Pith. sign in

REVIEW 3 major objections 4 minor 51 references

TOOD: Task-Aware Out-of-Distribution Score Calibration for Continual Learners

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Continual learners lose out-of-distribution detection through score miscalibration, not forgotten knowledge — and a training-free per-task re-calibration of energy scores recovers much of the loss.

desk verdict A genuinely useful post-hoc fix for energy-based OOD detection in continual learning, with a mostly honest mechanism story—but the headline dissociation claim is overstated and the joint-training control is missing. read the letter →

arxiv 2607.29592 v1 pith:XOSFMQFC submitted 2026-07-31 cs.CV cs.LG

Mostafa ElAraby , Samer B. Nashed , Liam Paull This is my paper
classification cs.CVcs.LG
keywords continuallearningout-of-distributiondetectionOODforgettingconfidencegapmanifoldcrowdingenergy-basedscorecalibrationclass-incremental
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

The paper tries to establish that a continual learning system's ability to reject out-of-distribution (OOD) inputs erodes through a failure largely distinct from classic catastrophic forgetting: score miscalibration. It identifies two mechanisms — the confidence gap, in which old-task logits shrink in absolute scale relative to newer-task logits as the output head expands, and manifold crowding, in which newly learned classes occupy feature space that previously served as OOD margin. It then shows that a training-free, post-hoc re-calibration of per-task energy scores, using per-task median and spread statistics from a small replay buffer, recovers a substantial portion of the OOD detection loss for energy-based detectors without retraining. If true, OOD robustness in class-incremental learning is substantially a calibration problem — a cheaper problem to fix than an irrecoverable loss of discriminative structure. The finding matters because it offers a practical, backward-compatible repair for a safety-critical failure mode of lifelong learning systems.

What carries the argument

The central object is the per-task energy decomposition with a robust location–scale re-anchoring inside a max operation. The logit vector is split by the training-time task-to-class partition into N task energies E_t(x) = log Σ_{c∈C_t} exp(h_c(x)); each is re-centered by subtracting its calibration-set median and re-scaled by its median absolute deviation (MAD), anchored at a reference task's statistics, and the final score is S(x) = max_t E_t^norm(x), with an optional margin bonus for a dominant top channel. Because ID samples activate one channel strongly while OOD samples activate none, the normalization inside the max moves ID and OOD scores in different directions — a transformation a

What would settle it

Take a near-OOD set that is semantically adjacent to one learned task and compute each sample's per-task energy E_t(x); if any such sample activates one task channel as strongly as in-distribution samples do, then TOOD's max-over-normalized-channels will score it as in-distribution — measuring whether its TOOD score distribution overlaps the ID distribution would directly test this premise. If the overlap is substantial while uncalibrated energy still separates the two, TOOD's reported gains should invert.

Watch

Extended reading notes

Core claim

Across several continual-learning methods, feature drift correlates strongly with accuracy forgetting but not with OOD deterioration, indicating OOD forgetting is a distinct phenomenon. Energy-based detectors fail through the confidence gap: as the classifier head expands, old-task logits shrink in absolute scale, so ID scores drift toward OOD scores even when old-task accuracy is preserved. Feature-based detectors fail through manifold crowding: new-task embeddings occupy latent regions that once served as OOD margin. TOOD splits the logit vector into per-task energy scores, normalizes each task's energy by its own median and spread statistics from a replay buffer, and scores by the maximum

Load-bearing premise

The method rests on the empirical assumption that an in-distribution sample strongly activates exactly one task's energy channel while an out-of-distribution sample activates none, and that each task's own score statistics can be estimated from a replay buffer or held-out split; if either premise fails, the per-task re-calibration can push out-of-distribution scores above the decision threshold or the method cannot be applied.

Editorial extensions

If this is right

  • OOD robustness should be measured per task across the whole learning stream, not just at the final checkpoint; current accuracy-forgetting metrics miss an independent failure mode.
  • For energy-based detectors, a substantial fraction of OOD forgetting is recoverable without retraining wherever the confidence gap dominates — a cheap, backward-compatible repair.
  • Methods whose output-head scale is already stabilized (for instance by explicit bias correction) stand to gain less, making TOOD a targeted repair rather than a universal boost.
  • Feature-based OOD detection requires complementary fixes such as margin-preserving replay or representation regularization, since re-centering scores cannot restore lost geometric margin.
  • TOOD needs only a few dozen calibration samples per task and no test-time task identity, so it applies to standard class-incremental settings with expanding heads.

