pith. sign in

arxiv: 2511.11934 · v3 · pith:SN7OK656new · submitted 2025-11-14 · 💻 cs.LG · cs.CV

A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts

Claudio C\'esar Claros Olivares , Austin J. Brockmeier This is my paper

Pith reviewed 2026-05-21 18:00 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords out-of-distribution detectionneural collapserepresentation learningCNNvision transformerbenchmarkingmisclassification detectionscore ranking
0
0 comments X

The pith

The competitive family of out-of-distribution detectors depends more on the learned representation than on the choice of scoring method.

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

This paper benchmarks out-of-distribution detection scores across CNN and ViT backbones trained on multiple image datasets including CIFAR-10, CIFAR-100, SuperCIFAR-100 and TinyImageNet. OOD test sets are grouped into near, mid and far regimes via CLIP semantic distances, and a statistical ranking pipeline with AURC and AUGRC metrics identifies top-performing cliques under different conditions. The central finding is that which score families win shifts primarily with properties of the learned representation rather than with score design. Simple probabilistic scores work best for misclassification detection across both architectures. On CNNs margin-based scores lead in near-OOD settings while geometry-aware scores improve as shifts grow more severe, and on fine-tuned ViTs reconstruction and residual scores dominate the top cliques. Neural collapse metrics on the last-layer features explain these patterns, and the authors add a PCA projection filter plus an NC-based method to shortlist good detectors without extra OOD data.

Core claim

The competitive detector family depends more on the learned representation than on score design alone. For both CNNs and ViTs, simple probabilistic scores dominate misclassification detection. On CNNs, margin-based scores are strongest in near-OOD regimes, while geometry-aware scores such as NNGuide, fDBD, and CTM become more competitive as shift severity increases. On fine-tuned ViTs, the top cliques are led mainly by reconstruction- and residual-based scores. These ranking shifts align with neural collapse metrics computed from the last-layer representation, and the authors propose a PCA-based projection-filtering procedure plus an NC-measurement approach that predicts a competitive short-

What carries the argument

Neural collapse metrics on last-layer representations that quantify prototype alignment with classifier weights and feature collapse, used to interpret and predict which detector families are competitive under different representation regimes.

If this is right

  • Simple probabilistic scores remain reliable for misclassification detection across CNNs and ViTs.
  • Margin-based scores are the strongest choice for near-OOD detection on CNNs.
  • Geometry-aware scores gain competitiveness as distribution shift severity increases on CNNs.
  • Reconstruction- and residual-based scores lead on fine-tuned vision transformers.
  • Neural collapse measurements from a trained classifier can shortlist competitive detectors without any OOD data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Users could compute neural collapse metrics on their own trained model to select a detector family before seeing any out-of-distribution examples.
  • The PCA projection filtering step may be worth testing as a lightweight way to improve other representation-based downstream tasks.
  • Training methods that increase neural collapse could indirectly strengthen out-of-distribution detection performance.

Load-bearing premise

That CLIP-derived semantic distances create meaningful and stable near/mid/far groupings that track genuine differences in detection difficulty, and that the multiple-comparison-controlled rank pipeline with AURC/AUGRC metrics avoids selection bias when naming top cliques.

What would settle it

An experiment on new architectures or training regimes in which detector-family rankings fail to track neural collapse measurements or in which the proposed PCA projection filter produces no consistent improvement.

Figures

Figures reproduced from arXiv: 2511.11934 by Austin J. Brockmeier, Claudio C\'esar Claros Olivares.

