Saving for the future: Enhancing generalization via partial logic regularization

Anh Nguyen; Kaizhu Huang; Qiufeng Wang; Xi Yang; Yijie Hu; Zhaorui Tan

REVIEW 4 major objections 5 minor 51 references

Saving for the future: Enhancing generalization via partial logic regularization

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper argues that visual classification with unknown classes is best modeled by partial logic, and that reserving an undefined part of feature space improves generalization; it introduces PL-Reg, a partial-logic regularization term, an

desk verdict The empirical recipe is real and consistent, but Proposition 1 is false as stated—the paper's central theoretical claim does not survive contact with a simple counterexample. read the letter →

arxiv 2508.15317 v1 pith:ZGQ7PQIM submitted 2025-08-21 cs.LG

Zhaorui Tan , Yijie Hu , Xi Yang , Qiufeng Wang , Anh Nguyen , Kaizhu Huang This is my paper

classification cs.LG

keywords partiallogicregularizationgeneralizedcategorydiscoverymulti-domaingeneralizationclassincrementallearningunknownclassesvisualclassification

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 the failure modes of visual classifiers on classes never seen at training time have a logical explanation: from the model's point of view at training time, statements about those classes are not false but undefined, and forcing them into a true/false scheme wastes capacity that later stages will need. To exploit this, it proposes PL-Reg, a regularization term that uses a learned mask to split the feature embedding into a defined part used for known-class logic and an undefined part reserved for future classes. The central theoretical claim, Proposition 1, is that under perfect fitting of known classes, a partial-logic model has generalization loss on unseen-unknown pairs no greater than a sentential-logic model. Experiments in generalized category discovery, multi-domain generalized category discovery, and long-tailed class incremental learning consistently show gains on unknown classes, at the cost of a small drop on known classes.

What carries the argument

The load-bearing object is partial logic LP, whose models assign each formula a truth value in {0,1,2}, with 2 meaning 'undefined'; a formula and its negation are both undefined exactly when either is. The method's mechanism is the partial logic mask M: a linear layer with sigmoid produces M, the embeddings are split into Z⊙M (defined) and Z⊙(1−M) (undefined), and both are classified, forcing the model to leave explicit feature-room for formulas that are not yet meaningful. Proposition 1, GL(fS,f∗,¬(Xk,Yk)) ≥ GL(fP,f∗,¬(Xk,Yk)), is the theorem that carries the generalization claim.

What would settle it

A direct check: on one GCD benchmark, take a fixed pretrained encoder and train two classifiers—one with L-Reg and one with PL-Reg—matching all other hyperparameters. Measure accuracy on samples from classes that are completely absent from training. Proposition 1 predicts PL-Reg's accuracy on those classes is at least L-Reg's whenever both models reach the same accuracy on known classes; if PL-Reg is ever worse while known-class accuracy is equal, the inequality is false. Conversely, on a closed-set task with no undefined classes, the theory predicts equal performance; finding a large PL-Reg g

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that classification tasks involving unknown classes are formal partial logics: there exist sentences about unseen labels whose truth value is neither true nor false, but 'undefined' (truth value 2). A model trained with ordinary sentential logic either assigns these undefined formulas a meaning or discards them, so every learned feature tends to be committed to a known class. PL-Reg instead generates a per-sample mask M, sends Z⊙M through the classifier as 'defined' logic and Z⊙(1−M) as 'undefined' logic, and trains both halves with a binary cross-entropy loss plus an entropy term that keeps the mask diverse; the original L-Reg loss is still

Load-bearing premise

The proof assumes the ideal model treats every pair made of an unseen image and an unseen label as 'undefined', and assumes both compared models fit the known data perfectly, so the advantage of the partial-logic model is very close to built into the comparison.