Reading between the lines

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

  • A direct testable extension is to pair TOOD with a margin-preserving replay objective; the paper's own diagnosis implies the two repairs are complementary, so the combination should recover most of the remaining OOD forgetting.
  • If the confidence-gap finding generalizes, a similar per-task, non-monotonic calibration could reformulate other output-based scores (for example, softmax confidence or temperature-scaled logits) in continual streams, though the paper's rank-preservation argument warns that only per-task transforms that act differently on ID and OOD samples can help.
  • The premise that OOD samples activate no task channel strongly is only tested on benchmark near/far OOD sets; adversarially chosen or semantically overlapping OOD that excites one task channel would stress where the max-over-channels scoring inverts.
  • For systems without replay, a held-out per-task ID split should suffice for calibration; the buffer-size saturation observed (50–100 samples) suggests this is feasible even in memory-constrained deployments.
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.

Referee Report

3 major / 4 minor

Summary. The paper studies out-of-distribution (OOD) detection in class-incremental learning (CIL). It introduces OOD forgetting metrics (D_i and D_avg) and average incremental AUROC, then identifies two mechanisms for OOD deterioration: the confidence gap (old-task logits shrink relative to new-task logits) and manifold crowding (new classes occupy the latent-space margin needed by distance-based detectors). To address the confidence gap, the paper proposes TOOD, a training-free post-hoc method that decomposes the logit vector into per-task energies, normalizes each task channel using task-wise ID statistics (mean or median/MAD), and scores a test sample by the maximum over normalized channels, optionally augmented by a margin term. The method is evaluated on CIFAR-10, CIFAR-100, and a 100-task ImageNet stream with several CIL backbones and OOD detectors. The paper reports that TOOD improves AUROC relative to uncalibrated energy in most settings, with largest gains for methods that exhibit strong logit-scale drift, and concludes that a substantial portion of OOD deterioration in CIL is due to score miscalibration rather than a complete loss of discriminative structure.

Significance. The paper has clear strengths: it formulates a useful metric for tracking OOD performance over a CL trajectory, gives an honest dissection of the confidence-gap mechanism with a controlled toy experiment (Appendix C), and backs the method with several careful ablations (per-task temperature scaling in Appendix D.2, partition robustness in Appendix D.3, anchor invariance in Appendix D.5, buffer-size sensitivity in Appendix D.4). Public code is provided. If the central attribution is correct, the result is practically valuable: a post-hoc, no-retraining intervention can recover a substantial part of OOD detection performance in CIL. However, two load-bearing points currently weaken the paper. First, the correlation statistic in Figure 1 (r=0.365, p=0.270) contradicts the abstract's claim that OODF is 'weakly anti-correlated' with classification performance; the reported value is positive and non-significant. Second, the paper never runs the sharpest control for the confidence-gap mechanism: applying the same per-task normalization to a jointly trained model that has no task-age logit drift. Without that control, the reported TOOD gains could be a generic benefit of per-group score no

