acceptodds
Under review as a conference paper at ICLR 2027

Faithful by Construction: Hierarchy-Aware Decision Rules for Better Mistakes

Abstract

Hierarchical classification considers settings where label classes are the leaves in a tree , called a taxonomy, and leverages that taxonomy to reduce mistake severity: a misclassified image should land on a similar class and not an unrelated one. Existing methods build the taxonomy into the training recipe, through a loss, soft targets, or feature regularization, but lack a principled account of precisely how these changes affect the decision rule. We ask what it means for a classifier's decision rule to incorporate taxonomic distances exactly. We define prototype rules, classifiers that decide by measuring distances to class prototypes in feature space, and -faithfulness, the property that squared distances between class prototypes match taxonomic distances exactly. We characterize which feature layouts and classifier heads achieve -faithfulness, and propose a simple frozen-head training objective that realizes -faithfulness by construction. An empirical evaluation shows that -faithful models offer competitive accuracy and the lowest mistake severity among several strong baselines. Through a neural collapse analysis of a simple mean-squared-error training objective, we identify two roadblocks that make it difficult to train a -faithful model without freezing the classifier head: we must anticipate and account for the effective backbone regularization strength in advance of training, and the distance-based decision rule is distorted by class-dependent offsets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
hierarchical classification, mistake severity, neural collapse, unconstrained features model, representation geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YFMdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TT-faithful prototype rules for hierarchical classification. A linear classifier is interpreted as assigning features to class prototypes under a Mahalanobis metric, potentially with class-dependent offsets; TT-faithfulness requires these offsets to vanish and the squared distances between prototypes to equal the corresponding taxonomic distances. Theorem 3 characterizes such rules through the pseudo-inverse of the Kron-reduced tree Laplacian. Based on this characterization, the proposed method constructs and freezes a faithful classifier head, then trains the encoder with a KL objective toward taxonomy-derived targets so that features approach their designated prototypes. Theorem 4 separately analyzes an MSE objective in the unconstrained features model and identifies class-dependent offsets and dependence on effective feature regularization as obstacles to learning the same geometry without freezing the head. Experiments on CIFAR-100, FGVC-Aircraft, and Stanford Cars compare against related baselines under uniform and CLIP-derived edge weights. The method obtains the lowest TreeDist@5 in all six evaluated dataset–taxonomy settings, with competitive accuracy on uniform taxonomies and some accuracy loss on CLIP taxonomies.

Critical strengths
  1. A precise decision-rule formulation of hierarchy-aware classification: The paper separates two requirements that are often conflated: the class prototypes must reproduce taxonomic distances, and the class-dependent offsets in the equivalent distance rule must vanish. Theorem 3 then gives an exact, constructive characterization of such TT-faithful prototype rules, while the frozen-head objective realizes the desired decision geometry by construction rather than merely encouraging it through a loss. Although the characterization draws on classical multidimensional scaling and effective resistance, this formulation is clean and useful. It provides a verifiable guarantee about the classifier's actual decision rule, not only about its targets or learned features.

  2. A substantial analysis of why the same guarantee is difficult with a learned head: Theorem 4 derives the geometry induced by an MSE objective with generic admissible targets and identifies two concrete obstacles: unavoidable class-dependent offsets on asymmetric taxonomies and distortion from an effective feature-regularization strength that is not known before training. The predicted geometries are checked across multiple datasets, architectures, targets, and regularizers in §K. The analysis motivates the frozen-head design and provides evidence that the theoretical phenomena occur in real networks. The main comparative experiments also obtain the lowest TreeDist@5 across all six evaluated taxonomy settings.

Critical weaknesses

The empirical evidence is limited to relatively shallow taxonomies: The evaluation covers three datasets with at most 196 leaf classes and taxonomies of only two or three levels. Bertinetto et al. (2020) and the closely related HAFrame work additionally evaluate on tieredImageNet-H (608 leaves, hierarchy height 12) and iNaturalist2019 (1,010 leaves, hierarchy height 7). Since the central contribution is a decision geometry that represents fine-grained taxonomic relationships exactly, comparisons on these deeper label trees would provide useful evidence that the observed benefits extend beyond shallow groupings. The current results are encouraging, but the empirical support for the method in the deeper-taxonomy regime remains limited.