Editorial extensions

If this is right

PL-Reg can be dropped onto existing GCD methods like PIM without changing their training objective, suggesting the partial-logic term is a portable regularizer rather than a new architecture.
If Proposition 1 holds, unknown-class accuracy should consistently improve relative to L-Reg in any target-shift task, not only the three benchmarks tested.
In CIL, session-specific masks imply the model saves different feature subsets per session, which is a concrete mechanism for mitigating catastrophic forgetting.
The known-class accuracy cost observed in experiments is a direct corollary of reserving feature capacity, so users must expect a tradeoff rather than a free lunch.
The theory predicts equality only when the task contains no undefined logic; tasks with no unknown classes should show no PL-Reg benefit.

Reading between the lines

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

An implication the authors leave implicit: the mask generator is effectively a learned gating mechanism, so any gated architecture (LSTM-style forget gates, for example) already has the raw material for partial logic; adding an explicit 'undefined' training signal could improve its generalization.
The paper assumes the oracle labels unseen-unknown pairs as undefined; a natural extension would test what happens when the test set contains classes that were known at training time but in new domains, where PL-Reg's predicted advantage should shrink.
A testable extension: vary the dimensionality of the reserved undefined part (i.e., how many mask entries are near zero) and measure unknown-class accuracy; the theory predicts a sweet spot where too little reservation collapses to L-Reg and too much reservation starves known-class logic.
Because the proof's inequality is driven by the extra truth value 2, a simpler implementation without the L-Reg loss on masked features might already capture part of the gain; the paper does not isolate this ablation.

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

Quick take: if you read this paper for the PL-Reg regularizer and the experiments, you'll find a simple, reproducible method with consistent gains on unknown classes across GCD, mDG+GCD, and CIL. If you read it for Proposition 1, the advertised proof that partial logic improves generalization, you'll find a proof that doesn't work. I agree with the reader's reject verdict, but the empirical part deserves more credit than a blanket dismissal.

What's actually new: the partial-logic framing—reserving an 'undefined' truth value so the model doesn't commit features to known classes—is a fresh way to think about open-world generalization. The implementation is straightforward: a sigmoid mask on embeddings, a binary defined/undefined classifier, and the L-Reg term. The paper ships code and runs three task families with standard backbones. The improvements in unknown-class accuracy are consistent, and the known-class trade-off is acknowledged in Section 5. That's honest, reproducible work.

Where it falls apart: Proposition 1. The proof compares the sets of values fS and fP can output, then treats that as if it bounds actual errors. Concretely, for (Xu,Yk), if f*=0, fS=1 and fP=2, the fS error is 1 and the fP error is 4, so the claimed inequality GL(fS) ≥ GL(fP) is simply false. For (Xu,Yu), the proof stipulates f*(Xu,Yu)=2 and assumes fP outputs 2, which is the conclusion wearing a costume. Nothing about training on known pairs constrains behavior on unseen pairs, so the 'perfect training' assumption does no work. The stress-test counterexample is correct, and I couldn't find a way to repair the statement without adding assumptions that defeat the purpose.

Secondary issue: Eq (11) minimizes the entropy of the per-sample mask distribution. The text says this avoids uniform masks and maintains diversity, but minimizing entropy pushes masks to be peaked, not diverse. It might still be a useful sparsity prior, but the description and the loss are at odds, and the paper doesn't clarify.

The experiments are real but tuned per dataset (Tables 2, 5, 13, 14), and the sensitivity analysis covers only one CIL setting. That's a moderate concern, not fatal.

Who should read this: people working on GCD or open-world recognition who want a cheap regularization trick. They can ignore the theory and test the mask. For the record, I'd send this to peer review—the empirical work is substantial enough to deserve referee time—but I'd expect the authors to either produce a correct theorem or drop the claim entirely and reframe as an empirical contribution.