major comments (3)
  1. [Abstract and §3, Fig. 1] The abstract and introduction state that OOD forgetting is 'weakly anti-correlated' with classification performance on previous tasks. Figure 1 reports r=0.365, p=0.270, which is a positive, non-significant correlation. This is internally inconsistent. The 'distinct mechanisms' framing (RQ1) is therefore stated more strongly than the evidence supports. Please report the sign correctly, avoid the word 'anti-correlated', and temper the conclusion to 'no significant correlation was detected in this small sample of CL methods'.
  2. [§5.2 and Appendix D] The paper's headline claim is that TOOD's improvement shows that a substantial portion of OOD deterioration is score miscalibration rather than loss of discriminative structure. This inference requires a control with no confidence gap: a model jointly trained on all classes, scored with the same task partition and per-task normalization. If TOOD still improves AUROC on such a model, its gain is a generic per-group normalization effect and does not specifically support the confidence-gap mechanism. The paper currently lacks this experiment; it would directly test the mechanistic claim and should be added.
  3. [§4.3, Eq. (8) and Appendix D.1] The default margin weight λ=0.5 is chosen by sweeping λ on CIFAR-10 using the same CL methods and OOD sets that appear in Table 1. This is selection on the evaluation benchmark. Since all main results use this value, the reported rankings may be optimistic. The qualitative conclusion is likely robust because λ=0 is a principled base case, but the paper should report the λ=0 results in the main tables or tune λ on a separate validation split and report both values.
minor comments (4)
  1. [Abstract vs §5.2] The abstract as supplied says TOOD ranks first or second in 'nine of ten' CIFAR configurations, while the body (Section 5.2 and the introduction) says 'eight of ten'. Please reconcile.
  2. [Figure 1] With only five CL methods, p=0.270 cannot support the conclusion that OODF and classification forgetting are unrelated; it only shows that this sample lacks power. Please soften the wording or add more methods to the correlation analysis.
  3. [§4.2 vs Appendix D.2] Section 4.2 says 'the recentering, not task-awareness on its own, is what drives the improvement,' while Appendix D.2 concludes the improvement is 'attributable to the per-task energy mechanism itself rather than access to the partition.' These can be reconciled, but the wording is confusing: the temperature baseline is also task-aware and also uses per-task energies, so the missing ingredient is the reference anchoring. Please clarify.
  4. [§4.3] The heuristic margin term is described as optional, and the paper analyzes its sensitivity, but the main text does not note that the fixed λ=0.5 was selected by a sweep on the same benchmark. Please add a caveat at the point of the default choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TOOD's calibration is fit only to ID replay statistics, and the anchor-invariance proof is self-contained.

full rationale

The derivation is self-contained. TOOD's per-task statistics (µ_t, median, MAD) are estimated from ID replay/held-out data (Sec 4.2, Eqs. 5-6), never from OOD labels or from the AUROC values being reported. The final score (Eq. 7) is a max over normalized per-task energies, and the paper explicitly proves that changing the reference task only applies a common affine map to all channels, which commutes with the max and preserves AUROC ranking (Sec 4.2, Appendix D.5). The central empirical claim—that OODF is distinct from accuracy forgetting—is supported by external measurements (CKA, nearest-neighbor margin, toy experiment) rather than by definition of the metrics. The toy experiment isolates the confidence gap from manifold crowding and shows TOOD recovers only the former (+5.1 Avg AUROC) and not the latter (−0.76 Avg AUROC), a falsifiable dissociation rather than a tautology. The λ=0.5 margin value is selected by sweeping the margin parameter on the same benchmarks; this is an explicit hyperparameter choice (Appendix D.1), not a fitted parameter renamed as a prediction, so it is a correctness/reproducibility concern rather than circularity. A missing joint-trained control would strengthen the mechanistic interpretation, but its absence does not make any equation reduce to an input. No load-bearing self-citations were found; the limitations stated in the conclusion (e.g., TOOD cannot fix manifold crowding, efficacy drops when the task partition is ambiguous) are openly acknowledged rather than hidden.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical entities are postulated. The 'confidence gap' and 'manifold crowding' are descriptive names for measured phenomena, and the paper explicitly states TOOD cannot fix manifold crowding.

free parameters (1)
  • margin weight lambda = 0.5
    Selected by sweeping lambda in [0,1] on CIFAR-10 and observing AUROC (Appendix D.1); used as a fixed default in all main experiments. A sweep-based choice on the evaluation benchmark is a fitted hyperparameter.
assumptions (4)
  • domain assumption Training-time task-to-class partition is available and fixed (each class belongs to exactly one task).
    TOOD groups logits into per-task energies using this partition (Section 4.1); standard CIL head expansion provides it.
  • domain assumption Per-task ID calibration sets B_t (replay buffer or held-out split) are available after training, with correct task labels.
    TOOD estimates median/MAD per task from B_t (Section 4.2); for replay-free methods this is an extra held-out ID split.
  • domain assumption OOD samples activate no task channel strongly; ID samples activate their own channel dominantly.
    This is the mechanism that makes the max over normalized per-task energies separate ID from OOD (Section 4.3); empirically tested but not proven.
  • standard math AUROC depends only on ranking; a common affine map applied to all normalized channels preserves ranking.
    Used in Section 4.2 anchor-invariance argument; mathematically correct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TOOD: Task-Aware Out-of-Distribution Score Calibration for Continual Learners." pith.science (2026). https://pith.science/paper/XOSFMQFC