Questions for authors
  1. For HXE (CLIP), how are the CLIP-derived non-uniform edge weights incorporated into HXE's edge weights λ(C)\lambda(C)? Bertinetto et al. (2020) define these weights from node depth, so the adaptation to arbitrary edge weights is unclear.
  2. The paper's empirical claim concerns reduced mistake severity, but the reported TreeDist@kk metrics are averaged over all test examples. Why not also use an error-conditional metric, analogous to the “hierarchical distance of a mistake” in Bertinetto et al. (2020), with LCA height replaced by dTd_T? Please clarify how the reported TreeDist@kk improvements distinguish less severe mistakes from fewer mistakes or better-ranked alternatives.
3 comments
AuthorsOpenReview ↗

Response to Reviewer YFMd

Thank you for your thoughtful review. We respond to each of your concerns below.

Need for deeper trees:

As mentioned in our overall response, we are working on adding two deeper trees (iNaturalist-2019, tieredImageNet-H) to our evaluation. Early results are promising.

How are non-uniform weights integrated in HXE (CLIP)?

This detail should’ve been clearly documented in our original submission, and we regret leaving it out. The HXE loss (equation 4 in [1]) includes a weight λ(Cℓ)\lambda(C^\ell) for each edge Cℓ→Cℓ+1C^\ell \to C^{\ell+1}, which is a function of a node’s height (equation 5 in [1]). We rescaled λ(Cℓ)\lambda(C^\ell) to wℓ⋅λ(Cℓ)w_{\ell} \cdot \lambda(C^\ell), where wℓw_{\ell} is the weight our tree places on the edge. In the uniform trees, wℓ=1w_{\ell} = 1 everywhere and we recover the original objective.

Mistake severity metric:

As mentioned in our overall response, we have added TreeDist|Error to Table 1 to disentangle mistake severity from accuracy. Our method maintains strong performance on the new metric.

Please let us know if you have any further concerns, we would be happy to discuss them with you.

[1] Luca Bertinetto, Romain Mueller, Konstantinos Tertikas, Sina Samangooei, and Nicholas A. Lord. Making better mistakes: Leveraging class hierarchies with deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12503–12512, 2020. Doi: 10.1109/CVPR42600.2020.01252.

Thank you for the response and for the additional experiments. Your clarification on how non-uniform weights enter HXE resolves my first question, and adding an error-conditional severity metric resolves my second: it separates mistake severity from accuracy, and I'm satisfied that the reported gains reflect less severe rather than merely fewer mistakes. I'd be glad to see the complete runs in a revision. I will maintain my recommendation.

Thank you for your reply. Unfortunately, we will be unable to provide full experimental results on the deeper taxonomies before the end of the discussion period; training models on these datasets takes much longer than on the shallower trees used in our initial submission. Our preliminary results on both iNaturalist-2019 and tieredImageNet-H show that our method can make better mistakes (as measured by TreeDist|Err) than HAFrame, a strong baseline. We will include full results in the final version of the paper.

Reviewer_T1JbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a classifier whose prototype distances exactly match distances in a class taxonomy, using a constructed and frozen classification head. It provides a graph-Laplacian characterization and neural-collapse analysis, and reports lower hierarchical distance metrics with generally competitive top-1 accuracy on three image-classification benchmarks.

Critical strengths

Clear geometric formulation and nontrivial analysis. The formulation separates inter-prototype distances from class-dependent decision offsets, and Theorem 3 gives a concise algebraic characterization through the reduced tree Laplacian.

Critical weaknesses

-Insufficient evidence that the improvement comes from T-faithfulness. The main idea is to make distances between class prototypes exactly match the taxonomy. However, compared with existing methods, the proposed approach changes both the classifier head and the training objective. This makes it difficult to determine whether the improvements come from exact T-faithfulness or other design choices. A more direct ablation study comparing (1) a frozen head without taxonomy information, (2) a frozen hierarchy-aware head without exact distance matching, and (3) the proposed T-faithful head, while keeping other training settings consistent could help to better understand the effect of T-faithfulness.

-The connection between the motivation and experimental results is not fully established. The theory guarantees that distances between class prototypes match the taxonomy, but this does not necessarily determine how actual classification errors are distributed. First, TreeDist@1 is influenced by classification accuracy, while TreeDist@5 includes predictions that are not the model's final decision. Second, some improvements in hierarchical distance come at the cost of lower accuracy, particularly with CLIP-weighted taxonomies.