Figure 1
Figure 1. Figure 1: Top-clique map for AURC/AUGRC metrics: rows are CSF; columns are evaluation regimes labeled “source→test, near, mid, far”. Within each column, connected dots indicate the Conover–Holm top clique (α=0.05). Larger cliques imply more methods are statistically tied. Shaded bands emphasize methods that repeatedly appear in top cliques across regimes. (Left) For VGG-13, probabilistic-derived CSF dominate the ID … view at source ↗
Figure 2
Figure 2. Figure 2: Conover-Holm p-values [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
read the original abstract

We present a systematic benchmark of out-of-distribution (OOD) detection CSFs through a representation-centric lens. Our study spans CNN and ViT backbones, multiple training paradigms, four image-classification source datasets (CIFAR-10, CIFAR-100, SuperCIFAR-100, and TinyImageNet), and OOD datasets grouped into near, mid, and far regimes using CLIP-derived semantic distances. To compare CSFs across these settings, we employ a multiple-comparison-controlled rank pipeline that identifies top cliques of statistically indistinguishable winners under threshold-free ranking metrics (AURC and AUGRC). The main empirical finding is that the competitive detector family depends more on the learned representation than on score design alone. For both CNNs and ViTs, simple probabilistic scores dominate misclassification detection. On CNNs, margin-based scores are strongest in near-OOD regimes, while geometry-aware scores such as NNGuide, fDBD, and CTM become more competitive as shift severity increases. On fine-tuned ViTs, the top cliques are led mainly by reconstruction- and residual-based scores. To interpret these ranking shifts, we analyze the last-layer representation using Neural Collapse (NC) metrics. The resulting picture is consistent across architectures: prototype- and boundary-aware scores become stronger when the representation is more collapsed and better aligned with classifier weights, whereas weaker-collapse regimes favor gradient- and manifold-based scores. Building on these insights, we propose two contributions: a simple PCA-based projection-filtering procedure that improves detector performance, and an approach that uses NC measurements computed from a trained classifier to predict its competitive out-of-distribution detector shortlist, without requiring any additional OOD data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a systematic empirical benchmark of out-of-distribution (OOD) detection scoring functions (CSFs) across CNN and Vision Transformer (ViT) backbones under various training paradigms and source datasets (CIFAR-10, CIFAR-100, SuperCIFAR-100, TinyImageNet). OOD datasets are grouped into near, mid, and far regimes using CLIP-derived semantic distances. A multiple-comparison-controlled rank pipeline is used to identify top cliques of detectors based on AURC and AUGRC metrics. The key finding is that the competitive detector families depend more on the learned representation than on the specific score design, with patterns such as probabilistic scores dominating misclassification detection, margin-based scores performing well in near-OOD for CNNs, and reconstruction/residual scores leading for fine-tuned ViTs. Neural Collapse (NC) metrics are employed to interpret these shifts, leading to proposals for a PCA-based projection-filtering procedure and an NC-based predictor for detector shortlists without additional OOD data.

Significance. If the results are robust, this study offers important insights into how representation properties influence the effectiveness of different OOD detection approaches, providing a more nuanced understanding beyond isolated comparisons. The incorporation of Neural Collapse analysis to explain performance variations and the development of practical tools like the projection filter and NC predictor represent constructive contributions. The use of statistical controls in ranking adds rigor to the benchmarking process.

major comments (2)
  1. [OOD regime partitioning (experimental setup or methods section describing dataset grouping)] The definition of near/mid/far OOD regimes is based exclusively on CLIP-derived semantic distances between source and OOD classes. However, there is no reported validation demonstrating that these distances correlate with actual detection difficulty, for instance by showing that baseline scores like MSP exhibit statistically different AURC or AUROC across the regimes independent of the evaluated CSFs. This is load-bearing for the central claim, as the observed ranking transitions (e.g., margin-based scores strongest in near-OOD on CNNs, geometry-aware scores more competitive with increasing shift severity) are interpreted as effects of representation and shift severity; without this correlation, the regime-specific findings risk being artifacts of the grouping method rather than genuine differences in detection challenges.
  2. [Neural Collapse analysis and interpretation of ranking shifts] In the section analyzing ranking shifts via Neural Collapse metrics, the link between representation collapse/alignment and score family performance (prototype-aware scores stronger under high collapse) is presented as consistent across architectures, but lacks explicit quantitative support such as correlation values or predictive regressions between NC quantities and clique membership or rank positions of score families.
minor comments (2)
  1. [Methods or experimental pipeline description] Provide more granular details on the exact statistical procedure for the multiple-comparison-controlled rank pipeline, including the test used for clique identification and the correction method, to support full reproducibility.
  2. [Results tables/figures on top cliques] In tables or figures showing top cliques per regime/architecture, include the number of underlying datasets or independent runs to allow readers to gauge the stability of the reported rankings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive review. The comments highlight important aspects of our experimental design and analysis that warrant clarification and strengthening. We address each major comment point by point below and outline the revisions we will make.

