Learning Continually from Low-shot Data Stream

Biao Wang; Canyu Le; Lei Zhang; Xihan Wei; Zhonggui Chen

REVIEW 2 major objections 5 minor 31 references

Learning Continually from Low-shot Data Stream

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read MetaCL, a model-agnostic algorithm, strengthens adaptation in continual low-shot learning by combining a Reptile-style multi-step inner loop with an adaptive gradient-balancing rule, and the paper reports consistent gains over EWC, PI…

desk verdict MetaCL is a genuinely new combination with consistent empirical gains, but the derivation that the Reptile inner loop optimizes Eq. 7 rests on an unjustified commutativity assumption and needs major revision. read the letter →

arxiv 1908.10223 v2 pith:KR3CFSCA submitted 2019-08-27 cs.LG stat.ML

Canyu Le , Xihan Wei , Biao Wang , Lei Zhang , Zhonggui Chen This is my paper · ORCID

classification cs.LGstat.ML

keywords continuallearninglow-shotmeta-learningcatastrophicforgettingparameterregularizationReptiledynamicbalancestrategyimageclassification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper addresses a combined problem that standard continual learning does not cover: models must learn from only a handful of examples per task and keep moving through a stream of tasks without forgetting. The proposed algorithm, MetaCL, wraps any parameter-regularization continual learner with two additions: a multi-step inner loop that extracts shared features from the small task data, and an adaptive rule that weights the forgetting-prevention gradient according to how much it conflicts with the current learning gradient. On Permuted MNIST, CIFAR-100, and CUB, the MetaCL variants consistently improve average accuracy, backward transfer, and forward adaptation over the three regularizers they are built on, with the largest gains appearing in the lowest-shot settings. The authors' argument is that strengthening adaptation, rather than only protecting old parameters, opens more room for the learning-forgetting trade-off. If correct, the method shows that continual low-shot learning can be made practical without storing old data or changing architecture.

What carries the argument

The central machinery is a pair of optimization tricks. First, a Reptile-style multi-step inner loop: a mini-batch is split into $m$ mini-bundles, the model takes $m$ small gradient steps $\theta_i = \theta_{i-1} - \alpha \ell'_{b_i}(\theta_{i-1})$, and the total displacement, scaled as $g_1 = (\theta - \theta_m)/(\alpha m)$, is used as the learning gradient. A Taylor expansion, relying on an exchangeability assumption, identifies this displacement as the gradient of an objective that rewards agreement between gradients of different mini-bundles, which the paper calls meta knowledge. Second, a dynamic balance rule solves the constrained problem $\min_{g_x} \frac{1}{2}\|g_x - g_1\|^2$ subject to $\langle g_x, g_2\rangle \ge 0$, where $g_2$ is the gradient of the preservation penalty; the closed form is $\lambda = \max(0, -g_1^T g_2 / (g_2^T g_2))$ and $g_x = g_1 + \lambda g_2$. This replaces the fixed hyperparameter $\beta$ with an adaptive weight that keeps the update close to learning while never increasing the penalty.

What would settle it

Compute the true gradient of the meta-objective in Eq. 7 with automatic differentiation on a small network and compare it with the inner-loop gradient $g_1 = (\theta - \theta_m)/(\alpha m)$ under the paper's $\alpha = 0.01$ and $m = 10$; if the directions disagree, or if permuting the order of mini-bundles changes $g_1$ noticeably, the mechanism the paper attributes to its gains is not the actual cause.

Watch

Extended reading notes

Core claim

The central claim is that continual low-shot learning can be improved by strengthening adaptation rather than only reducing forgetting. MetaCL does this in two steps. Inside each task it runs several small gradient updates on random mini-bundles and uses the total displacement as the learning gradient; by the Taylor argument this displacement approximates the gradient of an objective that rewards agreement between gradients of different mini-bundles, so the model pulls out common features without computing second derivatives. It then replaces the fixed regularization weight $\beta$ with a per-step $\lambda$ computed from the angle between the learning gradient $g_1$ and the regularization gradient $g_2$, so the update stays close to $g_1$ while never increasing the penalty. The paper reports that this combination yields higher average accuracy, less backward interference, and better forward adaptation than the underlying EWC, PI, and MAS baselines on Permuted MNIST, CIFAR-100, and CUB.

