- 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
-
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.
-
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
-
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.
-
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
-
(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.
-
(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?
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 we chose a value of 0.5, 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 0.05. As a simple default we set the final detection threshold to a value of 0.3 (Note that this is slightly higher than scoring an individual score of 0.5 in either the instability or out-of-distribution score due to the factor of λ=0.5). We will make the default values of these and all other hyperparameter more clear in a revised version of the paper.