@misc{pith2026260729592,
  author       = {Pith},
  title        = {Pith review of: TOOD: Task-Aware Out-of-Distribution Score Calibration for Continual Learners},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XOSFMQFC}},
  note         = {Machine review of arXiv:2607.29592}
}
read the original abstract

The primary challenge of continual learning (CL) systems is to learn new tasks while remaining performant on previously learned tasks. A similarly important though less well-studied aspect of CL systems is their ability to distinguish inputs that are unlikely to come from within the set of tasks the system has already encountered, often called out-of-distribution (OOD) detection. This paper presents several findings related to the dynamics of OOD detection in CL systems, causes of performance degradation over time which we call OOD forgetting (OODF), and proposed mitigation strategies for this degradation. Chiefly, we find the unintuitive result that OODF is only weakly anti-correlated with classification performance on previous tasks, suggesting that the underlying mechanisms producing OODF are distinct. Moreover, this effect is observed for both energy-based and feature-based OOD detection methods. Energy-based detectors suffer a drop in logit scale as additional tasks are learned, which we term the Confidence Gap, while feature-based detectors also degrade under a complementary effect we call Manifold Crowding. Motivated by these observations, we propose TOOD, a training-free post-hoc method that decomposes logits into per-task energy scores and re-calibrates them using replay-buffer statistics. Experiments on CIFAR-10, CIFAR-100, and a 100-task ImageNet-1K stream show that TOOD improves OOD detection performance over uncalibrated energy in most settings and ranks first or second in nine of ten CIFAR configurations, with the largest gains when the confidence gap is most severe. These results suggest that a substantial portion of OOD deterioration in continual learning arises from score miscalibration rather than from a complete loss of discriminative structure.

Figures

Figures reproduced from arXiv: 2607.29592 by the authors.

Figure 1
Figure 1. Feature drift explains accuracy forgetting but not OOD forgetting. Each point is a CL method on CIFAR￾10 (N = 5 tasks). Left: Feature drift (1 − CKA) vs. OOD deterioration (Davg) shows no significant relationship; r = 0.365, p = 0.270. Right: Feature drift vs. classification forgetting (F¯); r = 0.880, p < 0.001. Features drift appears to cause classification forgetting but not OOD detection, suggesting the latter i… view at source ↗
Figure 2
Figure 2. Continual learners retain classification accuracy on old tasks while losing OOD separability. Left: Task 0 classifier acc. (blue) remains high, while uncalibrated energy-based OOD AUROC (red dash) drops by 35 points. TOOD calibration (green) recovers most of that gap. Right: The mean ID energy score (blue) drifts from its T0 value toward the OOD score (red dash), collapsing the confidence gap at T3. TOOD (green) re-… view at source ↗
Figure 3
Figure 3. TOOD restores ID/OOD separability across CL methods. Each column corresponds to a different CL method. Top row: Uncalibrated Task 0 ID energy scores (blue) converge toward OOD scores (red) as new tasks are learned, eroding the detection margin. Bottom row: After TOOD’s per-task normalization, calibrated ID scores (green) maintain a clear separation from OOD scores throughout the task stream. Boxes span the interquar… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Effect of buffer size B (CIFAR-10, N = 5 tasks). Average area under the receiver operating characteristic curve (AUROC) (↑, top) and OOD deterioration (↓, bottom) saturate quickly. The default B = 200 (dashed line) is well within the stable performance plateau, confirm…
Figure 5
Figure 5. Figure 5: Manifold Crowding: new task embeddings erode the OOD detection margin without displacing old representations (iCaRL, CIFAR-10, 5 tasks). Top: the feature space is colonized task by task (t-SNE). Bottom: the resulting margin collapse is measured by nearest-neighbor dist…
Figure 6
Figure 6. Figure 6: TOOD selectively shifts the ID score distribution rightward while OOD scores remain anchored (iCaRL, CIFAR-10, N = 5 tasks). Each ridge is the energy distribution at one task step. Left (uncalibrated): the ID distribution (blue) shifts progressively leftward with each …
Figure 7
Figure 7. Figure 7: Controlled single-seed confidence-gap toy experiment. (a) the toy geometry: class clusters on a shell around [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Sensitivity to the margin parameter λ (CIFAR-10, N = 5). Left: Mean AUROC (↑); Right: Average OOD deterioration Davg (↓). High-drift methods (iCaRL, DER) peak near λ = 0.5 (dotted line). Low-drift methods (BiC) decrease monotonically for λ > 0, suggesting that the marg…
Figure 9
Figure 9. Figure 9: Score-level dissociation. Only logit-space TOOD ( [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 2 canonical work pages