Referee Report

4 major / 5 minor

Summary. The paper proposes PL-Reg, a regularization term based on partial logic, intended to improve generalization to unknown classes in Generalized Category Discovery (GCD), multi-domain GCD, and long-tailed Class Incremental Learning. It introduces a formal framework connecting visual classification to partial logic, states Proposition 1 claiming that models trained with partial-logic constraints have lower generalization loss than models trained with sentential-logic constraints on unseen pairs, and reports experiments showing gains in unknown-class accuracy across six GCD datasets, five mDG+GCD datasets, and two CIL settings. The code is made available.

Significance. If the theoretical result were valid, it would provide a principled justification for reserving undefined feature dimensions for future unknown classes. The empirical study is broad and generally well executed: it covers multiple tasks, includes reproducible code, reports sensitivity analysis, and provides visualizations. However, the central formal claim is not established: Proposition 1's proof is invalid, and the asserted inequality is false under the stated assumptions. The empirical results are plausible but cannot compensate for the unsupported proof. The contribution therefore reduces to an incremental regularizer whose theoretical motivation is not demonstrated.

major comments (4)

[§3.2.4 (Proposition 1, case (Xu,Yk))] The inequality '||fS(Xu,Yk) ∈ {0,1} − f*(Xu,Yk)||² ≥ ||fP(Xu,Yk) ∈ {0,1,2} − f*(Xu,Yk)||²' is not a consequence of the output alphabets. A larger output alphabet does not imply a smaller squared error, either pointwise or in expectation: with f*(Xu,Yk)=0, fS=1, fP=2, the fS error is 1 and the fP error is 4. Since the only training assumption constrains fS and fP on (Xk,Yk), their values on (Xu,Yk) are unconstrained. This case is therefore not proved.
[§3.2.4 (Proposition 1, case (Xu,Yu))] The proof stipulates f*(Xu,Yu)∈{0,2} and then asserts that fS has f*(Xu,Yu)∈{0} while fP has f*(Xu,Yu)∈{0,2}. This conflates the target's truth-value range with the realized outputs of the models. Nothing in the perfect-training assumption or in the training objective forces fP to output 2 on unseen-unknown pairs; the claimed advantage is essentially definitional rather than derived. The argument transfers a property of the target oracle to the trained model without justification.
[§3.2.4, Proposition 1 assumptions] The perfect-training assumption only gives zero loss on (Xk,Yk). The proof nevertheless concludes zero loss on (Xk,Yu) and compares errors on (Xu,Yk) and (Xu,Yu) as if the assumption constrained behavior on all unlabeled inputs. This is a load-bearing gap: generalization loss is defined on ¬(Xk,Yk), and no training signal determines fS or fP on those pairs.
[§3.3, Eq. (11)] LP2 = −(1/(B·dim)) Σ_i Σ_j m_ij log m_ij with m=softmax(M) over the last dimension. Since Σ_j m_ij=1, minimizing LP2 maximizes the entropy of each mask row, pushing m toward a uniform distribution over dimensions. This is the opposite of the stated intent to 'avoid uniformly distributed mask values' and 'maintain the diversity of the mask for each dimension'. As written, the regularizer's behavior contradicts its motivation; the formula or its description needs correction.

minor comments (5)

[§3.2.4, (Xk,Yu) case] 'f*(Xu,Yk) ∈ {0}' should presumably be 'f*(Xk,Yu) ∈ {0}'.
[§4.2.3, TerraIncognita] The text says GMDG+PL-Reg delivers 'the second-best accuracy of 8.93%' for all classes; this should be 48.93%.
[§4.3.3, ImageNet-Subset Shuffled Long-tailed] The claim 'improving the average accuracy from 49.2% (with L-Reg) to 48.6%' inverts the numbers in Table 16; the correct change is from 48.6% to 49.2%.
[§4.2.2] The text reports an average of 55.04% for GMDG+PL-Reg, while Table 7 reports 55.10%.
[Definitions 1–4 and Proposition 1] The notation mixes model functions with sets of admissible truth values (e.g., f*(Xu,Yk)∈{0,1} vs. a function value). Please clarify whether f* and f denote functions or sets of possible outputs, and state precisely how the squared norm is evaluated when the argument is a set.

