Pith. sign in

REVIEW 4 major objections 7 minor 99 references

Continual Learning Beyond Experience Rehearsal and Full Model Surrogates

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SPARC, a rehearsal-free, surrogate-free parameter-isolation method, matches or beats replay-based continual learners with a small fraction of their parameters.

desk verdict A rehearsal-free parameter-isolation method with strong benchmark results, but the Class-IL gains rest on an unablated activation-based re-normalization step that needs a no-normalization baseline and sensitivity analysis before it can be taken at face value. read the letter →

arxiv 2505.21942 v1 pith:2LR2CZW6 submitted 2025-05-28 cs.LG stat.ML

Prashant Bhat , Laurens Niesten , Elahe Arani , Bahram Zonooz This is my paper
classification cs.LGstat.ML
keywords continuallearningcatastrophicforgettingparameterisolationrehearsal-freedepthwiseseparableconvolutionsclass-incrementalweightre-normalizationcomplementarysystems
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 proposes SPARC, a continual learning method that deliberately avoids the two standard crutches: storing old training examples for replay, and keeping full-model copies as surrogates. Its central claim is that per-task sub-networks built from depthwise-separable convolutions, paired with a shared set of pointwise filters updated by an exponential moving average, can consolidate knowledge across tasks at a fraction of the cost. On Seq-TinyImageNet, SPARC reports 32.29% Class-IL accuracy against 23.47% for CLS-ER, a full-surrogate method, using about 6% of CLS-ER's parameters; on Seq-CIFAR100 it reports 49.03% against 43.80%, again with far fewer parameters. If this holds, it would make memory-efficient continual learning practical without sacrificing accuracy, since the method also works without any rehearsal buffer.

What carries the argument

The load-bearing mechanism is the depthwise-separable convolution split: each task's working memory uses per-channel depthwise filters and a half-task-specific, half-shared pointwise filter bank, so the task-agnostic semantic memory is literally a subset of pointwise filters updated by Eq. 4 with EMA coefficient alpha=0.99. A second mechanism, Eq. 5, re-normalizes each task's classifier weights and biases by kappa/eta where eta is drawn from the upper tail of that task's activation quartiles, which the paper argues equalizes activation magnitudes across sub-networks in Class-IL inference. The third component is per-task batch normalization statistics, which avoids the train-test mismatch that arises when the same BN layer sees non-stationary data. Together these parts keep SPARC's parameter growth linear but slow (1.04M to 3.62M for 5 to 20 tasks on Seq-CIFAR100).

What would settle it

Train SPARC on Seq-CIFAR100 (5 tasks) exactly as described but with the re-normalization in Eq. 5 disabled (or with kappa=1); if per-task Class-IL accuracies stay balanced and final average accuracy stays near 49%, the reported gains do not depend on the heuristic. Alternatively, measure the max activation set A^t; if the chosen eta lands at the cap a <= Q3 + IQR on every task, the scaling is doing little work.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a carefully chosen parameter-isolation architecture can deliver the benefits of complementary learning systems without either experience rehearsal or full-model surrogates. SPARC grows a small sub-network per task, replacing ordinary convolutions with depthwise-separable convolutions to keep that sub-network cheap; half of the pointwise filters in each block are task-specific, while the other half are shared across tasks and updated at task boundaries as an exponential moving average of the previous task's filters. After each task, the classifier weights and biases are re-normalized using a per-task scale derived from the distribution of maximum activations, which the paper reports removes task-recency bias and yields balanced per-task accuracy. Empirically the method reports the best Class-IL accuracy in its comparison table on Seq-CIFAR100 (5 tasks) and Seq-TinyImageNet (10 tasks), with parameter counts an order of magnitude below surrogate-based baselines.

Load-bearing premise

The load-bearing premise is that the activation-derived re-normalization, with its fixed constant kappa=5, equalizes task-specific classifier scales on new datasets and task orders without validation data, an effect never tested by ablation.

Editorial extensions

If this is right

  • SPARC can be deployed in memory-constrained settings such as edge devices without a replay buffer, eliminating privacy and storage concerns tied to rehearsal.
  • Long task sequences remain affordable: on Seq-CIFAR100, parameter count grows from 1.04M at 5 tasks to 3.62M at 20 tasks, versus 216.7M to 2645M for progressive networks.
  • The weight re-normalization provides a validation-free way to counteract recency bias, so parameter-isolation methods can be used for class-incremental inference where task identity is unknown.
  • Since task-agnostic filters are shared, the model performs cross-task consolidation implicitly, so it can improve current-task performance without storing previous data.
  • Performance scales with capacity: widening or deepening the backbone raises Class-IL accuracy (up to 52.48% on Seq-CIFAR100 at full width), giving users a direct efficiency-accuracy trade-off.