Questions for authors

-Reliability of the taxonomy. The method assumes that the given taxonomy accurately represents class relationships. How sensitive is the method to noisy or imperfect taxonomies? -In Table 2, the uncorrected bias is also used during training. A post-hoc bias-only comparison on the same learned encoder would more cleanly isolate decision offsets. -The method requires a feature dimension of at least K−1, where K is the number of classes. Since the experiments only cover datasets with fewer than 200 classes, it is unclear how well the method scales to thousands of classes.

2 comments
AuthorsOpenReview ↗

Response to Reviewer T1Jb

Thank you for your insightful commentary. We respond to each of your concerns below.

Insufficient evidence that the improvement comes from T-faithfulness:

As mentioned in our overall response, evaluating both “Ours (uniform)” and “Ours (CLIP)” on the CLIP taxonomies separated the effects of the frame from the effects of the KL loss. Taking your advice, we are using our KL loss to train new models with both “hierarchy-agnostic” frames and “hierarchy-aware but not faithful” frames. Preliminary results show that faithfulness offers performance improvements on both the uniform and CLIP FGVC-Aircraft taxonomies (numbers are in the overall response), and we will include full results in a camera-ready version of the paper. Thank you for suggesting these experiments.

We did not state how taxonomic errors are distributed:

We have added TreeDist|Error to Table 1, which measures the average severity of a mistake. Our method maintains strong performance on the new metric.

Some improvements in TreeDist@k come at the cost of lower accuracy:

After taking a second look at our logs, we found that the accuracy loss suffered by “Ours (CLIP)” can, in part, be explained by doing model selection with TreeDist@1, which can reward models both for making better mistakes as well as for making fewer mistakes. If we had done model selection with top-1 accuracy, “Ours (CLIP)” would have recovered about one point of validation accuracy on Aircraft, and still achieved better validation TreeDist@1 than any “Ours (uniform)” model on that dataset.

However, even if we had done model selection using top-1 accuracy, “Ours (CLIP)” would still be about one point behind “Ours (uniform)” on that metric. The real benefit to “Ours (CLIP)” lies in a significant reduction to mistake severity. We stress that this is not a generic consequence of training with non-uniform weights; CPCC and HXE’s CLIP variants did not reliably outperform their uniform variants when evaluated on the CLIP taxonomies. Together with our geometry-alignment plots (Appendix G.3), this supports the assertion that we are realizing the non-uniform CLIP geometry faithfully, and it is precisely this faithfulness that yields our superior mistake severity metrics.

How sensitive is the model to noisy or imperfect taxonomies?

This is a difficult question to answer in our setting, as the taxonomy is used for both training and evaluation. The metrics used here, as in much of the hierarchical classification literature, reward a model for aligning its decisions with a taxonomy even when the taxonomy is flawed. A taxonomy plays two roles: it offers prior knowledge about which classes are more or less similar (which gets incorporated during training) and it offers a cost model for evaluation (i.e., the cost of confusing class u for class v is d_T(u, v)).

However, we believe that the CLIP taxonomies are likely a better representation of visual similarity between classes than the uniform taxonomies, which lets us interpret the “Ours (uniform)” row of Table 1 as being faithful to a noisy and imperfect taxonomy. If we adopt that framing, the superior performance of “Ours (CLIP)” on CLIP evaluation shows that choosing the wrong taxonomy leads to worse outcomes. However, if we compare “Ours (uniform)” only to the baselines, we will find that it is competitive on all CLIP datasets’ mistake severity metrics, which could be read as evidence that faithfulness to an imperfect taxonomy’s frame is still better than no faithfulness at all.

Your comment would be more completely addressed by a new experiment: introducing increasingly large perturbations to a taxonomy (i.e., to its shape or its edge weights) to see how quickly performance degrades. This could only measure how quickly the model's decisions stop aligning with a reference taxonomy, however - it does not guarantee that the reference taxonomy is not itself noisy or imperfect. If time permits, we will run that experiment and share our results.

Post-hoc bias-only comparison on the same encoder:

As mentioned in the overall response, Table 2 has been updated to separate the effects of adding bias correction during training and adding it during inference.