Circularity Check

1 steps flagged · score 6.0 of 10

Proposition 1's claimed generalization advantage on unseen-unknown pairs is definitional: the oracle is assigned the partial-logic truth value 2 on (Xu,Yu), and the partial-logic model's lower loss is asserted from its output range, not from the training assumption.

self definitional [Section 3.2.4, Proposition 1 proof, case (Xu, Yu)]
"F or(Xu, Yu). In this case, Yu is unknown or undefined for the current situation and Xu may belong to Yk but not Yu or belong to purely undefined Yu. Thus, f ∗(Xu, Yu) ∈ {0, 2}. fS, however only allows defined relationships, thus f ∗(Xu, Yu) ∈ {0}, while f ∗(Xu, Yu) ∈ {0, 2}. It is obvious that: GL(fS, f∗, (Xu, Yu)) ≥ GL(fP , f∗, (Xu, Yu)), where the equality can only be achieved when every (Xu, Yu) forms an empty set."

The proposition's only training assumption is that both fS and fP are perfect on known pairs (Xk,Yk); nothing in that assumption forces fP to output 2 on (Xu,Yu). The proof instead compares output ranges: it stipulates f*(Xu,Yu) ∈ {0,2}, asserts fS outputs only 0 while fP can output 2, and then declares fP's loss no greater. If 'fP is a partial-logic model' is taken to mean, per Definition 3, that fP outputs 2 on undefined formulas, then the inequality holds because f* was assigned the same undefined truth value — the advantage is true by construction, not derived from learning on known data. If fP's actual output on (Xu,Yu) is not so constrained, the inequality is unsupported. Thus the central 'improved generalization' claim reduces to defining the oracle and the partial-logic model with

full rationale

The paper's central theoretical claim is Proposition 1 (Section 3.2.4). The (Xu,Yu) case of its proof makes the claimed advantage definitional: the target f* is stipulated to take the partial-logic value 2 on unseen-unknown pairs, and the partial-logic model fP is then assumed (via its output range, or via the semantics of Definition 3) to be able to match that value, while the sentential model fS cannot. That is not a consequence of the stated 'well trained on (Xk,Yk)' assumption; it is the conclusion restated as a definitional property of partial logic. Separately from circularity, the (Xu,Yk) case in the same proof is pointwise false: if f*(Xu,Yk)=0 and fP outputs 2, then fP's squared error is 4 while fS's error is at most 1, contradicting the asserted inequality. This reinforces that Proposition 1 is not established by the paper's own equations. The self-citations to L-Reg [41] are normal baseline-building and not load-bearing in a circular way; the experiments on GCD, mDG+GCD, and CIL are extensive and externally benchmarked, so the empirical contribution is not itself a fitted-input-renamed-as-prediction artifact. But because the paper's formal proof of improved generalization reduces to the definition of the partial-logic truth value on undefined pairs, the central theoretical claim is partially circular, warranting a score of 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central theoretical result rests on assumed truth-value assignments for unseen pairs and on the perfect-training condition. The method adds per-dataset tuneable weights and a pseudo-labeled defined/undefined classification loss with no external ground truth. The only 'invented' notion is the undefined truth value itself, which is borrowed from partial logic literature.

free parameters (3)

ωp1, ωp2, ωLReg (regularization weights) = vary per dataset/task (Tables 2, 5, 13, 14)
Chosen per dataset to maximize reported accuracy; this is post-hoc fitting.
lr-mult (mDG+GCD) = 0.5, 1.0, 5.0, 4.5, 1.0 per dataset
Learning-rate multiplier tuned per dataset, as reported in Table 5.
fine-tuning epochs in CIL = 50 epochs
Extended from the original UCIR training to allow convergence with added regularization; not justified by a principle.

