Summary
The motivation behind the work proposed in this paper is common among most out-of-distribution (OOD) data detection methods in machine learning: deploying an ML model could lead it to encounter near-in-distribution (NID) or OOD data. Evidently, this can cause severe errors in its ability to fulfil its intended task correctly. The authors propose a method, inspired by checksum functions, to add another output node to the neural network and modify the loss function to encourage the checksum node to flag OOD data.
Strengths
- Even though the proposed work lacks significant novelty, I do believe that there is some promise behind it. The, albeit limited results, show signs of the method working reasonably and with some further/major refinement could offer a nice solution.
- Figures throughout are useful and relevant. They are a positive inclusion.
- The discussion section is transparent and honest, openly addressing both the strengths and limitations of the study. This level of frankness is commendable.
Weaknesses
- Firstly, the work seems to be built upon work in [1]. The paper includes no reference to the work, and the novelty past it is very limited.
- The 'Related work & Background' section is severely limited. In the uncertainty quantification discussion, there are no remarks to very popular methods such as conformal prediction, test-time augmentation, evidential deep learning, variational inference, etc. I find this poor especially since some parallels can be drawn between EDL and this paper.
- The writing in the paper would benefit from improved organization and clarity. Currently, certain arguments or points appear prematurely in earlier sections (e.g., points relevant to Section 5 are raised in Section 3), which disrupts the logical flow of the content. This causes the text to read unevenly and creates an informal tone that may detract from the paper’s scientific rigour. I recommend restructuring to ensure each section clearly builds on the previous one.
- The use of the sum of predictions as a checksum for identifying in-distribution (ID) or out-of-distribution (OOD) data lacks sufficient mathematical rigour and robustness. While the authors imply that this approach checks for consistency to distinguish between ID and OOD cases, the method appears overly simplistic and insufficiently substantiated. Without clear justification or evidence supporting the effectiveness of this metric, it is unclear how well it can reliably capture the complexity of OOD detection.
- While the application of the proposed approach to Non-Local Thermodynamic Equilibrium (NLTE) calculations is both valid and relevant, it would benefit from broader experimental validation. Given the generalizable nature of the method, an additional evaluation on standard machine learning datasets or synthetic data would strengthen the study's claims and allow for a more comprehensive assessment of its applicability across domains. Benchmarking on common or synthetic datasets would also provide clearer comparisons to existing methods and better illustrate the versatility and robustness of the approach.
- In Figure 2, the boundary between ID and OOD data is abrupt, despite minimal differences near the threshold. Introducing a near-ID (NID) dataset could create a gradient at the boundary, allowing for a more nuanced transition and improving the model’s handling of data close to this threshold. I expect that the majority of false negatives reported in Table 1, could potentially show as NID data instead, therefore improving your results and increasing transparency into the limitations of the work.
- Was any ablation conducted on the threshold value to assess its impact? Were results better at a slightly lower threshold?
- The paper claims a novel method for exposing the model to random OOD data without biasing it toward a limited region, but there is no clear evidence or demonstration of this in the results. This claim would benefit from supporting experiments or explanations to substantiate it.
[1] Możejko, M., Susik, M. and Karczewski, R., 2018. Inhibited softmax for uncertainty estimation in neural networks. arXiv preprint arXiv:1810.01861.
Small weaknesses
- Separating the conclusion of the paper from the Discussion section could be useful for the readability of the paper.
- The paper devotes considerable attention to explaining basic, well-understood concepts, such as the use of a validation dataset. This level of detail seems unusual, as it may be redundant for the intended audience.
Questions
Can the authors comment on the following:
- Can you clarify how the sum of predictions as a checksum value reliably distinguishes between in-distribution (ID) and out-of-distribution (OOD) data? Is there a theoretical basis or empirical evidence to support its robustness?
- Have you considered validating the approach on standard machine learning or synthetic datasets to demonstrate its broader applicability?
- The paper mentions a novel method for exposing the model to random OOD data without biasing it. Could you provide more details or experimental evidence for this claim?
- Was any ablation conducted on the threshold values used in the model?
- Could you explore creating a "near ID" dataset to provide a smoother transition at the ID-OOD boundary and possibly improve results?