Adapting to large K:

As mentioned in our overall response, early results indicate that our method can adapt to larger trees with ~600-1000 leaves. Further, the newly-added Appendix L offers a simple way to identify the most faithful frame possible when d < K-1, along with a bound on how much any pairwise tree distances will be distorted by the low-dimensional approximation. We leave empirical study in this regime to future work.

Please let us know if you have any further concerns, we’d be happy to discuss them with you.

AuthorsOpenReview ↗

Follow-up Response to Reviewer T1Jb

Hello,

In our initial response to your concern that our experiments did not separate the effects of our KL loss from the effects of our T-faithful frames, we stated that we were working on ablations to address this. We have now updated Table 1 in our paper to contain new rows labeled "ETF+KL" and "HAFrame+KL", which respectively train hierarchy-agnostic ETF frames and hierarchy-aware but non-T-faithful HAFrame frames with our KL loss. These results indicate that improvements in mistake severity come from T-faithfulness rather than the KL loss.

We hope this addresses your concern. Please let us know if you have any further questions or concerns; we are happy to discuss them with you.

Reviewer_6eAJOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To reduce mistake severity in hierarchical classification, the proposed method introduces two rules: the classifier head should output distances from sample features to class prototypes, and the squared distances between prototypes should match taxonomic distances. These rules constrain the model according to the structure of the taxonomy. The authors first construct class prototypes from the taxonomy and set up a matching classifier head so that classification decisions satisfy both requirements. During training, the classifier head is frozen. The probability distribution obtained by passing each class prototype through this head is used as a soft target to supervise predictions for samples of that class, guiding the feature extractor to map samples close to their corresponding class prototypes. The experiments show that the method generally improves mistake severity metrics compared with the baselines. The improvements are more pronounced on CLIP taxonomies, although they may come at the cost of some classification accuracy.

Critical strengths
  1. Through detailed mathematical derivations and proofs, the authors provide a way to compute the prototypes and the matching classifier head parameters. This enables the model to classify samples based on feature-to-prototype distances while ensuring that squared distances between prototypes match taxonomic distances.

  2. The baselines are appropriately chosen, and their settings relevant to the proposed method are clearly described, including how each baseline uses the taxonomy to impose constraints and whether the classifier head parameters are learned or fixed. This makes it easy to compare the experimental metrics across different method configurations.

  3. The problem is clearly defined, and the motivation behind the approach is clearly explained. The authors provide detailed proofs that the optimal solutions obtained through training may exhibit class-dependent offsets, so the classifier head is not guaranteed to classify samples solely based on distances from sample features to prototypes.

Critical weaknesses

Bias correction may change the location of the distribution of h in feature space. The authors require h to have a geometric interpretation as a sample point whose distance to each prototype is used directly for classification. However, h is a learned representation and may have different distributions and interpretations under different bias settings. If h is not required to have this geometric interpretation, class-dependent offsets in the distance formulation of the classifier head's outputs may not constitute a methodological flaw.

The experiments show that bias correction improves performance, but it remains unclear whether these gains are related to the difficulty of fitting the training objective, the optimization process, or regularization.

Also, could this fitting difficulty be related to the training objective and the computational form of the classifier head? Could the prototypes and A be fixed, with distances from sample features to the prototypes computed directly to impose the same feature-space constraints, without explicitly constructing and freezing W and b?

Nevertheless, these concerns do not undermine the effectiveness of the proposed method. The experiments still show that setting the bias so that classification decisions correspond to distances from h to the prototypes can improve performance. There remains room for further discussion of the interpretation and underlying rationale.

Questions for authors

See the weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 6eAJ

Thank you for your comments on our paper. We will respond to each point you raised below.

Learned biases may have different interpretations and distributions:

We agree that in generic classifiers, learned biases do not constitute a methodological flaw. Our goal is to train models whose scores incorporate taxonomic distances exactly, and freezing the biases is the simplest way to achieve that. Learnable biases introduce class-dependent offsets to scores that upset this interpretation.

On whether features relocate under different bias settings: under our KL loss, they do not. The paper notes that the loss is minimized precisely when every sample’s feature vector lands exactly on its class’s prototype. This remains true no matter what bias is used. Empirically, a model trained without bias correction still improves when bias correction is added during inference (Table 2, “inference” columns). If features had adapted to the uncorrected head, adding bias correction at inference would not help.

If T-faithfulness is the goal, then learning biases is not a simple alternative to freezing them. As the last 3 paragraphs of section 5 explain, using a learned head yields class-dependent offsets for most taxonomies, and the realized geometry depends on effective regularization strength, which is unknown in advance of training. Freezing the head and the biases avoids both problems.

Experiments show that bias correction improves performance but it’s unclear where those gains came from:

As mentioned in the overall response, we updated Table 2 to separate the effects of bias correction during training and inference. Using bias correction during training does improve overall accuracy, but using it during inference causes additional improvements to overall accuracy as well as lower mistake severity, regardless of whether it was used during training.

Could the fitting difficulty be related to the training objective?

Theorem 3 assumes that every class is well-represented by a single point in feature space, its prototype. To leverage this result in practice, we need to learn a “collapsed” feature representation in which every class’s features approximately map onto a single point. The theory’s usefulness degrades when features are not collapsed; as Appendix G.5 shows, fitting a faithful readout to an insufficiently collapsed feature distribution costs accuracy.

To answer your question, we think the answer is yes: the KL loss is not integral to the method, and different loss functions (including those that ignore (W, b) and use A-distances) are possible. We would not be surprised if some of these offered more stable and effective training than our KL loss. We updated our draft to mention this as an avenue for future work.

Thanks again for your review, we are happy to discuss any further concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Response to Reviewers

We thank the reviewers for their comments on our paper. We were pleased to see that all reviewers found our paper clear and easy to follow, and that all reviewers said that they learned something interesting or useful from our work. We will respond to each review individually, and describe the overall changes and improvements to the paper here.

Metrics: Two reviewers noted that our TreeDist@k gains might result from making fewer mistakes rather than making better mistakes. We added TreeDist|Err, the mean taxonomic distance from a mis-predicted class and the true class, to our tables. Our method maintains strong performance on the new metric.

Scaling concerns: We are currently scaling our method to deeper trees with ~600-1000 leaves (discussed in the next comment). Nevertheless, the requirement that d≥K−1d \ge K-1 in Theorem 3 (where KK is the number of classes and dd the feature dimension) constrains the scale of taxonomies that admit faithful representations. Theorem 3(i) characterizes faithfulness in terms of a matrix GG being equal to a reduced Laplacian’s pseudo-inverse, LT†L_{\mathcal T}^\dagger. When d<K−1d < K-1, rank conditions prevent equality, but a simple application of the Eckart-Young-Mirsky theorem offers a closed-form expression for the matrix GG that minimizes the Frobenius distance to LT†L_{\mathcal T}^\dagger. Although no prototype rule can be faithful in this regime, we can identify the prototype rules that are as faithful as possible, and also provide a bound on the maximum distortion to taxonomic distances in the representation. We provide details in a new appendix, Appendix L, and leave the empirical study of the d<K−1d < K-1 regime to future work.

Role of the bias/offsets: Two reviewers questioned the role of the bias correction during training. We updated Table 2 to distinguish the effects of bias correction in training and inference. Training without bias correction typically yields worse accuracy than training with it, and evaluating without bias correction yields worse accuracy than with it (regardless of whether it was used during training). However, using bias correction during training does not decrease mistake severity, while adding it during inference does.

KL Loss Ablation: Reviewer T1Jb noted that our method’s gains could be coming from its KL loss rather than its faithful frame. Our original submission did have some evidence that the frame was at least partly responsible for the performance gain, although we did not highlight it: we evaluated both “Ours (uniform)” and “Ours (CLIP)” on the CLIP taxonomies, and “Ours (CLIP)” consistently achieved lower mistake severity. Since the only difference between these two methods is the frame, performance improvements can be attributed to the frame alone.

To more completely separate the effects of the frame from the effects of the KL loss, we took Reviewer T1Jb’s advice and trained new models with different frames and our KL loss. We added ETF+KL, which uses a simplex ETF frame [1] (a hierarchy-agnostic frame) and HAFrame+KL, which uses the hierarchy-aware but un-faithful frame proposed by [2]. For HAFrame+KL, we swept the exact same spacing parameters that the original authors did, so the only difference between HAFrame and HAFrame+KL is the training loss.

We present preliminary results for both the uniform and CLIP FGVC-Aircraft taxonomies here. Only one seed has been run for the two new methods; std’s are given in brackets for the other methods:

FGVC-Aircraft, Uniform
MethodAccuracy (↑)TreeDist|Err (↓)TreeDist@1 (↓)TreeDist@5 (↓)
Ours (Uniform)82.55 (0.95)3.86 (0.04)0.67 (0.03)3.44 (0.00)
HAFrame82.55 (0.35)4.20 (0.06)0.73 (0.02)3.50 (0.01)
HAFrame+KL81.343.920.733.50
ETF+KL81.404.300.804.23
FGVC-Aircraft, CLIP (TreeDist x 100)
MethodAccuracy (↑)TreeDist|Err (↓)TreeDist@1 (↓)TreeDist@5 (↓)
Ours (Uniform)82.66 (0.97)19.07 (0.63)3.30 (0.13)19.56 (0.16)
Ours (CLIP)81.12 (0.21)14.31 (0.80)2.70 (0.17)16.80 (0.02)
HAFrame82.48 (0.40)20.20 (0.59)3.54 (0.09)19.40 (0.005)
HAFrame+KL82.0918.343.2920.16
ETF+KL82.6020.553.5822.99

We are working on more experiments with the two new frames, both adding two more seeds on the FGVC-Aircraft taxonomies and getting results for the other two datasets. We will include the complete results in a camera-ready version of the paper.

(We will continue the overall response in a separate comment.)

AuthorsOpenReview ↗

Overall Response to Reviewers (Continued)

Deeper Trees (iNaturalist-2019, tieredImageNet-H):

Our original submission only considered shallow taxonomies with a modest number of leaves. We are working on adding two deeper taxonomies, iNaturalist-2019 and tieredImageNet-H, per reviewer YFMd’s suggestion.

On these trees, a small extension of our loss makes it competitive with HAFrame, our strongest baseline (numbers below). Our preliminary experiments showed that our unmodified KL loss introduced a trade-off between metrics: sharpening our soft targets boosted top-1 accuracy at the expense of TreeDist@5, while softening them boosted TreeDist@5 at a cost to top-1 accuracy. Since no single configuration optimized both, we kept the sharpened targets and added a hinge loss term, defined as follows:

Lhinge=1N∑i=1Nmax⁡k[14⋅dT(yi,k)−(zyi(gi)−zk(gi))]+,\mathcal{L}_{\mathrm{hinge}} = \frac{1}{N}\sum_{i=1}^{N} \max_{k} \Big[ \frac{1}{4} \cdot d_{\mathcal T}(y_i, k) - \big(z_{y_i}(g_i) - z_k(g_i)\big) \Big]_+ ,

where z(h)z(h) is the readout of a feature vector hh under the frozen frame (i.e., the logits). Every pair of classes has a linear decision boundary, and the hinge loss works to preserve half of the margin to the each such boundary. The KL term supervises fine-grained taxonomic relationships while the hinge supervises more distant relationships.

We report preliminary results on the validation split, along with the number of seeds:

iNat-2019Number of seedsTop-1 accuracy (↑)TreeDist@1 (↓)TreeDist@5 (↓)TreeDist@20 (↓)
HAFrame369.741.22712.30163.4994
Ours369.901.23392.37073.7380
Ours+Hinge370.001.20832.33323.5547
tiredImageNet-HSeedsTop-1 accuracy (↑)TreeDist@1 (↓)TreeDist@5 (↓)TreeDist@20 (↓)
HAFrame374.643.482810.188212.6120
Ours173.783.566610.281212.7857
Ours+Hinge174.273.48559.936412.4953

We stress that our method is not fully tuned yet (i.e., its best numbers came from the edges of its hyper-parameter grids), while the HAFrame runs used the full hyper-parameter grids from its original evaluation. However, our method is already competitive with theirs. We are actively working on a complete evaluation.

We will also reply individually to each reviewer. Thank you again for your consideration of our work.

[1] V. Papyan, X.Y. Han, & D.L. Donoho, Prevalence of neural collapse during the terminal phase of deep learning training, Proc. Natl. Acad. Sci. U.S.A. 117 (40) 24652-24663, https://doi.org/10.1073/pnas.2015509117 (2020).

[2] Tong Liang and Jim Davis. Inducing neural collapse to a fixed hierarchy-aware frame for reducing mistake severity, 2023. URL https://arxiv.org/abs/2303.05689

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.