Reading between the lines

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

  • The activation-quartile normalization is a post-hoc calibration technique that could transfer to other modular or parameter-isolated architectures, including prompt- or adapter-based continual learning, though the paper only tests it inside SPARC.
  • Varying the shared fraction of pointwise filters (currently fixed at half) is a natural experiment the paper does not run; the optimal ratio likely depends on task similarity and could further reduce parameter growth or improve forward transfer.
  • Because SPARC distributes capacity equally across tasks, its static allocation may over- or under-provision on uneven tasks; a similarity-based or difficulty-based allocation would be a direct extension.
  • The reported gains depend on comparing against baselines with much larger footprints; an apples-to-apples study with surrogate methods shrunk to SPARC's budget would clarify how much of the benefit comes from parameter isolation versus the DSC backbone.
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

4 major / 7 minor

Summary. The paper proposes SPARC, a rehearsal-free, surrogate-free continual-learning method. Task-specific 'working memories' are per-task depthwise-separable convolutional filter banks with task-specific batch normalization and a per-task classifier head; a task-agnostic 'semantic memory' is formed by sharing half of the pointwise filters across tasks and updating them with an exponential moving average at task boundaries (Eq. 4). A final classifier weight re-normalization (Eq. 5) rescales each task's classifier weights and biases by an activation-derived constant η computed from training-set FC maxima, with scale κ=5. The main empirical claims are: superior Class-IL accuracy on Seq-TinyImageNet (32.29% vs 23.47% for CLS-ER) and Seq-CIFAR100-5T (49.03% vs 43.80%) relative to rehearsal/surrogate baselines at a 200-sample buffer; parameter use of roughly 6% of full-model surrogates (1.04M, 1.90M, and 3.62M parameters for 5/10/20 tasks on CIFAR100); and competitive Task-IL accuracy. Ablations cover semantic consolidation (Table 5), width/depth scaling (Table 3), and EMA stability (Figure 4); additional analyses address forgetting, task-recency bias, and longer task sequences.

Significance. If correct, SPARC is a significant data point for efficient continual learning: it demonstrates that strict parameter isolation built on depthwise-separable convolutions, combined with a shared pointwise semantic memory, can reach or exceed the Class-IL accuracy of memory-intensive surrogate/replay systems at a fraction of the parameter budget. The paper has real experimental strengths: the backbone-matched control in Table 9 (DSC-equipped ER/DER++ still trail SPARC substantially), the complete-isolation ablation in Table 5, the EMA stability analysis in Figure 4, the parameter-growth table, and a candid limitations appendix. The principal weakness is attribution: the Class-IL gains rely on the Eq. 5 rescaling, which is never ablated against the raw model and whose free constants (κ, the IQR fence) are never swept, so it is unknown whether the headline results reflect the architecture, the calibration heuristic, or both. The Task-IL results, which are unaffected by the rescaling, partially mitigate this concern, but the central Class-IL claim needs the missing experiments before the conclusions can be taken at face value.

major comments (4)
  1. [§3.3, Eq. 5; §A.2; Tables 1–2] The stress-test concern is borne out by the manuscript as written: the headline Class-IL numbers (Seq-TinyImageNet 32.29% vs 23.47% for CLS-ER; Seq-CIFAR100-5T 49.03% vs 43.80%) depend on the activation-derived rescaling in Eq. 5, yet no configuration of SPARC without re-normalization is reported anywhere. Figure 6 compares re-normalization only against weight alignment (another rescaling method), not against the raw model, so the reader cannot tell whether the Class-IL advantage comes from the parameter-isolation architecture or from the specific rescaling recipe. Eq. 5 introduces two free choices — κ=5 and the IQR fence Q3+IQR — and the paper asserts in F.1 that no dataset-specific tuning is used, but no sensitivity analysis is provided for either choice. I request: (a) a no-re-normalization ablation on both main benchmarks; (b) a sweep over κ (e.g., 1, 2, 5, 10) and over the fence multiplier; and (c) an experiment applying the identical re-normalization to a strong surrogate-based baseline such as CLS-ER, to determine whether the improvement is architecture-specific or a generic post-hoc calibration gain. The Task-IL results are unaffected by this concern and help bound it, but they do not resolve it.
  2. [§3.2, Eq. 4; §3.4; §A.1] The semantic-memory update is specified inconsistently. Eq. 4 reads K̃c ← αK̃c + (1−α)K̃^{t−1} for t>2, while §3.4 and A.1 state that the EMA applies 'from the second task onward.' As written, the update at the end of task t never incorporates the pointwise filters trained on task t (it re-uses the already-frozen t−1 filters), which cannot be the intended algorithm; conversely the text's 'from the second task' contradicts the t>2 condition. Since α=0.99 and the update timing directly determine the consolidation behavior analyzed in Figure 4 and the results in Table 5, the exact update rule must be stated unambiguously: correct the index and the range, and specify whether the shared pointwise filters receive any gradient updates after task 1. As printed, the method is not reproducible.
  3. [Table 1; §4.1] The SPARC rows report standard deviations of 0.01–0.05 percentage points across three seeds (e.g., 32.29±0.01 and 65.66±0.01 on Seq-TinyImageNet; 49.03±0.05 on Seq-CIFAR100-5T). This is orders of magnitude below the seed-to-seed variation typically observed for deep continual-learning runs and far below the baseline variations in the same table (e.g., CLS-ER ±0.80). Please report the individual per-seed accuracies (or explain what the standard deviation was computed over); if the three runs share seeds or restrict stochasticity in some way, that must be stated, because it materially affects how the comparisons should be read.
  4. [Appendix E.3; §3.1; Table 4] Appendix E.3 states that SPARC 'assume[s] the knowledge of the number of tasks beforehand and equally distribute[s] total capacity among all tasks,' which is in tension with the main-text framing in §3.1 and §3.4 of allocating a fixed per-task subnetwork at each task boundary, and with the claim of linear growth. If the per-task filter budget (32/64/128/256, F.1) is fixed, no advance knowledge of the task count is needed and 'equal distribution' is vacuous; if, instead, per-task capacity is derived from an assumed total, that is a real deployment constraint that should appear in the main text. Please clarify which is the case, and reconcile Table 4's parameter growth with the stated linear-growth claim.
minor comments (7)
  1. [§4.2] The text refers to 'dynamic sparse methods (CLIP, NISPA, PackNet)'; the first name should be CLNP to match the legend of Figure 2.
  2. [Tables 3 and 5] The same SPARC configuration (width 0.5, depth 4) is reported as 49.03±0.05 in Table 3 and as 49.13±0.25 in Table 5; please reconcile.
  3. [Table 9] The '#Filters per task' entries [160, 320, 640, 1280] for ER and DER++ conflict with the text's statement that competitors are 'equipped with the same number of filters' as SPARC; presumably these are totals summed over the five tasks, but the caption should say so explicitly.
  4. [§A.3] The phrase 'Large Large Language Models (LLMs)' contains a duplicated word.
  5. [§F.3] The training-epoch exceptions list IMEX-Reg and DER++ w/ FPF, but neither method appears in Table 1; this appears to be leftover text from an earlier draft.
  6. [Figure 6 (left)] The legend labels 'Pre Weight Alignment' and 'Post Weight Alignment' in a panel whose caption says 'Effect of Weight Re-normalization'; the labels should refer to re-normalization.
  7. [General] No code or per-seed logs are released; given the ambiguity of Eq. 4 and the unusually low reported variance, releasing the code or per-seed accuracy logs would materially aid verification.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular reasoning: SPARC is an empirically evaluated architecture; its few heuristic choices are hand-set constants tested against external benchmarks, not derivations from the target claims.

