Hierarchical Selective Classification

Deploying deep neural networks for risk-sensitive tasks necessitates an uncertainty estimation mechanism. This paper introduces hierarchical selective classification, extending selective classification to a hierarchical setting. Our approach leverages the inherent structure of class relationships, enabling models to reduce the specificity of their predictions when faced with uncertainty. In this paper, we first formalize hierarchical risk and coverage, and introduce hierarchical risk-coverage curves. Next, we develop algorithms for hierarchical selective classification (which we refer to as "inference rules"), and propose an efficient algorithm that guarantees a target accuracy constraint with high probability. Lastly, we conduct extensive empirical studies on over a thousand ImageNet classifiers, revealing that training regimes such as CLIP, pretraining on ImageNet21k and knowledge distillation boost hierarchical selective performance.

Paper

Similar papers

Peer review

Reviewer 9FfT6/10 · confidence 3/52024-06-14

Summary

The authors propose hierarchical selective classification, a method that selects the hierarchical granularity of its prediction based on uncertainty.

Strengths

* The paper is well-written, and the proposed method is quite intuitive. * I like the idea that if uncertain, it makes sense to predict at a higher level of granularity. * The theoretical results & statements are sound. * Extensive experimental results showing the superiority of the proposed method to DARTS and a non-hierarchical baseline. * Applicability to pre-trained models

Weaknesses

* My biggest uncertainty is the similarity of this work to conformal prediction. To me, it seems that this method is very similar to conformal prediction, where the set of possible prediction sets is restricted via this pre-defined hierarchy. While, as far as I know, it has not been explored, it decreases the perceived novelty. * A weakness of the setting rather than the method is that it assumes the knowledge of the underlying hierarchy. As such, the applicability is somewhat limited. The paper would benefit from a way to unsupervisedly learn this hierarchy, e.g. based on classes whose predicted probabilities are positively correlated. * As also touched upon in the concluding remarks, the method is post-hoc rather than being optimized during training, thus, likely not performing up to the highest possible level. * Minor: Line 158-159 is a worded badly, similar to "... thus, we do A. Unlike others that do A, we do A+B".

Questions

* Could the authors please comment on the similarity to conformal prediction? * As such, how is [1] related and/or different? [1] Tyagi, Chhavi, and Wenge Guo. "Multi-label Classification under Uncertainty: A Tree-based Conformal Prediction Approach." Conformal and Probabilistic Prediction with Applications. PMLR, 2023.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are adequately discussed apart from the necessity of assuming knowledge of the underlying class hierarchy.

Reviewer XLCN5/10 · confidence 4/52024-06-19

Summary

The paper introduces a hierarchical selective classification technique that incorporates hierarchical risk and coverage. The authors additionally proposed an algorithm that guarantees target accuracy. Experimental results demonstrate the method's effectiveness.

Strengths

Hierarchical selective classification is a new area and therefore the current method is one of the first techniques to deal with such problem. Its application to critical settings can be substantial.

Weaknesses

• The need of a prior tree among classes can limit its usage for complex scenarios. The construction of such tree can be a non-trivial step for the applicability of the approach. • The main contribution looks an extension of previous methods for the hierarchical case.

Questions

• Regarding results like in Table 2, would be possible to calibrate the coverage (as done on selective networks) for fair comparison? • Have the authors thought about the building the hierarchical tree structure as a pre-processing step? I asked that because such prior is key for wide applicability. • A well know problem of selective approaches, exposed on [1] is that given the non-differentiability of selection mechanism the binary function g is replaced by a relaxed function g: X → [0, 1], that way not performing selection during training, but instead assigning a soft instance weight to each training sample. The same effect is observed in the proposed method? [1] Gumbel-Softmax Selective Networks, https://arxiv.org/pdf/2211.10564.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

• The requirement for a hierarchical class tree prior to using the method is a limitation. • The current experimental section is limited in terms of baseline approaches and datasets, making it challenging to gain a comprehensive understanding from the existing experiments.

Reviewer vmdY6/10 · confidence 3/52024-07-02

Summary

The paper introduces a new framework for selective classification called hierarchical selective classification. In a setting where a hierarchy in the classification task is present, the authors devise a selection strategy that considers confidence at different levels of the classification hierarchy. Extensive experimental analysis is performed over 1115 ImageNet classifiers.

