameter that allows to regulate the amount of data used in the computation of the scores. While taking a random subset of the data can change the precise numbers compared to the entire dataset, our experiments show that the effects almost never change the classification as "certain" or "uncertain" even when cutting out 50% of the training dataset in the evaluation.

Regarding the size of the evaluation dataset used in out experiments, we also want to clarify the evaluation process as the experiments show that the framework scales to large-scale datasets. For a specific dataset DD, we train the model on a large proportion of the data holding back some data for the evaluation. We then compute the set of adversarial examples from these hold-back data points and together with the ood data they constitute our evaluation dataset. In the evaluation, all data-dependent proxies are computed with respect to the training data DD, for instance, given a sample from the evaluation dataset the kNN algorithm searches for the kk nearest neighbours in DD.

We will make our evaluation process more clear in a revised version and will also include more evaluation statics and a complexity analysis.

The contribution of our paper is a global uncertainty visualization framework that can be applied to any model without restrictions. The goal of our experiments is therefore to show that this framework works well across different datasets, network architectures and to compare the effectiveness of different proxies (and their combinations). Therefore, every reported configuration is an instantiation of the framework. The results show that no proxy is the best across all experiments. As a recommended default we identify sampling + kNN, which is the most robust across benchmarks.

Out-of-distribution status is usually treated as a form of epistemic uncertainty, and so is instability under small perturbations (Hüllermeier, E., & Waegeman, W. (2021). Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Machine learning, 110(3), 457-506.). The out-of-distribution (OOD) score, quantifies the distributional atypicality of a sample with respect to a dataset.This atypicality can be computed purely from a dataset DD but it has been shown that certain model characteristics (including reconstruction errors of autoencoders or confidence and energy scores of classifiers) can also function as relative precise proxies and thus contribute a meaningful score (Ruff et al. (2021)).

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

Related papers

Discussion (0)

Sign in to comment.