full rationale

SPARC is presented as an empirical continual learning method, not as a derivation from first principles. The architecture components (DSC task-specific sub-networks, shared pointwise EMA consolidation, and activation-derived classifier rescaling in Eq. 5) are fixed design choices whose values are evaluated on standard benchmarks (Seq-CIFAR100, Seq-TinyImageNet, Seq-MiniImageNet) against external baselines. The weight re-normalization heuristic with kappa=5 is a hand-set hyperparameter, not a fitted parameter disguised as a prediction, and the paper does not derive a performance number from it by construction. Self-citations to CLS-ER, TAMiL, and TriRE are used as baselines or as prior building blocks (EMA consolidation), but the central claim does not rest on an unverified self-citation: SPARC's accuracy is measured independently on held-out test sets with standard CL protocols. The absence of a no-normalization ablation and the lack of sensitivity analysis for kappa and the IQR fence are legitimate robustness/correctness concerns, but they are not circularity because the reported numbers are external empirical outcomes rather than consequences of the method's definition. No equation in the paper reduces a claimed prediction to an input by construction, and no uniqueness or existence theorem is imported from the authors' own prior work to force the design. The paper is therefore essentially self-contained against external evidence, and the only mild self-relational aspect is the reuse of the authors' own CLS-inspired baselines, which does not make any result circular.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The empirical claims rest on a few hand-set hyperparameters (alpha, kappa, per-task filter widths) and domain assumptions about task boundaries, the re-normalization heuristic, and the sufficiency of depthwise separable convolutions. The two architectural modules, working and semantic memories, are implemented and ablated rather than purely hypothetical.

free parameters (3)
  • EMA decay alpha = 0.99
    Controls how slowly shared pointwise filters absorb new task information; authors fix alpha=0.99 and show in Figure 4 that faster aggregation lowers stability. Not derived; chosen by hand.
  • Weight re-normalization constant kappa = 5
    Scales task-specific classifier weights and biases in Eq. 5; fixed to 5 across all experiments with no ablation over kappa values.
  • Per-task filter widths = [32, 64, 128, 256] depthwise filters for layers 1-4
    Architecture capacity allocation per task; authors state this is a conscious choice to resemble ResNet-18 and ablate width and depth in Table 3, so it is hand-chosen.
assumptions (3)
  • domain assumption Task boundary and total task count are known before training; capacity is pre-partitioned equally among tasks.
    Stated in Section 3.1 and Appendix E.3; limits real-world applicability but defines the problem setting.
  • ad hoc to paper The interquartile-based max activation eta, with kappa=5, correctly rebalances classifier scales across tasks.
    Introduced for this paper only, with no theoretical justification and no ablation without it in main tables (Eq. 5, Section 3.3).
  • domain assumption Depthwise separable convolutions preserve enough representational capacity for continual learning tasks.
    Relies on prior literature (Chollet 2017, MobileNet) and the paper's own Table 9 ablation; presented as an established property.
invented entities (2)
  • Task-agnostic semantic memory (shared half of pointwise filters) independent evidence
    purpose: Consolidates cross-task knowledge via exponential moving average while keeping the model compact.
    Implemented and ablated in Table 5 and the stability analysis in Figure 4; it is a real module, not a hypothetical entity.
  • Task-specific working memories (depthwise filter sets per task) independent evidence
    purpose: Capture task-specific information with parameter isolation to prevent interference.
    Core architecture component shown to work across the evaluated benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Continual Learning Beyond Experience Rehearsal and Full Model Surrogates." pith.science (2026). https://pith.science/paper/2LR2CZW6