Strengths

The main strengths of the paper are: 1. the idea of applying selective classification in a hierarchical setting is novel; 2. the theoretical analysis relies on conformal prediction, which guarantees the soundness of the results; 3. the proposed framework can impact high-risk settings, as shown in the healthcare example.

Weaknesses

Overall, I think the paper is solid. My main concern is that the empirical evaluation could be improved, especially regarding motivations and attention to detail. A few examples: * I do not fully understand why the authors focus so much on showing how different training regimes affect HSC performance. I guess this improves the overall predictive performance of the (hierarchical) classifier, which is expected to impact the HSC task positively. * As the authors correctly claim, the training regimes were not optimized for hierarchical selective classification. Despite the clear computation-wise motivation, I argue that including regimes optimized for HSC would make the empirical evaluation sounder. * a few lines are off: for instance, I would argue that line 279, i.e., >CLIP achieves an exceptional improvement, surpassing 40% > does not match what is shown in Figure 4 (which shows an improvement below 40%).

Questions

I have a few questions/remarks regarding the paper. * Q1. I think the authors are not discussing an implicit (and, in my opinion, quite relevant) assumption of their strategy, i.e. the errors at different levels of the hierarchy are assumed to be the same. However, I argue this is not exactly the case in real life. For example, failing to distinguish a golden retriever from a labrador differs from failing to distinguish a dog from a spider. Can the authors elaborate on this point? * Q2. Can the authors discuss the points I highlighted as the main weakness?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The paper briefly discusses limitations. I think this section could be expanded, e.g. considering Q1.

Authorsrebuttal2024-08-07

Rebuttal Part 2 for Reviewer vmdY

*"Q1. I think the authors are not discussing an implicit (and, in my opinion, quite relevant) assumption of their strategy, i.e. the errors at different levels of the hierarchy are assumed to be the same. However, I argue this is not exactly the case in real life. For example, failing to distinguish a golden retriever from a labrador differs from failing to distinguish a dog from a spider. Can the authors elaborate on this point?"* That's a great question! In fact, we conducted all of our experiments using a hierarchical risk that considers scenarios like the ones you've illustrated. We chose not to include these results in the paper due to space constraints and for the sake of simplicity and readability. We developed a hierarchical risk that considers mistake severity with respect to the hierarchy, as follows: $R_h = 1 - \frac{\phi(LCA(\hat{v}, v))}{\phi(\hat{v})}$ Where $v$ is the ground truth node, $\hat{v}$ is the predicted node, $\phi(x)$ is the coverage of node $x$, and $LCA(x,y)$ is the lowest common ancestor of nodes $x$ and $y$. The most severe mistake occurs when $LCA(v,\hat{v})$ is the root of the hierarchy, resulting in a risk of 1. The higher the coverage of $LCA(\hat{v},v)$, the more $\hat{v}$ has in common with $v$ in hierarchical terms. One can safely assume that in any plausible hierarchy the LCA of (labrador, golden retriever), for example, dog, has significantly higher coverage compared to the LCA of (dog, spider), resulting in the former misclassification having lower risk compared to the latter. We provide below some of the results of our experiments on ImageNet, complementing Table 1 in the paper (all other tables and figures can also be made with this type of loss): | | ImageNet-1k (1115 models) | | |:--------------:|:-------------------------:|:-------------:| | Inference Rule | hAURC | Hier. Gain | | | $\times$ 1000 | (\%) | | Selective | 24.98$\pm$0.29 | - | | MC | 24.50$\pm$0.34 | 3.31$\pm$0.30 | | Climbing | 24.04$\pm$0.31 | 4.71$\pm$0.22 | The results align with those presented in the paper, where the risk is 0/1 loss. Climbing remains the inference rule with the best (lowest) hAURC and the highest hierarchical gain. It's theoretically possible to create examples where the hierarchical risk diverges from the 0/1 loss, such that models making less severe mistakes are penalized less by the hierarchical risk. However, in practice, the differences are smaller than we anticipated and are very highly correlated (which is interesting by itself). Do you think we should include these results and others with this risk in the revision? [1] Ido Galil, Mohammed Dabbah, Ran El-Yaniv. What Can We Learn From The Selective Prediction And Uncertainty Estimation Performance Of 523 Imagenet Classifiers, ICLR 2023. https://arxiv.org/pdf/2302.11874 [2] Guo et al. On Calibration of Modern Neural Networks. [3] Vladimir Vovk. Conditional validity of inductive conformal predictors. [4] Cattelan and Silva. How to Fix a Broken Confidence Estimator: Evaluating Post-hoc Methods for Selective Classification with Deep Neural Networks.