read point-by-point responses

  1. Referee: [OOD regime partitioning (experimental setup or methods section describing dataset grouping)] The definition of near/mid/far OOD regimes is based exclusively on CLIP-derived semantic distances between source and OOD classes. However, there is no reported validation demonstrating that these distances correlate with actual detection difficulty, for instance by showing that baseline scores like MSP exhibit statistically different AURC or AUROC across the regimes independent of the evaluated CSFs. This is load-bearing for the central claim, as the observed ranking transitions (e.g., margin-based scores strongest in near-OOD on CNNs, geometry-aware scores more competitive with increasing shift severity) are interpreted as effects of representation and shift severity; without this correlation, the regime-specific findings risk being artifacts of the grouping method rather than genuine.

    Authors: We agree that explicit validation of the regime partitioning would strengthen the manuscript and better support the interpretation of representation-dependent ranking shifts. While CLIP semantic distances follow established practices for defining semantic similarity in OOD literature, we did not include a direct check (e.g., MSP AURC differences across regimes with statistical tests). In the revision we will add this validation in the methods or results section, reporting AURC/AUROC for MSP (and optionally one or two other baselines) across near/mid/far regimes with appropriate multiple-comparison corrections, to confirm that the grouping aligns with measurable differences in detection difficulty. revision: yes

  2. Referee: [Neural Collapse analysis and interpretation of ranking shifts] In the section analyzing ranking shifts via Neural Collapse metrics, the link between representation collapse/alignment and score family performance (prototype-aware scores stronger under high collapse) is presented as consistent across architectures, but lacks explicit quantitative support such as correlation values or predictive regressions between NC quantities and clique membership or rank positions of score families.

    Authors: We appreciate this observation. The current analysis relies on qualitative consistency of patterns across architectures and settings, but we acknowledge that adding quantitative measures would make the link more rigorous. In the revised manuscript we will compute and report Pearson (or Spearman) correlations between key NC metrics (collapse, alignment, and variability) and the rank positions or clique membership indicators of score families. We will also include a brief regression analysis predicting score-family performance from NC quantities, with results presented in the Neural Collapse section or an appendix. revision: yes

Circularity Check

0 steps flagged

Empirical benchmarking with post-hoc NC analysis; no derivation reduces to fitted quantity by construction

full rationale

The paper is a systematic empirical benchmark of OOD detection scores across CNN/ViT architectures, training paradigms, and source datasets, with OOD regimes grouped by CLIP semantic distances and rankings obtained via multiple-comparison-controlled AURC/AUGRC pipelines. The central claim that competitive detector families depend more on learned representations than score design is supported by observed ranking shifts and post-hoc Neural Collapse metric analysis of last-layer representations. The proposed NC-based predictor for detector shortlists is explicitly described as an empirical observation and approach rather than a closed-form derivation or statistical fit that reuses the same data as a 'prediction.' No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the provided derivation chain; the work remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The study rests on the assumption that semantic distances from CLIP provide a valid proxy for OOD shift severity and that standard NC metrics capture the relevant representation properties. No free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption CLIP-derived semantic distances accurately group OOD datasets into near, mid, and far regimes that reflect meaningful detection difficulty differences
    Used to stratify results and interpret ranking shifts across shift severity.