@misc{pith2026250521942,
  author       = {Pith},
  title        = {Pith review of: Continual Learning Beyond Experience Rehearsal and Full Model Surrogates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LR2CZW6}},
  note         = {Machine review of arXiv:2505.21942}
}
read the original abstract

Continual learning (CL) has remained a significant challenge for deep neural networks as learning new tasks erases previously acquired knowledge, either partially or completely. Existing solutions often rely on experience rehearsal or full model surrogates to mitigate CF. While effective, these approaches introduce substantial memory and computational overhead, limiting their scalability and applicability in real-world scenarios. To address this, we propose SPARC, a scalable CL approach that eliminates the need for experience rehearsal and full-model surrogates. By effectively combining task-specific working memories and task-agnostic semantic memory for cross-task knowledge consolidation, SPARC results in a remarkable parameter efficiency, using only 6% of the parameters required by full-model surrogates. Despite its lightweight design, SPARC achieves superior performance on Seq-TinyImageNet and matches rehearsal-based methods on various CL benchmarks. Additionally, weight re-normalization in the classification layer mitigates task-specific biases, establishing SPARC as a practical and scalable solution for CL under stringent efficiency constraints.

Figures

Figures reproduced from arXiv: 2505.21942 by the authors.

Figure 1
Figure 1. The SPARC architecture, using ResNet-18 of 4 layers with 2 blocks each. Task-specific working memories (shown in white) efficiently capture task-relevant information, while the task-agnostic semantic memory (highlighted in red) consolidates knowledge across tasks. This design enables SPARC to effectively balance plasticity and stability, achieving scalable continual learning without the need for full model surrogate… view at source ↗
Figure 2
Figure 2. Comparison with parameter isolation approaches on Seq-CIFAR100 with 20 tasks. We report the final accuracy of each task after training on all tasks. stability of previous tasks. This task-agnostic information consolidation enables SPARC to remain parameter-efficient while closely approximating the performance of hard param￾eter isolation (see [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of relative performance and model size of different CL approaches in Seq-TinyImageNet 10 tasks with respect to a JOINT model in Class-IL (left) and Task-IL (right) settings. Compared Task-IL methods include ALASSO (Park et al., 2019), UCB (Ebrahimi et al., 2020), oEWC (Kirkpatrick et al., 2017), SI (Zenke et al., 2017), BMKP (Sun et al., 2023), and PackNet (Mallya & Lazebnik, 2018) While semantic informat… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Effect of semantic information consolidation on SPARC’s stability. Stability represents model stability at task t, quantified as the average performance across all preceding tasks. Due to space limitations, we provide additional analysis such as information on Class-IL…
Figure 5
Figure 5. Figure 5: Depiction of training and inference regimes in SPARC in Class-IL setting. A.2. Prominence of weight re-normalization in SPARC Continual learning approaches are prone to task recency bias - the tendency of a CL model to be biased towards classes from the most recent tas…
Figure 6
Figure 6. Figure 6: (Left) L2-Normed weights prior to and after weight re-normalization for each task in Seq-CIFAR100-5T. (Right) Comparison of weight re-normalization against weight alignment in SPARC. A.3. Comparison with PEFT techniques in Vision Transformers Parameter Efficient Fine-T…
Figure 7
Figure 7. Figure 7: Comparison of task-wise Class-IL performance of several methods on Seq-CIFAR100 divided into 5 tasks. SPARC shows balanced performance on all 5 tasks after the completion of the training phase. E. Further analysis E.1. Task-wise performance We present the final accurac…
Figure 8
Figure 8. Figure 8: Comparison of task-probabilities on Seq-CIFAR100 with 5 tasks. SPARC achieves a significantly better balance between tasks compared to other methods. 20 40 60 80 100 Number of classes 30 40 50 60 70 80 Top-1 Accuracy (%) Seq-CIFAR100 in 5 tasks ER DER++ LiDER SPARC 20 …
Figure 9
Figure 9. Figure 9: Class-IL (AT % ) performance on Seq-CIFAR100 with 5, 10, and 20 tasks respectively. The graphs compare the performance of SPARC to ER and DER++ across varying numbers of tasks. SPARC consistently maintains higher accuracy than ER and DER++ as the number of tasks increa…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

99 extracted references · 47 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Sparsity and Heterogeneous Dropout for Continual Learning in the Null Space of Neural Activations

    Abbasi, A., Nooralinejad, P., Braverman, V., Pirsiavash, H., and Kolouri, S. Sparsity and Heterogeneous Dropout for Continual Learning in the Null Space of Neural Activations . In Proceedings of The 1st Conference on Lifelong Learning Agents , pp.\ 617--628. PMLR, November 2022. ISSN: 2640-3498