assumptions (5)

domain assumption Partial-logic semantics from [4] apply to visual classification tasks
Section 3.2 assumes the five-tuple logic L(X,Y) can be constructed from image features and labels, and that the third truth value 'undefined' corresponds to reserved capacity for unknown classes.
ad hoc to paper Perfect-training assumption ||fS(Xk,Yk)-f*||²=0 and ||fP(Xk,Yk)-f*||²=0 in Proposition 1
Stated in Proposition 1; no mechanism guarantees this, and the proof then infers zero error on the unlabeled subset (Xk,Yu) from it without justification.
ad hoc to paper Target f* takes values in {0,1} on (Xu,Yk) and {0,2} on (Xu,Yu)
These value restrictions are assumed in the proof of Proposition 1 and are what force the inequality; they are not derived from any learning principle.
domain assumption Embedding dimensions are independent
Acknowledged in Section 5 (Limitations) as a preassumption of the mask-based logic construction; the paper admits this is a limitation.
domain assumption Single-label classification with finite label set
Stated in Notations, Section 3; restricts the theoretical claim to discrete classification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Saving for the future: Enhancing generalization via partial logic regularization." pith.science (2026). https://pith.science/paper/ZGQ7PQIM

@misc{pith2026250815317,
  author       = {Pith},
  title        = {Pith review of: Saving for the future: Enhancing generalization via partial logic regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZGQ7PQIM}},
  note         = {Machine review of arXiv:2508.15317}
}

read the original abstract

Generalization remains a significant challenge in visual classification tasks, particularly in handling unknown classes in real-world applications. Existing research focuses on the class discovery paradigm, which tends to favor known classes, and the incremental learning paradigm, which suffers from catastrophic forgetting. Recent approaches such as the L-Reg technique employ logic-based regularization to enhance generalization but are bound by the necessity of fully defined logical formulas, limiting flexibility for unknown classes. This paper introduces PL-Reg, a novel partial-logic regularization term that allows models to reserve space for undefined logic formulas, improving adaptability to unknown classes. Specifically, we formally demonstrate that tasks involving unknown classes can be effectively explained using partial logic. We also prove that methods based on partial logic lead to improved generalization. We validate PL-Reg through extensive experiments on Generalized Category Discovery, Multi-Domain Generalized Category Discovery, and long-tailed Class Incremental Learning tasks, demonstrating consistent performance improvements. Our results highlight the effectiveness of partial logic in tackling challenges related to unknown classes.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 46 canonical work pages

[1]

In: International Confer- ence on Machine Learning

Abbe, E., Bengio, S., Lotfi, A., Rizk, K.: Gen- eralization on the unseen, logic reasoning and degree curriculum. In: International Confer- ence on Machine Learning. pp. 31–60. PMLR (2023)

work page 2023
[2]

Advances in Neural Information Processing Systems 36 (2024)

Abbe, E., Cornacchia, E., Lotfi, A.: Provable advantage of curriculum learning on par- ity targets with mixed inputs. Advances in Neural Information Processing Systems 36 (2024)

work page 2024
[3]

arXiv preprint arXiv:2402.04875 (2024)

Ahuja, K., Mansouri, A.: On provable length and compositional generalization. arXiv preprint arXiv:2402.04875 (2024)

arXiv 2024
[4]

Studies in Logic, Springer, due to (2017)

Andr´ eka, H., N´ emeti, I., Sain, I.: Universal algebraic logic. Studies in Logic, Springer, due to (2017)

work page 2017
[5]

In: Proceed- ings of the AAAI Conference on Artificial Intelligence

Barbiero, P., Ciravegna, G., Giannini, F., Li´ o, P., Gori, M., Melacci, S.: Entropy-based logic explanations of neural networks. In: Proceed- ings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 6046–6054 (2022)

work page 2022
[6]

In: Proceedings of the European conference on computer vision (ECCV)