Load-bearing premise

The load-bearing premise is that the order in which small random data chunks are processed inside the inner loop can be swapped without changing the average update direction; this approximation, together with small step sizes, is what turns the inner-loop displacement into the gradient of the shared-feature meta-objective, and if it fails the claimed meta-knowledge mechanism has no support.

Editorial extensions

If this is right

Wrapping EWC, PI, or MAS with MetaCL raises final average accuracy in the reported low-shot settings; on CUB with 10 examples per class, EWC-MetaCL variants exceed plain EWC by more than 10 percentage points.
MetaCL improves forward adaptation in most configurations, meaning knowledge from earlier tasks helps the model learn the current task faster than a randomly initialized reference model.
The dynamic balance variant MetaCL-$\lambda$ generally matches or outperforms the fixed-balance MetaCL-$\beta$, with more stable gradient directions as tasks accumulate.
The gains shrink as per-task data grows from 20 to 500 examples per class, consistent with the claim that the adaptation mechanism matters most when data are scarce.
The method is model-agnostic and memory-efficient, requiring neither architecture changes nor storage of old task data.

Reading between the lines

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

As an extension the paper does not test, the adaptive-balance rule could be applied to any two competing gradient terms, such as trading a task loss against a fairness or safety penalty, since it requires only the two gradient vectors.
As an editorial inference, the inner loop's 'meta knowledge' is a within-task shared-feature effect, not cross-task meta-learning; coupling MetaCL with actual cross-task meta-training could push the low-shot limit further, but that remains untested.
If the exchangeability approximation is the true engine, intentionally ordering mini-bundles, for example grouping same-class samples, should weaken MetaCL's gains; running that experiment would isolate the mechanism from the raw benefit of multiple inner steps.
The paper reports vision benchmarks only; whether the adaptation benefit transfers to non-vision streams or to tasks with structured outputs is an open question the method does not address.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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: the paper is a solid empirical study wrapped in an overclaimed theoretical package. The combination of a Reptile-style inner loop with EWC/PI/MAS and an adaptive balance weight is new, and the gains on low-shot streams are consistent. But the proof that the inner loop optimizes the meta-objective relies on a commutativity assumption that doesn't hold for general loss functions. The authors should fix or soften that claim, add missing hyperparameters, and broaden the baseline comparison. Still, it deserves a real referee.

What's good: the problem—continual learning where each task has only tens of examples—is real and underexplored. The method is model-agnostic and memory-free, which is attractive in practice. The dynamic balance derivation (Eqs. 3-5) is correct for the stated constrained optimization: projecting g1 to keep the regularizer from increasing is a sensible idea, and the closed-form lambda is simple and cheap. The experiments show consistent improvements over fine-tuning, EWC, PI, and MAS across Permuted MNIST, CIFAR-100, and CUB, with confidence intervals. The FA (forward adaptation) metric is a nice addition. The fact that MetaCL w/o reg already helps on CIFAR-100 and CUB suggests the Reptile inner loop does something useful even without regularization.

Where it's soft: the main problem is Eq. 13. The stress-test note is right: random sampling of mini-bundles does not imply that l''_bi(theta) l'_bk(theta) equals l''_bk(theta) l'_bi(theta). That is an algebraic condition that fails for generic neural network losses. Without it, the symmetrized term in Eq. 13 does not follow, so g1 is not the gradient of the inner-product objective in Eq. 7. This is a load-bearing step in the "implicitly satisfies Eq. 7" argument. The empirical results can still stand—Reptile is known to provide useful inductive bias—but the theoretical framing overclaims. The authors could try to rescue it by taking expectations over random mini-bundles, which might justify a symmetrized term on average, or they could present the inner loop as a heuristic.