    2022

  3. [3]

    Expert gate: Lifelong learning with a network of experts

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

    2017

  4. [4]

    Online continual learning with maximal interfered retrieval

    Aljundi, R., Belilovsky, E., Tuytelaars, T., Charlin, L., Caccia, M., Lin, M., and Page-Caccia, L. Online continual learning with maximal interfered retrieval. Advances in neural information processing systems, 32, 2019

    2019

  5. [5]

    Learning fast, learning slow: A general continual learning method based on complementary learning system

    Arani, E., Sarfraz, F., and Zonooz, B. Learning fast, learning slow: A general continual learning method based on complementary learning system. In International Conference on Learning Representations, 2022

    2022

  6. [6]

    and Laughlin, S

    Attwell, D. and Laughlin, S. B. An energy budget for signaling in the grey matter of the brain. Journal of Cerebral Blood Flow & Metabolism, 21 0 (10): 0 1133--1145, 2001

    2001

  7. [7]

    C., Dale, D., Dong, N., Duquenne, P.-A., Elsahar, H., Gong, H., Heffernan, K., Hoffman, J., et al

    Barrault, L., Chung, Y.-A., Meglioli, M. C., Dale, D., Dong, N., Duquenne, P.-A., Elsahar, H., Gong, H., Heffernan, K., Hoffman, J., et al. Seamlessm4t-massively multilingual & multimodal machine translation. arXiv preprint arXiv:2308.11596, 2023

    arXiv 2023

  8. [8]

    Barth, A. L. and Poulet, J. F. Experimental evidence for sparse firing in the neocortex. Trends in neurosciences, 35 0 (6): 0 345--355, 2012

    2012

Show all 99 references
  1. [9]

    Mitigating interference in the knowledge continuum through attention-guided incremental learning

    Bhat, P., Renjith, B., Arani, E., and Zonooz, B. Mitigating interference in the knowledge continuum through attention-guided incremental learning. arXiv preprint arXiv:2405.13978, 2024 a

    2024 arXiv

  2. [10]

    IMEX - Reg : Implicit - Explicit Regularization in the Function Space for Continual Learning , April 2024 b

    Bhat, P., Renjith, B., Arani, E., and Zonooz, B. IMEX - Reg : Implicit - Explicit Regularization in the Function Space for Continual Learning , April 2024 b . arXiv:2404.18161 [cs]

    2024 arXiv

  3. [11]

    S., Zonooz, B., and Arani, E

    Bhat, P. S., Zonooz, B., and Arani, E. Consistency is the key to further mitigating catastrophic forgetting in continual learning. In Conference on Lifelong Learning Agents, pp.\ 1195--1212. PMLR, 2022 a

    2022

  4. [12]

    S., Zonooz, B., and Arani, E

    Bhat, P. S., Zonooz, B., and Arani, E. Task agnostic representation consolidation: a self-supervised based continual learning approach. In Chandar, S., Pascanu, R., and Precup, D. (eds.), Proceedings of The 1st Conference on Lifelong Learning Agents, volume 199 of Proceedings ...

    2022

  5. [13]

    S., Zonooz, B., and Arani, E

    Bhat, P. S., Zonooz, B., and Arani, E. Task-aware information routing from common representation space in lifelong learning. In The Eleventh International Conference on Learning Representations, 2023

    2023

  6. [14]

    On the effectiveness of lipschitz-driven rehearsal in continual learning

    Bonicelli, L., Boschini, M., Porrello, A., Spampinato, C., and Calderara, S. On the effectiveness of lipschitz-driven rehearsal in continual learning. Advances in Neural Information Processing Systems, 35: 0 31886--31901, 2022

    2022

  7. [15]

    Brown, T. B. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020

    2005 arXiv

  8. [16]

    Dark experience for general continual learning: a strong, simple baseline

    Buzzega, P., Boschini, M., Porrello, A., Abati, D., and Calderara, S. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33: 0 15920--15930, 2020

    2020

  9. [17]

    New insights on reducing abrupt representation change in online continual learning

    Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., and Belilovsky, E. New insights on reducing abrupt representation change in online continual learning. In International Conference on Learning Representations, 2021 a

    2021

  10. [18]

    Reducing representation drift in online continual learning

    Caccia, L., Aljundi, R., Asadi, N., Tuytelaars, T., Pineau, J., and Belilovsky, E. Reducing representation drift in online continual learning. arXiv preprint arXiv:2104.05025, 2021 b

    2021 arXiv

  11. [19]

    Co2l: Contrastive continual learning

    Cha, H., Lee, J., and Shin, J. Co2l: Contrastive continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 9516--9525, 2021

    2021

  12. [20]

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

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

    2018

  13. [21]

    Xception: Deep Learning with Depthwise Separable Convolutions , April 2017

    Chollet, F. Xception: Deep Learning with Depthwise Separable Convolutions , April 2017. arXiv:1610.02357 [cs]

    2017 arXiv

  14. [22]

    A continual learning survey: Defying forgetting in classification tasks

    De Lange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., and Tuytelaars, T. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence, 44 0 (7): 0 3366--3385, 2021

    2021

  15. [23]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  16. [24]