  1. [1]

    In or out? fixing imagenet out-of-distribution detection evaluation

    Julian Bitterwolf, Maximilian M \" u ller, and Matthias Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , volume 202 of ...

    2023

  2. [2]

    D ark E xperience R eplay for C ontinual L earning

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. D ark E xperience R eplay for C ontinual L earning. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pp.\ 15320--15330, 2020

    2020

  3. [3]

    Efficient lifelong learning with A-GEM

    Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with A-GEM . In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019 a . URL https://openreview.net/forum?id=Hkf2\_sC5FX

    2019

  4. [4]

    O n tiny episodic memories in continual learning

    Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. O n tiny episodic memories in continual learning. ArXiv preprint, abs/1902.10486, 2019 b . URL https://arxiv.org/abs/1902.10486

    arXiv 1902

  5. [5]

    Describing textures in the wild

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014 , pp.\ 3606--3613. IEEE Computer Society, 2014. doi:10.1109/CVPR.2014.461. URL https://doi.org/10.1109/CVPR.2014.461

    doi:10.1109/cvpr.2014.461 2014

  6. [6]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li - Jia Li, Kai Li, and Fei - Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA , pp.\ 248--255. IEEE Computer Society, 2009. doi:10.1109/CVPR.2009.5206848. URL https://doi.or...

    arXiv 2009

  7. [7]

    Extremely simple activation shaping for out-of-distribution detection

    Andrija Djurisic, Nebojsa Bozanic, Arjun Ashok, and Rosanne Liu. Extremely simple activation shaping for out-of-distribution detection. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. URL https://openreview.net/pdf?id=ndYXTEL6cZz

    2023

  8. [8]

    Toy models of superposition

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022

    arXiv 2022

Show all 51 references
  1. [9]

    An introduction to roc analysis

    Tom Fawcett. An introduction to roc analysis. Pattern recognition letters, 27 0 (8): 0 861--874, 2006

    2006

  2. [10]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017 , volume 70 o...

    2017

  3. [11]

    Out-of-distribution detection for continual learning: Design principles and benchmarking

    Srishti Gupta, Riccardo Balia, Daniele Angioni, Fabio Brau, Maura Pintor, Ambra Demontis, Alessandro Sebastian, Salvatore Mario Carta, Fabio Roli, and Battista Biggio. Out-of-distribution detection for continual learning: Design principles and benchmarking. arXiv preprint arXi...

    2025 doi

  4. [12]

    Buffer-free class-incremental learning with out-of-distribution detection

    Srishti Gupta, Daniele Angioni, Maura Pintor, Ambra Demontis, Lea Sch \"o nherr, Fabio Roli, and Battista Biggio. Buffer-free class-incremental learning with out-of-distribution detection. Pattern Recognition, 172: 0 112441, 2026. doi:10.1016/j.patcog.2025.112441. URL https://...

    2026

  5. [13]

    Controlling neural collapse enhances out-of-distribution detection and transfer learning

    Md Yousuf Harun, Jhair Gallardo, and Christopher Kanan. Controlling neural collapse enhances out-of-distribution detection and transfer learning. arXiv preprint arXiv:2502.10691, 2025. doi:10.48550/arXiv.2502.10691. URL https://arxiv.org/abs/2502.10691

    2025 doi

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016 , pp.\ 770--778. IEEE Computer Society, 2016. doi:10.1109/CV...

    2016 doi

  7. [15]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net,...