Other issues: beta and gamma are never reported, which makes reproduction painful. The "state-of-the-art" claim is based on comparing against only three regularization methods; replay-based methods like GEM/A-GEM (which use memory) or other rehearsal-free projection methods are not included. Sampling the first K images per class is not a random subsample and could introduce order effects. And K=1,5 are deferred, which is honest, but it means the "low-shot" claim is really about K=10-20.

Bottom line: this paper is useful to people working on continual few-shot learning and regularization. It should go to peer review, but it needs major revision on the theory, complete hyperparameters, and a more honest scope of comparison. I'd accept the empirical findings as preliminary evidence that the combination helps. My recommendation for the editor: send it to a venue that allows any form of rebuttal, and require the authors to either fix the derivation or explicitly label the inner loop as a heuristic, and to release code. If they do, it could become a solid publication.

Referee Report

2 major / 5 minor

Summary. The paper proposes MetaCL, an algorithm for continual low-shot learning that combines a Reptile-style multi-step inner-loop update over mini-bundles of the current task's limited data with a dynamic balance strategy that adaptively sets the regularization weight λ for parameter-importance-based continual learning methods. The method is instantiated on EWC, PI, and MAS, and evaluated on Permuted MNIST, CIFAR-100, and CUB under low-shot protocols. The reported results indicate that MetaCL variants improve average accuracy, backward transfer, and forward adaptation over the base regularizers and over fine tuning.

Significance. If the central claims hold, the paper addresses a realistic and under-studied setting at the intersection of continual learning and few-shot learning. The dynamic balance projection in Eqs. (3)-(5) is a clean, model-agnostic technique that can be grafted onto any parameter-regularization method, and the experimental study is moderately thorough: three datasets, three base regularizers, two MetaCL variants, repeated runs with confidence intervals, and ablations over shot number. The main weakness is theoretical: the derivation that the inner loop implicitly optimizes the meta-objective in Eq. (7) rests on an unjustified commutativity assumption, so the 'meta-knowledge' mechanism is not established even though the reported empirical gains may still be real. The dynamic balance component appears mathematically sound and is the more defensible contribution.

major comments (2)

[Strengthening Adaptation, Eq. (13)] The assertion that random sampling of mini-bundles implies the operator equality l''_{b_i}(θ) l'_{b_k}(θ) = l''_{b_k}(θ) l'_{b_i}(θ) is not justified. Random sampling makes the bundle indices exchangeable in distribution, but it does not make the Hessian-gradient products commute pathwise; for a generic loss, these matrices and vectors do not commute. Consequently, the symmetrized expression in Eq. (13) does not follow from Eq. (12), and g1 in Eq. (14) is not the gradient of the inner-product objective in Eq. (7). The stated small-step and small-m caveat controls the Taylor remainder but not this algebraic gap. The authors should either provide a valid proof that the expected gradients match (which would require stating the result in expectation and correctly handling the unsymmetrized sum), or revise the text to present the inner loop as a heuristic inspired by Eq. (6), with empirical validation of the mechanism.
[Abstract and Introduction (meta-knowledge claim)] The paper's narrative that the inner loop 'learns meta knowledge' is stronger than what is demonstrated. In the continual low-shot setup there is no meta-training dataset, and Eq. (6) maximizes agreement among gradients of data points from the current task only; this is a within-task regularizer, not cross-task meta-knowledge in the usual meta-learning sense. Unless the authors clarify the intended definition and supply evidence that this within-task gradient agreement transfers across tasks (for example, through a forward-transfer or initialization analysis), the abstract and introduction overstate the result.

minor comments (5)