    Uncertainty-guided continual learning with bayesian neural networks

    Ebrahimi, S., Elhoseiny, M., Darrell, T., and Rohrbach, M. Uncertainty-guided continual learning with bayesian neural networks. In International Conference on Learning Representations, 2020

    2020

  17. [25]

    Continual learning via neural pruning

    Golkar, S., Kagan, M., and Cho, K. Continual learning via neural pruning. arXiv preprint arXiv:1903.04476, 2019

    1903 arXiv

  18. [26]

    Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning

    Goswami, D., Liu, Y., Twardowski, B., and van de Weijer, J. Fecam: Exploiting the heterogeneity of class distributions in exemplar-free continual learning. Advances in Neural Information Processing Systems, 36, 2024 a

    2024

  19. [27]

    Resurrecting old classes with new data for exemplar-free continual learning

    Goswami, D., Soutif-Cormerais, A., Liu, Y., Kamath, S., Twardowski, B., van de Weijer, J., et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 28525--2853...

    2024

  20. [28]

    Network decoupling: From regular to depthwise separable convolutions

    Guo, J., Li, Y., Lin, W., Chen, Y., and Li, J. Network decoupling: From regular to depthwise separable convolutions. In BMVC, 2018

    2018

  21. [29]

    Depthwise Convolution is All You Need for Learning Multiple Visual Domains , February 2019

    Guo, Y., Li, Y., Feris, R., Wang, L., and Rosing, T. Depthwise Convolution is All You Need for Learning Multiple Visual Domains , February 2019. arXiv:1902.00927 [cs]

    2019 arXiv

  22. [30]

    Gurbuz, M. B. and Dovrolis, C. Nispa: Neuro-inspired stability-plasticity adaptation for continual learning in sparse networks. In International Conference on Machine Learning, pp.\ 8157--8174. PMLR, 2022

    2022

  23. [31]

    A., and Pascanu, R

    Hadsell, R., Rao, D., Rusu, A. A., and Pascanu, R. Embracing change: Continual learning in deep neural networks. Trends in cognitive sciences, 24 0 (12): 0 1028--1040, 2020

    2020

  24. [32]

    Deep residual learning for image recognition

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

    2016

  25. [33]

    Pyramidal cell communication within local networks in layer 2/3 of rat neocortex

    Holmgren, C., Harkany, T., Svennenfors, B., and Zilberter, Y. Pyramidal cell communication within local networks in layer 2/3 of rat neocortex. The Journal of physiology, 551 0 (1): 0 139--153, 2003

    2003

  26. [34]

    C., Wang, Z., and Lin, D

    Hou, S., Pan, X., Loy, C. C., Wang, Z., and Lin, D. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  27. [35]

    Parameter-efficient transfer learning for nlp

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pp.\ 2790--2799. PMLR, 2019

    2019

  28. [36]

    Inverted residuals and linear bottlenecks: Mobile networks for classification, detection and segmentation

    Howard, A., Zhmoginov, A., Chen, L.-C., Sandler, M., and Zhu, M. Inverted residuals and linear bottlenecks: Mobile networks for classification, detection and segmentation. In Proc. CVPR, pp.\ 4510--4520, 2018

    2018

  29. [37]

    G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., and Adam, H

    Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., and Adam, H. MobileNets : Efficient Convolutional Neural Networks for Mobile Vision Applications , April 2017. arXiv:1704.04861 [cs]

    2017 arXiv

  30. [38]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  31. [39]

    Compacting, picking and growing for unforgetting continual learning

    Hung, C.-Y., Tu, C.-H., Wu, C.-E., Chen, C.-H., Chan, Y.-M., and Chen, C.-S. Compacting, picking and growing for unforgetting continual learning. Advances in Neural Information Processing Systems, 32, 2019 a

    2019

  32. [40]

    C., Lee, J.-H., Wan, T

    Hung, S. C., Lee, J.-H., Wan, T. S., Chen, C.-H., Chan, Y.-M., and Chen, C.-S. Increasingly packing multiple facial-informatics modules in a unified deep-learning model via lifelong learning. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pp.\ ...

    2019

  33. [41]

    Birt: bio-inspired replay in vision transformers for continual learning

    Jeeveswaran, K., Bhat, P., Zonooz, B., and Arani, E. Birt: bio-inspired replay in vision transformers for continual learning. In Proceedings of the 40th International Conference on Machine Learning, pp.\ 14817--14835, 2023

    2023

  34. [42]

    Kang, H., Mina, R. J. L., Madjid, S. R. H., Yoon, J., Hasegawa-Johnson, M., Hwang, S. J., and Yoo, C. D. Forget-free continual learning with winning subnetworks. In International Conference on Machine Learning, pp.\ 10734--10750. PMLR, 2022

    2022

  35. [43]

    C., Lo, W.-Y., et al

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023

    2023 arXiv

  36. [44]

    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. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114 0 (13): 0 3521--3526, 2017

    2017

  37. [45]

    Learning multiple layers of features from tiny images

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

    2009

  38. [46]

    Krizhevsky, A., Sutskever, I., and Hinton, G. E. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60 0 (6): 0 84--90, May 2012. ISSN 0001-0782, 1557-7317. doi:10.1145/3065386

    2012 doi

  39. [47]

    P., Chakravarthi Raja, S., Cheney, N., Clune, J., et al

    Kudithipudi, D., Aguilar-Simon, M., Babb, J., Bazhenov, M., Blackiston, D., Bongard, J., Brna, A. P., Chakravarthi Raja, S., Cheney, N., Clune, J., et al. Biological underpinnings for lifelong learning machines. Nature Machine Intelligence, 4 0 (3): 0 196--210, 2022

    2022

  40. [48]

    Investigating multilingual nmt representations at scale

    Kudugunta, S., Bapna, A., Caswell, I., and Firat, O. Investigating multilingual nmt representations at scale. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EM...