    2017

  8. [16]

    MOS: towards scaling out-of-distribution detection for large semantic space

    Rui Huang and Yixuan Li. MOS: towards scaling out-of-distribution detection for large semantic space. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021 , pp.\ 8710--8719. Computer Vision Foundation / IEEE , 2021. doi:10.1109/CV...

    2021

  9. [17]

    R obust statistics

    Peter J Huber. R obust statistics. In International encyclopedia of statistical science, pp.\ 1248--1251. Springer, 2011

    2011

  10. [18]

    Continual learning based on OOD detection and task masking

    Gyuhak Kim, Sepideh Esmaeilpour, Changnan Xiao, and Bing Liu. Continual learning based on OOD detection and task masking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2022, New Orleans, LA, USA, June 19-20, 2022 , pp.\ 3855--3865....

    2022

  11. [19]

    Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A

    James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska - Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. O vercoming catastrophic forget...

    2016 arXiv

  12. [20]

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey E. Hinton. Similarity of neural network representations revisited. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 201...

    2019

  13. [21]

    L earning M ultiple L ayers of F eatures from T iny I mages

    Alex Krizhevsky. L earning M ultiple L ayers of F eatures from T iny I mages. Technical report, University of Toronto, 2009

    2009

  14. [22]

    Fine-tuning can distort pretrained features and underperform out-of-distribution

    Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 ....

    2022

  15. [23]

    Kylberg texture dataset v

    Gustaf Kylberg. Kylberg texture dataset v. 1.0. Centre for Image Analysis, Swedish University of Agricultural Sciences and …, 2011

    2011

  16. [24]

    A S imple and E ffective B aseline for O ut-of-distribution D etection

    Kimin Lee, Kibok Lee, Honglak Lee, and In So Kweon. A S imple and E ffective B aseline for O ut-of-distribution D etection. In Advances in Neural Information Processing Systems (NeurIPS), volume 31, 2018

    2018

  17. [25]

    L earning W ithout F orgetting

    Zhizhong Li and Derek Hoiem. L earning W ithout F orgetting. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling (eds.), Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV , volume 9908 of Lecture ...

    2016 doi

  18. [26]

    Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . Op...

    2018

  19. [27]

    Owens, and Yixuan Li

    Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria - Florina Balcan, and Hsuan - Tien Lin (eds.), Advances in Neural Information Processing Systems 33: Annual Confer...

    2020

  20. [28]

    A valanche: A n E nd-to-end L ibrary for C ontinual L earning

    Vincenzo Lomonaco, Lorenzo Pellegrini, Andrea Cossu, Antonio Carta, Gabriele Graffieti, Tyler L Hayes, Matthias De Lange, Marc Masana, Jary Pomponi, Gido M van de Ven, et al. A valanche: A n E nd-to-end L ibrary for C ontinual L earning. I n 2021 I E E E . In CVF Conference on...

    2021

  21. [29]

    Gradient episodic memory for continual learning

    David Lopez - Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 30:...

    2017

  22. [30]

    O pen C I L : B enchmarking out-of-distribution detection in class incremental learning

    Wenjun Miao, Guansong Pang, Trong - Tung Nguyen, Ruohuan Fang, Jin Zheng, and Xiao Bai. O pen C I L : B enchmarking out-of-distribution detection in class incremental learning. Pattern Recognit., 171: 0 112163, 2026. doi:10.1016/J.PATCOG.2025.112163. URL https://doi.org/10.101...

    2026

  23. [31]

    Revisiting the calibration of modern neural networks

    Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaug...

    2021

  24. [32]

    H ow D oes F ine-tuning I mpact O ut-of-distribution D etection for V ision-language M odels? Int

    Yifei Ming and Yixuan Li. H ow D oes F ine-tuning I mpact O ut-of-distribution D etection for V ision-language M odels? Int. J. Comput. Vis., 132 0 (2): 0 596--609, 2024. doi:10.1007/S11263-023-01895-7. URL https://doi.org/10.1007/s11263-023-01895-7

    2024 doi

  25. [33]

    R eading D igits in N atural I mages with U nsupervised F eature L earning

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. R eading D igits in N atural I mages with U nsupervised F eature L earning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, volume 2011, pp.\ 5, 2011

    2011

  26. [34]

    Nearest neighbor guidance for out-of-distribution detection

    Jaewoo Park, Yoon Gyo Jung, and Andrew Beng Jin Teoh. Nearest neighbor guidance for out-of-distribution detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pp.\ 1686--1695. IEEE , 2023. doi:10.1109/ICCV51070.2023.00...