[Eq. (14)] The subscript in the second term of Eq. (14) is written as b_t but should be b_k; this appears to be a typo that obscures the derivation.
[Algorithm 2 and Dynamic Balance Strategy] The text states that adding a small constant γ > 0 to the adaptive weight λ fortifies knowledge preservation, but Algorithm 2 does not include γ in the pseudocode, and no value of γ is reported. The manuscript should specify where γ is added and provide its value for reproducibility.
[Eq. (16), Backward Transfer] The formula for BT is ambiguous: the expression min_{i∈{1,...,n−1}} a_{n,j} − a_{i,j} is not the standard backward-transfer measure in the cited literature, and the range of the minimum is unclear. Please clarify the definition and its relationship to the standard measure.
[Tables and text] There are several typographical errors, including 'Fune tine' in Table 3 and 'Permutated MNIST' in the text; these should be corrected.
[Appendix B and hyperparameters] Appendix B does not state how β is selected for the fixed-balance methods (base regularizers and MetaCL-β), nor how γ is set for MetaCL-λ; without this information the comparisons could be sensitive to hyperparameter choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MetaCL's method and results are evaluated against external baselines, and no claimed prediction reduces by construction to a fitted input or to a self-citation.

full rationale

The paper's central technical claim is that the Reptile-style inner loop in Algorithm 1 implicitly optimizes the objective in Eq. 7. This connection is made through a Taylor expansion and an exchangeability assumption in Eq. 13, namely that l''_b_i(theta) l'_b_k(theta) = l''_b_k(theta) l'_b_i(theta). That assumption is mathematically questionable and not justified by random sampling alone, so the derivation may be unsound. However, unsoundness is a correctness concern, not circularity: Eq. 7 is stated independently of the algorithm, and the algorithm is not defined in terms of the claim it is meant to establish. The dynamic balance weight lambda in Eq. 5 is derived from a constrained quadratic program and computed from the current gradients g1 and g2; it is not fitted to the test set or renamed from fitted data. The method is compared with external baselines (EWC, PI, MAS, and fine tuning), and the meta-learning inspiration is explicitly attributed to Reptile (Nichol et al. 2018), an external algorithm. There is no load-bearing self-citation chain, no fitted input called a prediction, and no known result merely renamed. Tuned hyperparameters such as beta, gamma, alpha, and eta are standard and do not make the empirical comparisons circular. Therefore the appropriate finding is no significant circularity, with a score of 0.

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

The central claim rests on a few approximations and hyperparameters. The most fragile is the exchangeability assumption in Eq. 13, which is not generally valid. The method also depends on standard domain assumptions about importance measures and on unreported values for beta and gamma.

free parameters (5)

beta (fixed balance coefficient) = not reported
Balances current task loss and regularization in MetaCL-beta; the paper criticizes fixed beta but does not disclose the value used in experiments.
gamma (constant added to lambda) = not reported
Mentioned as a small constant to further fortify knowledge preservation, but value not specified and not included in Algorithm 2.
m (number of inner-loop mini-bundles) = 10
Mini-batch size 100, mini-bundle size 10, giving m=10 inner steps; affects the meta-gradient approximation.
alpha (inner step size) = 0.01
SGD learning rate for the Reptile-style inner loop; chosen without a reported tuning procedure.
eta (outer step size) = 0.001
Adam learning rate for the outer update; standard choice but a free parameter.

assumptions (5)

standard math Quadratic programming duality (Dorn 1960) applies to the constrained optimization in Eq. 3.
Used to derive the dual problem and the closed-form solution for lambda in Eq. 4-5.
domain assumption The objective is locally linear around small optimization steps, allowing the constrained formulation in Eq. 3.
The paper states 'Suppose the objective function is locally linear (it happens around small optimization steps)' to justify the projection formulation.
ad hoc to paper Taylor expansion with higher-order terms neglected is valid for the inner-loop updates.
Used in Algorithm analysis to approximate l'_b_i(theta_{i-1}) and to derive the meta-gradient expression; validity depends on small inner steps.
ad hoc to paper The exchangeability assumption l''_b_i(theta) l'_b_k(theta) = l''_b_k(theta) l'_b_i(theta) holds for randomly sampled mini-bundles.
Stated in Algorithm analysis (Eq. 13); this is not generally true and is crucial for rewriting the gradient sum as a derivative of an inner product.
domain assumption Parameter importance measures (from EWC, PI, MAS) accurately capture the knowledge of previous tasks under the continual low-shot setting.
This is the standard assumption underlying all parameter-regularization methods and is adopted without additional justification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Continually from Low-shot Data Stream." pith.science (2026). https://pith.science/paper/KR3CFSCA