    2019

  41. [49]

    and Yang, X

    Le, Y. and Yang, X. Tiny imagenet visual recognition challenge. CS 231N, 7 0 (7): 0 3, 2015

    2015

  42. [50]

    Learning from students: Online contrastive distillation network for general continual learning

    Li, J., Ji, Z., Wang, G., Wang, Q., and Gao, F. Learning from students: Online contrastive distillation network for general continual learning. In Proc. 31st Int. Joint Conf. Artif. Intell., pp.\ 3215--3221, 2022

    2022

  43. [51]

    and Hoiem, D

    Li, Z. and Hoiem, D. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40 0 (12): 0 2935--2947, 2017

    2017

  44. [52]

    and Li, W.-J

    Liang, Y.-S. and Li, W.-J. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 23638--23647, 2024

    2024

  45. [53]

    Self-improving reactive agents based on reinforcement learning, planning and teaching

    Lin, L.-J. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine Learning, 8 0 (3): 0 293--321, May 1992. ISSN 1573-0565. doi:10.1007/BF00992699

    1992 doi

  46. [54]

    F., Cheng, K.-T., and Chen, M.-H

    Liu, S.-Y., Wang, C.-Y., Yin, H., Molchanov, P., Wang, Y.-C. F., Cheng, K.-T., and Chen, M.-H. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024

    2024 arXiv

  47. [55]

    More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning

    Liu, Y., Parisot, S., Slabaugh, G., Jia, X., Leonardis, A., and Tuytelaars, T. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceeding...

    2020

  48. [56]

    Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning

    Mai, Z., Li, R., Kim, H., and Sanner, S. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3589--3599, 2021

    2021

  49. [57]

    and Lazebnik, S

    Mallya, A. and Lazebnik, S. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.\ 7765--7773, 2018

    2018

  50. [58]

    W., Abdellah, M., Sanchez, C

    Markram, H., Muller, E., Ramaswamy, S., Reimann, M. W., Abdellah, M., Sanchez, C. A., Ailamaki, A., Alonso-Nanclares, L., Antille, N., Arsever, S., et al. Reconstruction and simulation of neocortical microcircuitry. Cell, 163 0 (2): 0 456--492, 2015

    2015

  51. [59]

    D., and Van De Weijer, J

    Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A. D., and Van De Weijer, J. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (5): 0 5513--5533, 2022

    2022

  52. [60]

    L., McNaughton, B

    McClelland, J. L., McNaughton, B. L., and O'Reilly, R. C. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review, 102 0 (3): 0 419, 1995

    1995

  53. [61]

    and Cohen, N

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

    1989

  54. [62]

    The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects

    Mermillod, M., Bugaiska, A., and Bonin, P. The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects. Frontiers in psychology, 4: 0 504, 08 2013. doi:10.3389/fpsyg.2013.00504

    2013

  55. [63]

    I., Chaudhry, A., Yin, D., Hu, H., Pascanu, R., Gorur, D., and Farajtabar, M

    Mirzadeh, S. I., Chaudhry, A., Yin, D., Hu, H., Pascanu, R., Gorur, D., and Farajtabar, M. Wide neural networks forget less catastrophically. In International Conference on Machine Learning, pp.\ 15699--15717. PMLR, 2022

    2022

  56. [64]

    Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift

    Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., and Snoek, J. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information processing systems, 32, 2019

    2019

  57. [65]

    I., Kemker, R., Part, J

    Parisi, G. I., Kemker, R., Part, J. L., Kanan, C., and Wermter, S. Continual lifelong learning with neural networks: A review. Neural Networks, 113: 0 54--71, 2019

    2019

  58. [66]

    Park, D., Hong, S., Han, B., and Lee, K. M. Continual learning by asymmetric loss approximation with single-side overestimation. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 3335--3344, 2019

    2019

  59. [67]

    Fetril: Feature translation for exemplar-free class-incremental learning

    Petit, G., Popescu, A., Schindler, H., Picard, D., and Delezoide, B. Fetril: Feature translation for exemplar-free class-incremental learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp.\ 3911--3920, 2023

    2023

  60. [68]

    Dualnet: Continual learning, fast and slow

    Pham, Q., Liu, C., and Hoi, S. Dualnet: Continual learning, fast and slow. Advances in Neural Information Processing Systems, 34: 0 16131--16144, 2021 a

    2021

  61. [69]

    Continual normalization: Rethinking batch normalization for online continual learning