    2023

  27. [35]

    Anatomy of catastrophic forgetting: Hidden representations and task semantics

    Vinay Venkatesh Ramasesh, Ethan Dyer, and Maithra Raghu. Anatomy of catastrophic forgetting: Hidden representations and task semantics. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021. URL htt...

    2021

  28. [36]

    Sylvestre - Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pp.\ 5533--5...

    2017 doi

  29. [37]

    A da S C A L E : A daptive S caling for OOD D etection

    Sudarshan Regmi. A da S C A L E : A daptive S caling for OOD D etection. ArXiv preprint, abs/2503.08023, 2025. URL https://arxiv.org/abs/2503.08023

    2025

  30. [38]

    Progressive neural networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. ArXiv preprint, abs/1606.04671, 2016. URL https://arxiv.org/abs/1606.04671

    2016 arXiv

  31. [39]

    Overcoming catastrophic forgetting with hard attention to the task

    Joan Serr \` a , Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Jennifer G. Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm...

    2018

  32. [40]

    DICE: L everaging S parsification for O ut-of-distribution D etection

    Yiyou Sun and Yixuan Li. DICE: L everaging S parsification for O ut-of-distribution D etection. In Shai Avidan, Gabriel J. Brostow, Moustapha Ciss \' e , Giovanni Maria Farinella, and Tal Hassner (eds.), Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, ...

    2022 doi

  33. [41]

    React: Out-of-distribution detection with rectified activations

    Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan (eds.), Advances in Neural Information Processing Systems 34: Annual Con...

    2021

  34. [42]

    Out-of-distribution detection with deep nearest neighbors

    Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesv \' a ri, Gang Niu, and Sivan Sabato (eds.), International Conference on Machine Learning, ICML 2022, 17...

    2022

  35. [43]

    Three scenarios for continual learning

    Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019

    1904 arXiv

  36. [44]

    V i M : O ut- O f-distribution with V irtual-logit M atching

    Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. V i M : O ut- O f-distribution with V irtual-logit M atching. ArXiv preprint, abs/2203.10807, 2022. URL https://arxiv.org/abs/2203.10807

    2022 arXiv

  37. [45]

    Large scale incremental learning

    Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pp.\ 374--382. Computer Vision Foundation ...

    2019

  38. [46]

    G eneralized out-of-distribution detection: A survey

    Jize Yang, Kai Zhou, and Yixuan Li. G eneralized out-of-distribution detection: A survey. ArXiv preprint, abs/2110.11334, 2021. URL https://arxiv.org/abs/2110.11334

    2021 arXiv

  39. [47]

    M N I S T handwritten digit database

    LeCun Yann. M N I S T handwritten digit database. ATT Labs., 2010

    2010

  40. [48]

    Continual learning through synaptic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017 , volume 70 of P...

    2017

  41. [49]

    O pen O O D v1.5: E nhanced B enchmark for O ut-of-distribution D etection

    Jingyang Zhang, Jingkang Yang, Pengyun Wang, Haoqi Wang, Yueqian Lin, Haoran Zhang, Yiyou Sun, Xuefeng Du, Yixuan Li, Ziwei Liu, Yiran Chen, and Hai Li. O pen O O D v1.5: E nhanced B enchmark for O ut-of-distribution D etection. J. Data-centric Mach. Learn. Res., 2: 0 (3):1--3...

    2024

  42. [50]

    M aintaining discrimination and fairness in class incremental learning

    Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shutao Xia. M aintaining discrimination and fairness in class incremental learning. 2020 I E E E . In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 13205--13214, 2019

    2020

  43. [51]

    P laces: A 10 M illion I mage D atabase for S cene R ecognition

    Bolei Zhou, \` A gata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. P laces: A 10 M illion I mage D atabase for S cene R ecognition. IEEE Trans. Pattern Anal. Mach. Intell. , 40 0 (6): 0 1452--1464, 2018. doi:10.1109/TPAMI.2017.2723009. URL https://doi.org/10.110...

    2018

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.