@misc{pith2026190810223,
  author       = {Pith},
  title        = {Pith review of: Learning Continually from Low-shot Data Stream},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KR3CFSCA}},
  note         = {Machine review of arXiv:1908.10223}
}

read the original abstract

While deep learning has achieved remarkable results on various applications, it is usually data hungry and struggles to learn over non-stationary data stream. To solve these two limits, the deep learning model should not only be able to learn from a few of data, but also incrementally learn new concepts from data stream over time without forgetting the previous knowledge. Limited literature simultaneously address both problems. In this work, we propose a novel approach, MetaCL, which enables neural networks to effectively learn meta knowledge from low-shot data stream without catastrophic forgetting. MetaCL trains a model to exploit the intrinsic feature of data (i.e. meta knowledge) and dynamically penalize the important model parameters change to preserve learned knowledge. In this way, the deep learning model can efficiently obtain new knowledge from small volume of data and still keep high performance on previous tasks. MetaCL is conceptually simple, easy to implement and model-agnostic. We implement our method on three recent regularization-based methods. Extensive experiments show that our approach leads to state-of-the-art performance on image classification benchmarks.

Figures

Figures reproduced from arXiv: 1908.10223 by the authors.

**Figure 2.** The difference between two strategies. The dy [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The average accuracy changes as more tasks are learned with different K. The parameter regularization based methods [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The learning speed and average accuracy compari [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: The gradient angles (hg1, gxi, hg2, gxi) change in optimization procedure. The dynamic balance strategy can provide a more stable and better compromise. Conclusion In this paper, we generalize the standard continual learning to low-shot scenario. The low-shot setting i…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 23 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...
[2]

Aljundi, R.; Babiloni, F.; Elhoseiny, M.; Rohrbach, M.; and Tuytelaars, T. 2018. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV) , 139--154

work page 2018
[3]

Aljundi, R.; Chakravarty, P.; and Tuytelaars, T. 2017. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 3366--3375

work page 2017
[4]

Aljundi, R.; Rohrbach, M.; and Tuytelaars, T. 2018. Selfless sequential learning. arXiv preprint arXiv:1806.05421

arXiv 2018
[5]

K.; Ajanthan, T.; and Torr, P

Chaudhry, A.; Dokania, P. K.; Ajanthan, T.; and Torr, P. H. 2018a. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European Conference on Computer Vision (ECCV) , 532--547

work page
[6]

Chaudhry, A.; Ranzato, M.; Rohrbach, M.; and Elhoseiny, M. 2018b. Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420

arXiv
[7]

Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , 248--255. Ieee

2009
[8]

Dorn, W. S. 1960. Duality in quadratic programming. Quarterly of Applied Mathematics 18(2):155--162

work page 1960

Show all 31 references

[9]

Finn, C.; Abbeel, P.; and Levine, S. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 , 1126--1135. JMLR. org

2017
[10]

Gidaris, S., and Komodakis, N. 2018. Dynamic few-shot visual learning without forgetting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 4367--4375

2018
[11]

He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , 770--778

2016
[12]

P., and Ba, J

Kingma, D. P., and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

2014 arXiv
[13]

A.; Milan, K.; Quan, J.; Ramalho, T.; Grabska-Barwinska, A.; et al

Kirkpatrick, J.; Pascanu, R.; Rabinowitz, N.; Veness, J.; Desjardins, G.; Rusu, A. A.; Milan, K.; Quan, J.; Ramalho, T.; Grabska-Barwinska, A.; et al. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13):3521--3526

2017
[14]