pith-pipeline@v0.9.0 · 5848 in / 1310 out tokens · 45466 ms · 2026-05-21T18:00:40.625829+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 3 internal anchors

  1. [1]

    B., Belkhir, N., Popescu, S., Manzanera, A., and Franchi, G

    Ammar, M. B., Belkhir, N., Popescu, S., Manzanera, A., and Franchi, G. Neco: Neural collapse based out-of- distribution detection.ArXiv Preprint ArXiv:2310.06823,

    work page arXiv

  2. [2]

    Learning Confidence for Out-of-Distribution Detection in Neural Networks

    DeVries, T. and Taylor, G. W. Learning confidence for out-of-distribution detection in neural networks.ArXiv Preprint ArXiv:1802.04865,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Kernel PCA for out- of-distribution detection: Non-linear kernel selections and approximations.ArXiv Preprint ArXiv:2505.15284,

    Fang, K., Tao, Q., He, M., Lv, K., Yang, R., Hu, H., Huang, X., Yang, J., and Cao, L. Kernel PCA for out- of-distribution detection: Non-linear kernel selections and approximations.ArXiv Preprint ArXiv:2505.15284,

    work page arXiv

  4. [4]

    Bias-Reduced Uncertainty Estimation for Deep Neural Classifiers

    Geifman, Y ., Uziel, G., and El-Yaniv, R. Bias-reduced uncertainty estimation for deep neural classifiers.ArXiv Preprint ArXiv:1805.08206,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    W., and Palm, C

    Gutbrod, M., Rauber, D., Nunes, D. W., and Palm, C. Open- MIBOOD: Open medical imaging benchmarks for out-of- distribution detection.ArXiv Preprint ArXiv:2503.16247,

    work page arXiv

  6. [6]

    A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks

    Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks.ArXiv Preprint ArXiv:1610.02136,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Scaling out- of-distribution detection for real-world settings.ArXiv Preprint ArXiv:1911.11132,

    9 A Systematic Analysis of Out-of-Distribution Detection Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., and Song, D. Scaling out- of-distribution detection for real-world settings.ArXiv Preprint ArXiv:1911.11132,

    work page arXiv 1911

  8. [8]

    F., L¨uth, C

    Jaeger, P. F., L¨uth, C. T., Klein, L., and Bungert, T. J. A call to reflect on evaluation practices for failure detection in image classification.ArXiv Preprint ArXiv:2211.15259,

    work page arXiv

  9. [9]

    and Qin, Y

    Liu, L. and Qin, Y . Fast decision boundary based out-of- distribution detector.ArXiv Preprint ArXiv:2312.11536,

    work page arXiv

  10. [10]

    Massey, J. L. Guessing and entropy. InProceedings Of 1994 IEEE International Symposium On Information Theory, pp

    work page 1994

  11. [11]

    D., and Thanh-Tung, H

    Ngoc-Hieu, N., Hung-Quang, N., Ta, T.-A., Nguyen-Tang, T., Doan, K. D., and Thanh-Tung, H. A cosine similarity- based method for out-of-distribution detection.ArXiv Preprint ArXiv:2306.14920,

    work page arXiv

  12. [12]

    Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein

    Pope, P., Zhu, C., Abdelkader, A., Goldblum, M., and Gold- stein, T. The intrinsic dimension of images and its impact on learning.ArXiv Preprint ArXiv:2104.08894,

    work page arXiv

  13. [13]

    J., L ¨uth, C

    Traub, J., Bungert, T. J., L ¨uth, C. T., Baumgartner, M., Maier-Hein, K. H., Maier-Hein, L., and Jaeger, P. F. Over- coming common flaws in the evaluation of selective clas- sification systems.ArXiv Preprint ArXiv:2407.01032,

    work page arXiv

  14. [14]