Beery, S., Van Horn, G., Perona, P.: Recog- nition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456–473 (2018)

work page 2018
[7]

Boix-Adsera, E., Saremi, O., Abbe, E., Ben- gio, S., Littwin, E., Susskind, J.: When can transformers reason with abstract symbols? arXiv preprint arXiv:2310.09753 (2023)

work page Pith review arXiv 2023
[8]

Advances in Neural Information Processing Systems 33, 2445–2457 (2020)

Boudiaf, M., Ziko, I., Rony, J., Dolz, J., Piantanida, P., Ben Ayed, I.: Information maximization for few-shot learning. Advances in Neural Information Processing Systems 33, 2445–2457 (2020)

work page 2020

Show all 51 references

[9]

In: Interna- tional Conference on Learning Representa- tions (2022), https://openreview.net/forum? id=O-r8LOR-CCA

Cao, K., Brbic, M., Leskovec, J.: Open- world semi-supervised learning. In: Interna- tional Conference on Learning Representa- tions (2022), https://openreview.net/forum? id=O-r8LOR-CCA

2022
[10]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Caron, M., Touvron, H., Misra, I., J´ egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650–9660 (2021)

2021
[11]

In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision

Chiaroni, F., Dolz, J., Masud, Z.I., Mitiche, A., Ben Ayed, I.: Parametric information maximization for generalized category discov- ery. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision. pp. 1729–1739 (2023)

2023
[12]

In: 2009 IEEE Conference on Computer Vision and Pattern Recognition

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248–255. Ieee (2009)

2009
[13]

In: Proceedings of the IEEE Interna- tional Conference on Computer Vision

Fang, C., Xu, Y., Rockmore, D.N.: Unbiased metric learning: On the utilization of mul- tiple datasets and web images for softening bias. In: Proceedings of the IEEE Interna- tional Conference on Computer Vision. pp. 1657–1664 (2013)

2013
[14]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Fini, E., Sangineto, E., Lathuili` ere, S., Zhong, Z., Nabi, M., Ricci, E.: A unified objective for novel class discovery. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9284–9292 (2021) 21

2021
[15]

The journal of machine learning research 17(1), 2096–2030 (2016)

Ganin, Y., Ustinova, E., Ajakan, H., Ger- main, P., Larochelle, H., Laviolette, F., Marc- hand, M., Lempitsky, V.: Domain-adversarial training of neural networks. The journal of machine learning research 17(1), 2096–2030 (2016)
[16]

arXiv preprint arXiv:2007.01434 (2020)

Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. arXiv preprint arXiv:2007.01434 (2020)

2007 arXiv
[17]

IEEE Transactions on Pattern Analy- sis and Machine Intelligence (2021)

Han, K., Rebuffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A.: Autonovel: Automatically discovering and learning novel visual cate- gories. IEEE Transactions on Pattern Analy- sis and Machine Intelligence (2021)

2021
[18]

Neural Computation MIT-Press (1997)

Hochreiter, S.: Long short-term memory. Neural Computation MIT-Press (1997)

1997
[19]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incremen- tally via rebalancing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 831–839 (2019)

2019
[20]

In: Uncertainty in Artifi- cial Intelligence

Hu, S., Zhang, K., Chen, Z., Chan, L.: Domain generalization via multidomain dis- criminant analysis. In: Uncertainty in Artifi- cial Intelligence. pp. 292–302. PMLR (2020)

2020
[21]

European Conference on Computer Vision (ECCV) (2022)

Junbum, C., Kyungjae, L., Sungrae, P., Sanghyuk, C.: Domain generalization by mutual-information regularization with pre- trained models. European Conference on Computer Vision (ECCV) (2022)

2022
[22]

In: Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision

Kalla, J., Biswas, S.: Robust feature learning and global variance-driven classifier align- ment for long-tail class incremental learn- ing. In: Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision. pp. 32–41 (2024)

2024
[23]