Krizhevsky, A.; Hinton, G.; et al. 2009. Learning multiple layers of features from tiny images. Technical report, Citeseer

2009
[15]

LeCun, Y. 1998. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/

1998
[16]

Lee, S.-W.; Kim, J.-H.; Jun, J.; Ha, J.-W.; and Zhang, B.-T. 2017. Overcoming catastrophic forgetting by incremental moment matching. In Advances in neural information processing systems , 4652--4662

2017
[17]

Li, Z., and Hoiem, D. 2017. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence 40(12):2935--2947

2017
[18]

Lopez-Paz, D., et al. 2017. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems , 6467--6476

2017
[19]

McCloskey, M., and Cohen, N. J. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation , volume 24. Elsevier. 109--165

1989
[20]

Munkhdalai, T., and Yu, H. 2017. Meta networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 , 2554--2563. JMLR. org

2017
[21]

Nichol, A.; Achiam, J.; and Schulman, J. 2018. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999

2018 arXiv
[22]

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

2017
[23]

A.; Rabinowitz, N

Rusu, A. A.; Rabinowitz, N. C.; Desjardins, G.; Soyer, H.; Kirkpatrick, J.; Kavukcuoglu, K.; Pascanu, R.; and Hadsell, R. 2016. Progressive neural networks. arXiv preprint arXiv:1606.04671

2016 arXiv
[24]

Sabour, S.; Frosst, N.; and Hinton, G. E. 2017. Dynamic routing between capsules. In Advances in neural information processing systems , 3856--3866

2017
[25]

Santoro, A.; Bartunov, S.; Botvinick, M.; Wierstra, D.; and Lillicrap, T. 2016. Meta-learning with memory-augmented neural networks. In International conference on machine learning , 1842--1850

2016
[26]

Snell, J.; Swersky, K.; and Zemel, R. 2017. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems , 4077--4087

2017
[27]

H.; and Hospedales, T

Sung, F.; Yang, Y.; Zhang, L.; Xiang, T.; Torr, P. H.; and Hospedales, T. M. 2018. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 1199--1208

2018
[28]

Vinyals, O.; Blundell, C.; Lillicrap, T.; Wierstra, D.; et al. 2016. Matching networks for one shot learning. In Advances in neural information processing systems , 3630--3638

2016
[29]

Wah, C.; Branson, S.; Welinder, P.; Perona, P.; and Belongie, S. 2011. The Caltech-UCSD Birds-200-2011 Dataset . Technical Report CNS-TR-2011-001, California Institute of Technology

2011
[30]

Yoon, J.; Yang, E.; Lee, J.; and Hwang, S. J. 2018. Lifelong learning with dynamically expandable networks

2018
[31]

Zenke, F.; Poole, B.; and Ganguli, S. 2017. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 , 3987--3995. JMLR. org

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

[2] [2]

Aljundi, R.; Babiloni, F.; Elhoseiny, M.; Rohrbach, M.; and Tuytelaars, T. 2018. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV) , 139--154

work page 2018

[3] [3]

Aljundi, R.; Chakravarty, P.; and Tuytelaars, T. 2017. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 3366--3375

work page 2017

[4] [4]

Aljundi, R.; Rohrbach, M.; and Tuytelaars, T. 2018. Selfless sequential learning. arXiv preprint arXiv:1806.05421

arXiv 2018

[5] [5]

K.; Ajanthan, T.; and Torr, P

Chaudhry, A.; Dokania, P. K.; Ajanthan, T.; and Torr, P. H. 2018a. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European Conference on Computer Vision (ECCV) , 532--547

work page

[6] [6]

Chaudhry, A.; Ranzato, M.; Rohrbach, M.; and Elhoseiny, M. 2018b. Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420

arXiv

[7] [7]

Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , 248--255. Ieee

2009

[8] [8]

Dorn, W. S. 1960. Duality in quadratic programming. Quarterly of Applied Mathematics 18(2):155--162

work page 1960