    OpenOOD v1.5: Enhanced Benchmark for Out -of- Distribution Detection,

    Zhang, J., Yang, J., Wang, P., Wang, H., Lin, Y ., Zhang, H., Sun, Y ., Du, X., Li, Y ., Liu, Z., et al. Openood v1.5: Enhanced benchmark for out-of-distribution detection. ArXiv Preprint ArXiv:2306.09301,

    work page arXiv

  15. [15]

    Training Paradigms, CFS Baselines and Variations A.1

    10 A Systematic Analysis of Out-of-Distribution Detection A. Training Paradigms, CFS Baselines and Variations A.1. Computing Infrastructure All experiments were executed on an internal GPU cluster.CNNruns (VGG-13 trained from scratch) were scheduled on NVIDIA T4 GPUs, whileViTruns (fine-tuned from a large pretrained model) were scheduled on NVIDIA A100 GP...

    work page 2019

  16. [16]

    This leads to the loss LDG(W;D train, o) :=− 1 |Dtrain| P (xi,yi)∈Dtrain log o pyi(xi) +p K+1(xi) . When pK+1 =0 (no abstention), LDG reduces to cross-entropy up to an additive constant (since logo adds to the true-class 11 A Systematic Analysis of Out-of-Distribution Detection logit). The head is linear, g(z) =W z+b, W∈R (K+1)×D , b∈R K+1, so the method ...

    work page 2019

  17. [17]

    and Deep Gamblers (Liu et al., 2019), ConfidNet (Corbi`ere et al., 2019; Corbiere et al.,

    work page 2019

  18. [18]

    Prototype matching in feature space consists of quantifying the similarity between a sample x and the last-layer trained weights {w1, . . . ,wK}. Therefore the similarity to the closest trained weight is CTM(x) = max k≤C sim wk,h . Alternatively, we can compute class means µc train and score by similarity to the closest class mean, CTMmean(x) = maxk≤C sim...

    work page 2023

  19. [19]

    Higher Energy score typically indicates higher uncertainty

    The energy score is defined as Energy(x) =−Tlog PC k=1 exp g(h)k/T , with temperature T >0 . Higher Energy score typically indicates higher uncertainty. A.3.3. MAXIMUMSOFTMAXRESPONSE(MSR) (HENDRYCKS& GIMPEL, 2016)ANDMAXIMUMLOGITSCORE (MLS) (HENDRYCKS ET AL.,

    work page 2016

  20. [20]

    Lower values indicate atypical inputs

    A baseline confidence score given by the maximum predicted probabilityMSR(x) = maxk≤C pk, widely used for OOD detection. Lower values indicate atypical inputs. Similarly, MLS is a confidence score measured in the logit space, MLS(x) = maxk≤C g(h)k,often more stable than softmax under temperature changes. A.3.4. PREDICTIVEENTROPY(PE), GENERALIZEDENTROPY(GE...

    work page 2019

  21. [21]

    Generalized Entropy (GEN) (Liu et al., 2023).GEN is a post-hoc OOD score that uses the softmax probabilities of a trained classifier

    of the predictive distributionPE(x) =H p(x) =− PC k=1 pk logp k,with larger entropy signaling higher uncertainty. Generalized Entropy (GEN) (Liu et al., 2023).GEN is a post-hoc OOD score that uses the softmax probabilities of a trained classifier. Let p(1) ≥ · · · ≥p (K) denote the probabilities sorted in descending order for a given input x. For sensitiv...

    work page 2023

  22. [22]

    collision probability,

    quantifies the expected number of guesses to identify the true class when labels are guessed in decreasing probability pk(x): if p(1) ≥ · · · ≥p (K) are sorted, then GE(x) = PC k=1 kp (k), with larger values denoting higher uncertainty. Predictive Collision Entropy (PCE) (Granese et al., 2021).PCE measures prediction uncertainty via thecollision (order-2 ...

    work page 2021

  23. [23]