Reviewer h9cA6/10 · confidence 3/52024-07-10

Summary

The paper proposes an extension of selective classification following a class hierarchy to reduce the specificity of model prediction when there is a high uncertainty. In particular, if the prediction confidence of a class is smaller than a predefined threshold, the proposed algorithm would proceed towards a higher class level in the hierarchical structure (the parent node), until the confidence of the considering node exceeds that threshold. The paper also formulises hierarchical risk and coverage, so that the area under curve can be used as a metric to benchmark different selective classification methods. An extensive number of pretrained classifiers on ImageNet dataset are then used to evaluate the proposed method and show promising results. The paper also include a PAC-like theoretical result, so that when finding the optimal threshold, one can select appropriate hyper-parameters to achieve their desired outcome with certain confidence level.

Strengths

The paper goes into details to provide an adequate background about selective classification, the definition of heirarchical risk and coverage as well as its area under curve as a metric to quantify the performance of hierarchical-based selective classification. It also links to previous studies in the same subfield. In general, the paper is well written and easy to follow. The paper also includes a theoretical result on the guarantee of the learning algorithm when one wants to find the optimal thresholding value for their hierarchical selective classification. This simple theoretical results does strengthen the paper. The paper also include an extensive number of experiments and ablation studies to provide insights into the newly-proposed method.

Weaknesses

The paper relies on the setting with the following assumptions: - It is an inference rule. This means that the algorithm is used at test time only. If this could be even integrated into training is a plus. - It needs a validation set to find the optimal hyper-parameter $\theta$, or the threshold (partly mentioned in the conclusion). It is understandable because there is no training involve here, so there is a need for that. However, in some cases, there may not be additional data available.

Questions

Could the authors clarify if it can also be integrated into training a whole model to perform hierarchical selective classification?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer 9FfT2024-08-08

Thank you for your response. The responses have cleared up my uncertainties, and I have adjusted my score. Regarding the end-to-end training, I believe the extension is interesting and might benefit future work that is derived from the current manuscript. As such, I would appreciate this extension in either the Appendix or as a separate work, whereby I leave it up to you to decide which you believe to be more feasible.

Reviewer vmdY2024-08-10

I thank the authors for their responses. I have a couple of remarks/comments: > From the results below, it seems to us that the improvement in hierarchical performance is significantly more pronounced than the one in accuracy. If I get how improvements are computed (i.e. $\Delta(Metric)=\frac{|Metric_{new} - Metric_{orig}|}{Metric_{orig}}$), I think it is not that straightforward to compare the % improvements of two different metrics. I feel this is affected by the original baseline value starting point. Given an absolute improvement, the relative improvement increases as the baseline value decreases, and vice versa. I.e., an absolute improvement of $\delta$ on baseline values $b_1, b_2, b_1 << b2$ results in relative improvements $r_1, r_2, r_1 >> r_2$. Since the lower the hAURC, the better, and the higher the accuracy, the better (as in the just presented example); seeing higher percentage improvements might be misleading, and I would be cautious about the conclusions here. > We initially chose not to include this analysis in the paper due to concerns about the paper being already packed with too much information. Do you think it would be preferable to include these results in the revision? I think these results could be safely added to the Appendix, as they could explain why we see improvements in hAURC through different training regimes. > However, in practice, the differences are smaller than we anticipated and are very highly correlated (which is interesting by itself). Do you think we should include these results and others with this risk in the revision? Concerning the results w.r.t. the hierarchical loss, the authors could add the results in the Appendix. In light of these comments, I am still in favour of a positive score for this paper.

Area Chair ne2v2024-08-11

Dear Reviewers, The discussion period ends within 3 days. Authors made tremendous efforts on providing responses to your concerns. So, if you haven’t yet, please check responses and express your opinions whether you want to initiate discussion or accept the authors’ responses. Best, Your AC

Reviewer XLCN2024-08-11

I appreciate the authors' rebuttal. I would raise my score from 4 to 5 in response to the authors' rebuttal. I will also discuss with the other fellow reviewers and/or AC.

Reviewer h9cA2024-08-12

Comments by Reviewers h9cA

Thank you, the authors, for clarifying my concerns. I have a positive view of the paper, although I am not very familiar to the research field. Hence, I keep my rating as is.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC