Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift

Fahimeh Fooladgar; Michael Cooper; Minh Nguyen Nhat To; Mohamed Harmanani; Parvin Mousavi; Paul F RWilson; Purang Abolmaesumi; Rahul G. Krishnan; Viet Nguyen

REVIEW 3 major objections 6 minor 1 cited by

Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that replacing a trained network's final linear layer with an ensemble of diverse per-class prototype classifiers improves worst-group accuracy under subpopulation shift, even when subgroup annotations are unavailable.

desk verdict A genuinely new last-layer ensemble method with strong numbers, but the 'no subgroup annotations' claim is undercut by validation WGA selection, so it needs protocol fixes before the headline is credible. read the letter →

arxiv 2505.23027 v1 pith:EXDHAH3D submitted 2025-05-29 cs.LG cs.AIcs.CV

Minh Nguyen Nhat To , Paul F RWilson , Viet Nguyen , Mohamed Harmanani , Michael Cooper , Fahimeh Fooladgar , Purang Abolmaesumi , Parvin Mousavi

show 1 more author

Rahul G. Krishnan

This is my paper

classification cs.LGcs.AIcs.CV

keywords diverseprototypicalensemblessubpopulationshiftworst-groupaccuracyspuriouscorrelationsattributeimbalancegeneralizationlast-layerretrainingensemblediversity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that subpopulation shift can be countered without knowing the subgroup structure in advance. Instead of reweighting examples by group membership, it swaps the final linear layer of a trained network for an ensemble of prototype classifiers, with several prototypes per class, trained to disagree with one another. The claim is that this diversified prototypical ensemble discovers latent subgroups on its own and, by averaging predictions across members, keeps the worst-case subgroups from being sacrificed to majority accuracy. If true, robustness to spurious correlations, attribute imbalance, and attribute generalization becomes a modular last-layer fix that can be added to standard training pipelines. Across nine real-world datasets, the method reports higher worst-group accuracy than prior methods, with and without subgroup labels available for validation.

What carries the argument

The load-bearing object is the Diverse Prototypical Ensemble: for each of the $K$ classes, $N$ learnable prototypes $p_j^{(k)}$ replace the weight vectors of the final linear layer. Each prototype is a point in the frozen feature space, and classification uses a softmax over scaled Euclidean distances between the input feature and the class prototypes, averaged over the $N$ ensemble members. Training is sequential: at stage $n$, new prototypes are optimized with the classification loss plus $L_{IPS}$, which penalizes the absolute inner product between distinct prototypes of the same class, while earlier prototypes are frozen. Each new member is trained on a different class-balanced (or, when available, attribute-balanced) bootstrap of the validation data, so diversity is imposed twice, once by the loss and once by the sampling. This mechanism is what lets the ensemble cover different decision boundaries and latent subgroups.

What would settle it

Rerun the no-annotation experiments but choose all hyperparameters and checkpoints using only average validation accuracy or another criterion computable without group labels, then measure worst-group accuracy on the test set. If DPE's advantage over ERM* shrinks or vanishes, the annotation-free claim fails; if the margin persists, the claim survives. A cheaper audit is to inspect the released code to see whether the validation-fold worst-group accuracy used for selection is computed from known subgroup labels.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a sequentially trained ensemble of per-class prototypes, each fitted on its own balanced bootstrap subset and pushed apart by an inter-prototype similarity loss, acts as an implicit subgroup learner. Across nine benchmark datasets, the method reports an average worst-group accuracy of 73.9% without subgroup annotations, against 57.7% for ERM and 65.2% for DFR; with a stronger augmented backbone, worst-group accuracy reaches 94.1% on Waterbirds, 84.6% on CelebA, and 63.0% on Living17. When subgroup labels are available on a validation subset, the average rises to 83.0%, with 94.1% on Waterbirds, 90.3% on CelebA, and 91.7% on MetaShift. The authors interpret these numbers as evidence that prototype diversity, rather than representation quality alone, is what recovers underrepresented subpopulations.

Load-bearing premise

The no-annotation claim rests on tuning and selecting the model by worst-group accuracy on a validation fold, but that metric is defined over known subgroups; if the labels needed to compute it are unavailable, DPE's gains may depend on group information after all.

Editorial extensions

If this is right

If DPE works as claimed, worst-group robustness becomes a plug-in head: it can be applied to any frozen ERM feature extractor without retraining the backbone.
The no-annotation results suggest that explicit subgroup identification, a bottleneck for methods like JTT and CnC, is not needed for training-time robustness.
DPE's largest margins appear in attribute-imbalance and attribute-generalization settings such as CheXpert, NICO++, and Living17, where the paper says most existing methods fail to beat ERM.
The ensemble-size ablation shows improvements saturate around 15 prototypes, so the gains do not require hundreds of classifiers.
When validation subgroup labels are available, DPE can be combined with group-balanced subsampling to push worst-group accuracy higher, so annotations remain useful when they exist.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the ensemble trains on precomputed features, DPE could be stacked on any feature extractor, including self-supervised ones; the paper only lists this as future work, but the modularity makes it a natural next test.
The nearest-prototype structure could be used as a pseudo-grouping: validation points closest to different prototypes might seed audit sets or group-balanced training, an extension the paper does not pursue.
The annotation-free claim is only as strong as the selection procedure: the paper tunes and selects models by worst-group accuracy on a validation fold, which requires knowing the subgroup partition, so a fully label-free variant would need a selection criterion that does not use group membership.
The sequential freeze-previous-prototypes schedule resembles residual or boosting-style fitting in feature space; whether the ordering matters, or whether an unordered diversity regularizer gives the same coverage, is a testable extension the paper leaves open.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the thing you should know: the method is a real variant of last-layer retraining for subpopulation shift, and the reported worst-group gains are mostly real, but the headline 'no subgroup annotations' claim does not survive contact with Section 4.4. The authors select hyperparameters and the final model by validation worst-group accuracy, and WGA is defined over group partitions (Appendix A.2). That means group labels are doing selection work in Table 1. The training stage only needs class labels, so this is a selection leak, not a training leak—but it changes the claim from 'works without subgroup annotations' to 'works when you use subgroup labels to tune and pick the model.' That is the DFR/CRT setup, which also uses a labeled validation set.

What is actually new: a per-class prototype ensemble, trained sequentially, with an inter-prototype similarity loss and bootstrap subsets, replacing the linear head. The ablations are informative: diversity via IPS plus random subsets beats fixed subsets, prototype ensemble beats linear ensemble, and gains saturate around N=15. Code is linked. The paper runs the usual benchmarks and gets strong numbers on Waterbirds, CelebA, and Living17. The limitations section is honest about the lack of theory.

Soft spots beyond the selection leak. (1) CheXpert is missing for ERM+DPE in Table 1 even though the paper says nine datasets in both settings. (2) The sensitivity claim 'within 1–2%' is contradicted by their own Appendix C.5: Living17 WGA ranges from 55.3 to 63.0 across temperature values, about 7.7 points, not 1–2%. (3) The exact 50.0 with zero standard deviation on NICO++ for both ERM+DPE and ERM*+DPE looks odd; maybe it is a degenerate case, but it needs explanation. (4) Comparing DPE on a stronger ERM* backbone against SubpopBench baselines on an ERM backbone conflates backbone and method; the top/bottom split in Table 1 partially acknowledges this, but the 'average 73.9% vs ERM 57.7%' sentence in Section 5.1 mixes the two. Also, the paper itself says 'does not require subgroup attribute annotations' at the second stage, which is true but weaker than 'without subgroup annotations' in the abstract.

Verdict: worth refereeing, not worth accepting as is. The protocol needs fixing or the claim needs reframing. If the authors rerun the no-annotation setting with a label-free selection criterion, or explicitly say that group labels are used to tune a last-layer retraining approach like DFR, the paper becomes solid. I would send it to review, with heavy revision expected.

Referee Report

3 major / 6 minor

Summary. The paper proposes Diversified Prototypical Ensembles (DPE), a method that replaces the linear classification head of a pretrained feature extractor with an ensemble of distance-based prototype classifiers. Each prototype is trained sequentially on a class-balanced subset of the validation set with an inter-prototype similarity loss intended to encourage diversity. The authors claim that DPE improves worst-group accuracy under subpopulation shift without requiring subgroup annotations, and they evaluate on nine real-world datasets from the SubpopBench suite in both annotated and unannotated settings. The main reported results are that DPE improves worst-group accuracy over ERM and several baselines in a no-annotation protocol, and that adding subgroup annotations further improves performance.

Significance. If the no-annotation claim held, DPE would be a practically valuable contribution to robustness under subpopulation shift, since existing methods often require group labels for training or model selection. The paper also provides a useful modular design: the method is computationally light (only the classification head is trained on precomputed features) and the code is released. The ablation studies on ensemble size, diversification strategies, and hyperparameter sensitivity are informative and suggest the method is relatively easy to use. However, the headline claim of operating without subgroup annotations is not supported by the evaluation protocol, as detailed below, so the actual contribution is currently closer to a last-layer retraining method that, like DFR, relies on a group-labeled validation set for hyperparameter and model selection.

major comments (3)

[Section 4.4, Table 1, Section 5.1] The central claim that DPE improves worst-group accuracy without subgroup annotations is undercut by the model selection protocol. Section 4.4 states that 'both hyperparameter tuning and model selection are based on the worst-group accuracy within the validation fold.' Since worst-group accuracy is the minimum accuracy over subgroups (Appendix A.2), computing it requires subgroup labels on the validation fold. Thus the results in Table 1 are not obtained under an annotation-free protocol: group labels are used indirectly to choose hyperparameters (such as τ, α, and the number of prototypes) and to select the final model among seeds. This is a selection leak rather than a training leak, but it invalidates the abstract's claim that DPE works 'without prior knowledge of subgroup identities.' I recommend that the authors either re-run the no-annotation experiments with a label-free selection criterion (for example, using average validation accuracy, a fixed default hyperparameter setting, or a validation criterion that does not require group partitions), or explicitly re-scope the contribution to state that DPE requires group-labeled validation data for tuning and selection while only the ensemble training itself is annotation-free. The current wording conflates these two very different settings.
[Table 1, Section 5.1] Table 1 reports a dash for ERM+DPE on CheXpert, yet Section 5.1 states that 'on difficult datasets like CHEXPERT, DPE achieves a WGA of 76.8%.' That number is from ERM*+DPE in the bottom half, not from the ERM+DPE row. This is a reporting inconsistency that matters because the paper claims evaluation on 'nine real-world datasets' and reports an average WGA of 73.9%, but the table does not contain a complete set of ERM+DPE results. Please clarify which backbone and which sets of datasets are used for each reported average, and either fill in the missing entries or adjust the text to avoid implying a complete nine-dataset comparison for ERM+DPE.
[Section 4.2 and Table 1] The labeling of Table 1 as 'without subgroup annotations' is confusing because several baselines listed there, such as DFR and CRT, are known to use subgroup labels for validation and retraining (as acknowledged in the paper's related-work description of DFR). Section 4.2 says that the group-annotated validation set is available 'for model selection, hyperparameters tuning, and re-training the classifier head' and that DPE is evaluated 'in the absence of explicit subgroup annotations.' This creates an ambiguity about whether the no-annotation setting is defined only by the training stage or also by the tuning and selection stage. The distinction is critical for a fair comparison and for the scientific claim. I ask the authors to make the protocol definition explicit and to ensure that all compared methods are treated under the same conditions with respect to tuning and selection.

minor comments (6)

[Equation (4)] The notation in the inter-prototype similarity loss is hard to follow: L_IPS is written per ensemble member n, but the summation indices i and j range over ensemble members, and the normalization includes both n and d. Please clarify which prototypes are frozen and which are being optimized at each stage, and align the notation with Algorithm 1.
[Algorithm 1 and Section 3] Algorithm 1 says 'Sample class-balanced subset D_sub from D_val' in Stage 2, but the text mentions that group-balanced subsets are used when annotations are available. Please state in the algorithm both variants and when each is used.
[Figure 1] Figure 1 is very dense and the text labels are difficult to read at the printed size. Please enlarge the figure or simplify the caption so that the pipeline steps (feature extractor, prototype ensemble, IPS loss, and the visualization of centroids) are legible.
[Table 1] The column header 'C IVILCOMMENTS' contains an erroneous space, and several table cells (e.g., NICO++ WGA of 50.0 for both ERM+DPE and ERM*+DPE) look suspicious; please double-check these values and add a footnote explaining any rounding or reporting conventions.
[Section 5.1] When reporting the average WGA of 73.9%, specify exactly which rows and datasets are averaged. The table contains dashes and a mix of ERM+DPE and ERM*+DPE results, so the reader cannot reproduce the average.
[Appendix C.7] The exploratory analysis using ChatGPT to interpret prototype-closest samples is qualitative and potentially non-reproducible. If it is kept, please describe the exact prompt, the version of ChatGPT used, and how the authors verified the reliability of the generated interpretations.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical benchmark paper with no circular derivation; the no-annotation tuning protocol issue is a soundness concern, not circular reasoning.

full rationale

This is an empirical benchmark paper, not a formal derivation, and I find no step in which a reported result reduces by construction to its own inputs. DPE's training procedure (Algorithm 1) optimizes a prototypical classification loss with an inter-prototype similarity loss on class-balanced validation subsets; the reported worst-group accuracies are measured on held-out test splits from SubpopBench, so the numbers are not fitted quantities renamed as predictions. The only self-citation (Ginsberg et al. 2022, co-authored by R. Krishnan) is used as general motivation for ensemble diversity and is not load-bearing for the main claim. The paper's Limitations section explicitly acknowledges the absence of a formal theoretical explanation and the qualitative nature of the prototype-subgroup alignment analysis, which is appropriately framed as exploratory. The most salient issue is in Section 4.4: 'both hyperparameter tuning and model selection are based on the worst-group accuracy within the validation fold,' while worst-group accuracy, as defined in Appendix A.2, requires subgroup partitions. This means the Table 1 'without subgroup annotations' results may rely on group labels for model selection, so the annotation-free claim is not fully established by the reported protocol. However, this is a validity or soundness concern, not circular reasoning: tuning hyperparameters on validation worst-group accuracy is standard practice, and it does not make the test-set predictions equivalent to the tuning objective by definition. No equation in the paper is defined in terms of the quantity it claims to predict, and no uniqueness theorem is imported from the authors' own prior work.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical or ontological entities are introduced; the prototypes are learned parameters in a classifier head. The main ledger items are tuned hyperparameters and domain assumptions about ERM features, class-balanced sampling, and the use of validation WGA for selection.

free parameters (4)

Number of prototypes per class N = 15
Set from a saturation ablation in Section 6.3; N=15 gives most of the gain, while N=25 and N=40 add nothing.
Temperature tau = Inverse temperature 10 to 40, dataset-specific
Tuned on validation worst-group accuracy per Section 4.4 and Appendix C.5; not a fixed constant.
IPS loss weight alpha = 1e5 to 5e5 depending on dataset
Listed per dataset in Table 11 and selected by validation WGA tuning.
Validation subset sampling scheme = Unspecified
Class-balanced or group-balanced subsets are sampled from validation data, but subset size and resampling protocol are not reported in the text, making exact reproduction dependent on the code.

assumptions (4)

domain assumption ERM-trained feature representations retain enough subgroup information for last-layer retraining to be effective
Invoked in Section 3, citing Izmailov et al. and Kirichenko et al.; the entire two-stage design depends on frozen ERM features being sufficient.
domain assumption Class-balanced sampling of the validation set is a sufficient proxy for unknown subgroup structure
Stage 2 in Algorithm 1 samples class-balanced subsets when group labels are unavailable; no evidence is given that class balance alone covers the worst subgroups.
ad hoc to paper Worst-group accuracy on the validation fold can be computed without subgroup annotations
Section 4.4 states model selection and hyperparameter tuning use validation WGA, which requires group labels; this is inconsistent with the no-annotation setting in Table 1.
ad hoc to paper Sequential freezing of earlier prototypes while optimizing the current one produces the intended diversity
Algorithm 1 and Eq. 4 rely on this optimization schedule; the paper offers no theory, only an empirical ablation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift." pith.science (2026). https://pith.science/paper/EXDHAH3D

@misc{pith2026250523027,
  author       = {Pith},
  title        = {Pith review of: Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EXDHAH3D}},
  note         = {Machine review of arXiv:2505.23027}
}

read the original abstract

The subpopulationtion shift, characterized by a disparity in subpopulation distributibetween theween the training and target datasets, can significantly degrade the performance of machine learning models. Current solutions to subpopulation shift involve modifying empirical risk minimization with re-weighting strategies to improve generalization. This strategy relies on assumptions about the number and nature of subpopulations and annotations on group membership, which are unavailable for many real-world datasets. Instead, we propose using an ensemble of diverse classifiers to adaptively capture risk associated with subpopulations. Given a feature extractor network, we replace its standard linear classification layer with a mixture of prototypical classifiers, where each member is trained to classify the data while focusing on different features and samples from other members. In empirical evaluation on nine real-world datasets, covering diverse domains and kinds of subpopulation shift, our method of Diverse Prototypical Ensembles (DPEs) often outperforms the prior state-of-the-art in worst-group accuracy. The code is available at https://github.com/minhto2802/dpe4subpop

Figures

Figures reproduced from arXiv: 2505.23027 by the authors.

**Figure 1.** High-level overview of our method. (1) Binary classification with implicit (unannotated) subgroups. We aim to natively detect and correct for subpopulation shifts without prior subgroup knowledge. (2) Given a frozen feature extractor, f(·), we train (3) an ensemble of N prototype classifiers for each of the K classes to identify distinct sub-groups. These classifiers are trained using LIPS (Equation 4) to maximize p… view at source ↗

**Figure 2.** Motivation of DPE. (a) The synthetic training data consists of two classes, with major subgroups containing Attribute 1 and minority subgroups containing Attributes 2 and 3. Training a single model on the entire dataset leads to suboptimal decision boundaries, focusing primarily on the major subgroups; (b, c, d) as the number of models in the prototypical ensemble increases, where each member is trained to classify … view at source ↗

**Figure 3.** Worst-group improvement over ERM* when using DPE with and without subgroup annotations [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Linear ensemble versus prototypical ensemble with known and unknown attributes. NICO++ is not included in the plot since in our experiments, the worst-group accuracy of the linear ensemble on this dataset is zero. improvement as ∆N = WGAN −WGA1 WGA1 × 100, where WGAN d…

**Figure 2.** Figure 2: Class/Attribute distribution in CELEBA dataset. CHEXPERT (Irvin et al., 2019) is a large-scale medical dataset consisting of chest radiographs. The dataset presents a significant class imbalance problem, with certain rare medical conditions being underrepresented in th…

**Figure 1.** Figure 1: Class/Attribute distribution in WATERBIRDS dataset. CELEBA (Liu et al., 2015) is a large-scale facial attribute dataset, widely used to explore the impact of attribute imbalance. The target task is to classify images based on hair color (blond vs. non-blond) with the …

**Figure 3.** Figure 3: Class/Attribute distribution in CHEXPERT dataset. a) Each color represents one subgroup; b) Example images from the corresponding subgroups [PITH_FULL_IMAGE:figures/full_fig_p014_3.png]

**Figure 6.** Figure 6: Attribute distribution in NICO++ dataset. a) Each color represents one subgroup; b) Example images from the corresponding subgroups. BALANCED ACCURACY (BA) Balanced accuracy is designed to mitigate the effects of class imbalance by averaging the accuracy across all cl…

**Figure 7.** Figure 7: extends the results shown in [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

**Figure 8.** Figure 8: presents pairwise cosine similarity matrices for the first five prototypes within the ensemble, visualized separately for the “Landbird” and “Waterbird” classes. These matrices quantify the similarity between prototype embeddings, where values closer to 0 indicate gr…

**Figure 9.** Figure 9: Waterbird Prototypes. Each row depicts the top-10 validation samples closest to one of the prototypes learned for the WATERBIRDS class. Using ChatGPT for cluster interpretation, we observe that the prototypes induce structured prototype-subgroup alignment that meaningf…

**Figure 10.** Figure 10: Landbird Prototypes. Each row shows the top-10 validation samples closest to one of the prototypes from our Diversified Prototypical Ensemble (DPE) model on the LANDBIRDS class of the WATERBIRDS dataset. ChatGPT was used to analyze each prototype’s semantic structure.…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Beyond Metadata: CAPRA for Hidden Subgroup Analysis under Missing Metadata in Medical Imaging
eess.IV 2026-07 conditional novelty 6.0 of 10

CAPRA calibrates image-derived semantic proxy axes on a small labeled split into a reusable subgroup interface for failure auditing and domain-dependent robust transfer without deployment metadata.

Reference graph

Works this paper leans on

15 extracted references · 11 canonical work pages · cited by 1 Pith paper

[1]

We observe that inference time increases only slightly as the number of prototypes increases, from 0.0031s with 15 prototypes to 0.0045s with 100 prototypes

Table 12 shows the runtime and memory usage as a function of the number of prototypes per class. We observe that inference time increases only slightly as the number of prototypes increases, from 0.0031s with 15 prototypes to 0.0045s with 100 prototypes. Memory usage grows more noticeably, increasing from 0.2032 GB to 0.8517 GB. In comparison, a standard ...

work page
[3]

Deep en- sembles: A loss landscape perspective

Fort, S., Hu, H., and Lakshminarayanan, B. Deep en- sembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757,

arXiv 1912
[5]

Kirichenko, P., Izmailov, P., and Wilson, A. G. Last layer re-training is sufficient for robustness to spurious correla- tions. In ICML 2022: Workshop on Spurious Correlations, Invariance and Stability,

work page 2022
[9]

M., and McLaughlin, N

Mitchell, J., del Rinc ´on, J. M., and McLaughlin, N. Un- learning from experience to avoid spurious correlations. arXiv preprint arXiv:2409.02792,

arXiv
[10]

Pagliardini, M., Jaggi, M., Fleuret, F., and Karimireddy, S. P. Agree to disagree: Diversity through disagreement for better transferability. In ICLR 2022-International Conference on Learning Representations,

work page 2022
[11]

The caltech-ucsd birds-200-2011 dataset

Wah, C., Branson, S., Welinder, P., Perona, P., and Belongie, S. The caltech-ucsd birds-200-2011 dataset

work page 2011
[12]

Time and memory benchmarking on RTX6000 with k=1000 classes (ResNet-50, batch size = 1). Model Head # Prototypes Time per Batch (s) GPU Memory (GB) DPE 15 0.0031 0.2032 DPE 20 0.0033 0.2413 DPE 30 0.0033 0.3176 DPE 100 0.0045 0.8517 Linear (Baseline) N/A 0.0032 0.1040 C.4. Different Types of Ensemble-based Techniques In this ablation study, we compare the...

work page
[13]

Wotboost: Weighted oversampling technique in boosting for imbal- anced learning

Zhang, W., Ramezani, R., and Naeim, A. Wotboost: Weighted oversampling technique in boosting for imbal- anced learning. In 2019 IEEE International Conference on Big Data (Big data), pp. 2523–2531. IEEE,

work page 2019

Show all 15 references

[2015]

and Ludermir, T

Macˆedo, D. and Ludermir, T. Enhanced isotropy maxi- mization loss: Seamless and high-performance out-of- distribution detection simply replacing the softmax loss. arXiv preprint arXiv:2105.14399,

arXiv
[2018]

Nuanced metrics for measuring unintended bias with real data for text classification

Borkan, D., Dixon, L., Sorensen, J., Thain, N., and Vasser- man, L. Nuanced metrics for measuring unintended bias with real data for text classification. In Companion Pro- ceedings of the 2019 World Wide Web Conference , pp. 491–500,

2019
[2019]

V ., Lazarevic, A., Hall, L

Chawla, N. V ., Lazarevic, A., Hall, L. O., and Bowyer, K. W. Smoteboost: Improving prediction of the minority class in boosting. In Knowledge Discovery in Databases: PKDD 2003: 7th European Conference on Principles and Practice of Knowledge Discovery in Databases, Cavtat- Dub...

2003
[2020]

Ginsberg, T., Liang, Z., and Krishnan, R. G. A learning based hypothesis test for harmful covariate shift. arXiv preprint arXiv:2212.02742,

arXiv
[2021]

Learning compact and effective distance metrics with diversity regularization

Xie, P. Learning compact and effective distance metrics with diversity regularization. In Machine Learning and Knowledge Discovery in Databases: European Confer- ence, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15, pp. 610–624. Springer,

2015
[2022]

and Zou, J

Liang, W. and Zou, J. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. arXiv preprint arXiv:2202.06523,

arXiv
[2024]

Diversify and disambiguate: Learning from underspecified data

Lee, Y ., Yao, H., and Finn, C. Diversify and disambiguate: Learning from underspecified data. In ICML 2022: Work- shop on Spurious Correlations, Invariance and Stability,

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

We observe that inference time increases only slightly as the number of prototypes increases, from 0.0031s with 15 prototypes to 0.0045s with 100 prototypes

Table 12 shows the runtime and memory usage as a function of the number of prototypes per class. We observe that inference time increases only slightly as the number of prototypes increases, from 0.0031s with 15 prototypes to 0.0045s with 100 prototypes. Memory usage grows more noticeably, increasing from 0.2032 GB to 0.8517 GB. In comparison, a standard ...

work page

[2] [3]

Deep en- sembles: A loss landscape perspective

Fort, S., Hu, H., and Lakshminarayanan, B. Deep en- sembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757,

arXiv 1912

[3] [5]

Kirichenko, P., Izmailov, P., and Wilson, A. G. Last layer re-training is sufficient for robustness to spurious correla- tions. In ICML 2022: Workshop on Spurious Correlations, Invariance and Stability,

work page 2022

[4] [9]

M., and McLaughlin, N

Mitchell, J., del Rinc ´on, J. M., and McLaughlin, N. Un- learning from experience to avoid spurious correlations. arXiv preprint arXiv:2409.02792,

arXiv

[5] [10]

Pagliardini, M., Jaggi, M., Fleuret, F., and Karimireddy, S. P. Agree to disagree: Diversity through disagreement for better transferability. In ICLR 2022-International Conference on Learning Representations,

work page 2022

[6] [11]

The caltech-ucsd birds-200-2011 dataset

Wah, C., Branson, S., Welinder, P., Perona, P., and Belongie, S. The caltech-ucsd birds-200-2011 dataset

work page 2011

[7] [12]

Time and memory benchmarking on RTX6000 with k=1000 classes (ResNet-50, batch size = 1). Model Head # Prototypes Time per Batch (s) GPU Memory (GB) DPE 15 0.0031 0.2032 DPE 20 0.0033 0.2413 DPE 30 0.0033 0.3176 DPE 100 0.0045 0.8517 Linear (Baseline) N/A 0.0032 0.1040 C.4. Different Types of Ensemble-based Techniques In this ablation study, we compare the...

work page

[8] [13]

Wotboost: Weighted oversampling technique in boosting for imbal- anced learning

Zhang, W., Ramezani, R., and Naeim, A. Wotboost: Weighted oversampling technique in boosting for imbal- anced learning. In 2019 IEEE International Conference on Big Data (Big data), pp. 2523–2531. IEEE,

work page 2019