Advances in neural informa- tion processing systems 23 (2010)

Krause, A., Perona, P., Gomes, R.: Discrim- inative clustering by regularized information maximization. Advances in neural informa- tion processing systems 23 (2010)

2010
[24]

In: Proceedings of the IEEE Inter- national Conference on Computer Vision workshops

Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine-grained cate- gorization. In: Proceedings of the IEEE Inter- national Conference on Computer Vision workshops. pp. 554–561 (2013)

2013
[25]

Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)

2009
[26]

Advances in neural information processing systems 25 (2012)

Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolu- tional neural networks. Advances in neural information processing systems 25 (2012)

2012
[27]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Kukleva, A., Kuehne, H., Schiele, B.: Gen- eralized and incremental few-shot learn- ing by explicit learning and calibration without forgetting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9020–9029 (2021)

2021
[28]

In: Proceedings of the IEEE International Conference on Computer Vision

Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5542–5550 (2017)

2017
[29]

In: Proceedings of the AAAI conference on artificial intelligence

Li, Y., Gong, M., Tian, X., Liu, T., Tao, D.: Domain generalization via conditional invari- ant representations. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)

2018
[30]

In: Proceedings of the European conference on computer vision (ECCV)

Li, Y., Tian, X., Gong, M., Liu, Y., Liu, T., Zhang, K., Tao, D.: Deep domain general- ization via conditional invariant adversarial networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 624–639 (2018)

2018
[31]

Advances in Neural Information Processing Systems 36 (2024)

Li, Z., Huang, Y., Li, Z., Yao, Y., Xu, J., Chen, T., Ma, X., Lu, J.: Neuro- symbolic learning yielding logical constraints. Advances in Neural Information Processing Systems 36 (2024)

2024
[32]

arXiv preprint arXiv:2102.02038 (2021)

Liu, L., Zhou, T., Long, G., Jiang, J., Dong, X., Zhang, C.: Isometric propagation net- work for generalized zero-shot learning. arXiv preprint arXiv:2102.02038 (2021)

2021 arXiv
[33]

In: European Conference on Computer Vision

Liu, X., Hu, Y.S., Cao, X.S., Bagdanov, A.D., Li, K., Cheng, M.M.: Long-tailed class incre- mental learning. In: European Conference on Computer Vision. pp. 495–512. Springer (2022)

2022
[34]

In: Proceedings of the 5th Berkeley Symposium on Mathemat- ical Statistics and Probability

MacQueen, J.: Classification and analysis of multivariate observations. In: Proceedings of the 5th Berkeley Symposium on Mathemat- ical Statistics and Probability. pp. 281–297 (1967)

1967
[35]

IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5513–5533 (2022)

Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A.D., Van De Weijer, J.: Class-incremental learning: survey and per- formance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5513–5533 (2022)

2022
[36]

In: Proceedings of 22 the IEEE/CVF International Conference on Computer Vision

Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi- source domain adaptation. In: Proceedings of 22 the IEEE/CVF International Conference on Computer Vision. pp. 1406–1415 (2019)

2019
[37]

In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition

Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2001– 2010 (2017)

2001
[38]

Advances in Neural Information Processing Systems 32 (2019)

Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for contin- ual learning. Advances in Neural Information Processing Systems 32 (2019)

2019
[39]

arXiv preprint arXiv:1906.05372 (2019)

Tan, K.C., Liu, Y., Ambrose, B., Tulig, M., Belongie, S.: The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372 (2019)

2019 arXiv
[40]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Tan, Z., Yang, X., Huang, K.: Rethinking multi-domain generalization with a general learning objective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23512– 23522 (June 2024)

2024
[41]

In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024)

Tan, Z., Yang, X., Wang, Q., Nguyen, A., Huang, K.: Interpret your decision: Logical reasoning regularization for generalization in visual classification. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024)

2024
[42]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12183–12192 (2020)

2020
[43]