    NeCo’s new observation eatblishes ID/OOD orthogonality, which implies that OOD features concentrate near the origin after projection onto the ID subspace

    This method is motivated by the Neural Collapse phenomena (Papyan et al., 2020), which unveils geometric properties that manifest at the end of the training process. NeCo’s new observation eatblishes ID/OOD orthogonality, which implies that OOD features concentrate near the origin after projection onto the ID subspace. This method fits PCA on ID features ...

    work page 2020

  24. [24]

    Both quantities are evaluated on CLIP embeddings;smallervalues indicate that DOOD is closer to the ID manifold

    with a polynomial kernel k(u,v) = (u ⊤v+c) d: \MMD 2 = 1 n(n−1) P i̸=i′ k(z i,z i′) + 1 m(m−1) P j̸=j ′ k(z ′ j,z ′ j′)− 2 nm P i,j k(z i,z ′ j). Both quantities are evaluated on CLIP embeddings;smallervalues indicate that DOOD is closer to the ID manifold. Class-aware distances.For ID class c∈ {1, . . . , K} , define the (normalized) image- prototype µc ...

    work page 1957

  25. [25]

    For finite samples, the Iman–Davenport F -approximation is recommended (Iman & Davenport, 1980): FF = (N−1)Q N(k−1)−Q ∼F k−1,(k−1)(N−1)

    kX j=1 ¯R 2 j −3N(k+ 1), (optionally applying a standard tie correction within blocks). For finite samples, the Iman–Davenport F -approximation is recommended (Iman & Davenport, 1980): FF = (N−1)Q N(k−1)−Q ∼F k−1,(k−1)(N−1) . IfF F exceeds the critical value at levelα, we rejectH 0 and proceed with post-hoc pairwise comparisons. Conover post-hoc & Bron–Ke...

    work page 1980

  26. [26]

    top groups,

    6N , T ij = | ¯Ri − ¯Rj| SE , two-sided p-values are obtained from the normal (or t) reference, and multiplicity is controlled across all k 2 pairs using Holm’s step-down procedure (Holm, 1979). To summarize statistically indistinguishable winners, construct anindifference graph G= (V, E) with nodes V={1, . . . , k} (methods) and edges (i, j)∈E iff the ad...

    work page arXiv 1979

  27. [27]

    For Figure 1, we only report the first layer for all the possible scenarios. CTM Confidence Energy GEN MSR fDBD CTM Confidence Energy GEN MSR fDBD 1.000 0.000 0.001 0.000 0.052 0.133 0.000 1.000 0.000 0.228 0.052 0.000 0.001 0.000 1.000 0.000 0.000 0.037 0.000 0.228 0.000 1.000 0.002 0.000 0.052 0.052 0.000 0.002 1.000 0.001 0.133 0.000 0.037 0.000 0.001 ...

    work page 2020

  28. [28]

    This implies that the collapsed ID feature space is maximally sparse in terms of angular distribution

    In the other hand, Maximal Angular Margin dictates that class means form a Simplex ETF, maximizing the separation angle θij between any distinct classes i, j: cos(µi,µ j) =− 1 K−1 ∀i̸=j . This implies that the collapsed ID feature space is maximally sparse in terms of angular distribution. For an OOD sample xOOD lying in the subspace orthogonal to the ID ...

    work page 1950

  29. [29]

    Equiangularity dictates that any pair of class means are equally spacedcosu(i, j) =β,∀i̸=j , meaning that for an off-target logitg(h) j =w ⊤ j h+b k ≈w ⊤ j µk =α∥µ k∥2β=αR 2β

    This uniformity prevents class-conditional bias, where some ID classes might otherwise have naturally higher energy (and thus higher False Positive Rates) than others due to varying feature norms. Equiangularity dictates that any pair of class means are equally spacedcosu(i, j) =β,∀i̸=j , meaning that for an off-target logitg(h) j =w ⊤ j h+b k ≈w ⊤ j µk =...

    work page 2025