In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7492–7501 (2022)

2022
[44]

In: Pro- ceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition

Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Pro- ceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition. pp. 5018–5027 (2017)

2017
[45]

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

2011
[46]

arXiv preprint arXiv:2404.00560 (2024)

Xiao, C., Liu, B.: A theory for length gener- alization in learning to reason. arXiv preprint arXiv:2404.00560 (2024)

2024 arXiv
[47]

International Journal of Computer Vision pp

Xuan, S., Yang, M., Zhang, S.: Incremen- tal model enhancement via memory-based contrastive learning. International Journal of Computer Vision pp. 1–19 (2024)

2024
[48]

Inter- national Journal of Computer Vision 131(2), 552–571 (2023)

Yuan, J., Ma, X., Chen, D., Kuang, K., Wu, F., Lin, L.: Domain-specific bias filtering for single labeled domain generalization. Inter- national Journal of Computer Vision 131(2), 552–571 (2023)

2023
[49]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

Zhou, D.W., Wang, Q.W., Qi, Z.H., Ye, H.J., Zhan, D.C., Liu, Z.: Class-incremental learn- ing: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

2024
[50]

In: The Eleventh International Conference on Learning Representations (2022)

Zhou, D.W., Wang, Q.W., Ye, H.J., Zhan, D.C.: A model or 603 exemplars: Towards memory-efficient class-incremental learning. In: The Eleventh International Conference on Learning Representations (2022)

2022
[51]

Advances in Neural Information Pro- cessing Systems 34, 14306–14318 (2021) 23

Zhu, F., Cheng, Z., Zhang, X.y., Liu, C.l.: Class-incremental learning via dual augmen- tation. Advances in Neural Information Pro- cessing Systems 34, 14306–14318 (2021) 23

2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

In: International Confer- ence on Machine Learning

Abbe, E., Bengio, S., Lotfi, A., Rizk, K.: Gen- eralization on the unseen, logic reasoning and degree curriculum. In: International Confer- ence on Machine Learning. pp. 31–60. PMLR (2023)

work page 2023

[2] [2]

Advances in Neural Information Processing Systems 36 (2024)

Abbe, E., Cornacchia, E., Lotfi, A.: Provable advantage of curriculum learning on par- ity targets with mixed inputs. Advances in Neural Information Processing Systems 36 (2024)

work page 2024

[3] [3]

arXiv preprint arXiv:2402.04875 (2024)

Ahuja, K., Mansouri, A.: On provable length and compositional generalization. arXiv preprint arXiv:2402.04875 (2024)

arXiv 2024

[4] [4]

Studies in Logic, Springer, due to (2017)

Andr´ eka, H., N´ emeti, I., Sain, I.: Universal algebraic logic. Studies in Logic, Springer, due to (2017)

work page 2017

[5] [5]

In: Proceed- ings of the AAAI Conference on Artificial Intelligence

Barbiero, P., Ciravegna, G., Giannini, F., Li´ o, P., Gori, M., Melacci, S.: Entropy-based logic explanations of neural networks. In: Proceed- ings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 6046–6054 (2022)

work page 2022

[6] [6]

In: Proceedings of the European conference on computer vision (ECCV)

Beery, S., Van Horn, G., Perona, P.: Recog- nition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456–473 (2018)

work page 2018

[7] [7]

Boix-Adsera, E., Saremi, O., Abbe, E., Ben- gio, S., Littwin, E., Susskind, J.: When can transformers reason with abstract symbols? arXiv preprint arXiv:2310.09753 (2023)

work page Pith review arXiv 2023

[8] [8]

Advances in Neural Information Processing Systems 33, 2445–2457 (2020)

Boudiaf, M., Ziko, I., Rony, J., Dolz, J., Piantanida, P., Ben Ayed, I.: Information maximization for few-shot learning. Advances in Neural Information Processing Systems 33, 2445–2457 (2020)

work page 2020