    Pham, Q., Liu, C., and Steven, H. Continual normalization: Rethinking batch normalization for online continual learning. In International Conference on Learning Representations, 2021 b

    2021

  62. [70]

    Connectionist models of recognition memory: constraints imposed by learning and forgetting functions

    Ratcliff, R. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological review, 97 0 (2): 0 285, 1990

    1990

  63. [71]

    Online structured laplace approximations for overcoming catastrophic forgetting

    Ritter, H., Botev, A., and Barber, D. Online structured laplace approximations for overcoming catastrophic forgetting. Advances in Neural Information Processing Systems, 31, 2018

    2018

  64. [72]

    A., Rabinowitz, N

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

    2016 arXiv

  65. [73]

    Gradient Projection Memory for Continual Learning , March 2021

    Saha, G., Garg, I., and Roy, K. Gradient Projection Memory for Continual Learning , March 2021. arXiv:2103.09762 [cs]

    2021 arXiv

  66. [74]

    MobileNetV2 : Inverted Residuals and Linear Bottlenecks , March 2019

    Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and Chen, L.-C. MobileNetV2 : Inverted Residuals and Linear Bottlenecks , March 2019. arXiv:1801.04381 [cs]

    2019 arXiv

  67. [75]

    Synergy between synaptic consolidation and experience replay for general continual learning

    Sarfraz, F., Arani, E., and Zonooz, B. Synergy between synaptic consolidation and experience replay for general continual learning. In Chandar, S., Pascanu, R., and Precup, D. (eds.), Proceedings of The 1st Conference on Lifelong Learning Agents, volume 199 of Proceedings of M...

    2022

  68. [76]

    Sparse Coding in a Dual Memory System for Lifelong Learning , December 2022 b

    Sarfraz, F., Arani, E., and Zonooz, B. Sparse Coding in a Dual Memory System for Lifelong Learning , December 2022 b . arXiv:2301.05058 [cs]

    2022 arXiv

  69. [77]

    and Ye, M

    Shi, W. and Ye, M. Prototype reminiscence and augmented asymmetric knowledge aggregation for non-exemplar class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 1772--1781, 2023

    2023

  70. [78]

    and Mallat, S

    Sifre, L. and Mallat, S. Rigid- Motion Scattering for Texture Classification , March 2014. arXiv:1403.1687 [cs]

    2014 arXiv

  71. [79]

    S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., and Kira, Z

    Smith, J. S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., and Kira, Z. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Visio...

    2023

  72. [80]

    Decoupling learning and remembering: A bilevel memory framework with knowledge projection for task-incremental learning

    Sun, W., Li, Q., Zhang, J., Wang, W., and Geng, Y.-a. Decoupling learning and remembering: A bilevel memory framework with knowledge projection for task-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20186--20195, 2023

    2023

  73. [81]

    Going Deeper With Convolutions

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going Deeper With Convolutions . pp.\ 1--9, 2015

    2015

  74. [82]

    Gcr: Gradient coreset based replay buffer selection for continual learning

    Tiwari, R., Killamsetty, K., Iyer, R., and Shenoy, P. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 99--108, 2022

    2022

  75. [83]

    Trire: A multi-mechanism learning paradigm for continual knowledge retention and promotion

    Vijayan, P., Bhat, P., Arani, E., and Zonooz, B. Trire: A multi-mechanism learning paradigm for continual knowledge retention and promotion. arXiv preprint arXiv:2310.08217, 2023

    2023 arXiv

  76. [84]

    A comprehensive survey of continual learning: Theory, method and application

    Wang, L., Zhang, X., Su, H., and Zhu, J. A comprehensive survey of continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487, 2023

    2023 arXiv

  77. [85]

    S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning

    Wang, Y., Huang, Z., and Hong, X. S-prompts learning with pre-trained transformers: An occam’s razor for domain incremental learning. Advances in Neural Information Processing Systems, 35: 0 5682--5695, 2022 a

    2022

  78. [86]

    Sparcl: Sparse continual learning on the edge

    Wang, Z., Zhan, Z., Gong, Y., Yuan, G., Niu, W., Jian, T., Ren, B., Ioannidis, S., Wang, Y., and Dy, J. Sparcl: Sparse continual learning on the edge. Advances in Neural Information Processing Systems, 35: 0 20366--20380, 2022 b

    2022

  79. [87]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.-Y., Ren, X., Su, G., Perot, V., Dy, J., et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pp.\ 631--648. Springer, 2022 c

    2022

  80. [88]

    Learning to prompt for continual learning

    Wang, Z., Zhang, Z., Lee, C.-Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., and Pfister, T. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 139--149, 2022 d

    2022

  81. [89]

    Large scale incremental learning

    Wu, Y., Chen, Y., Wang, L., Ye, Y., Liu, Z., Guo, Y., and Fu, Y. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 374--382, 2019

    2019

  82. [90]

    Yoon, J., Yang, E., Lee, J., and Hwang, S. J. Lifelong learning with dynamically expandable networks. In International Conference on Learning Representations, 2018

    2018

  83. [91]

    Semantic drift compensation for class-incremental learning

    Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., and van de Weijer, J. Semantic drift compensation for class-incremental learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 6980--6989, 2020. doi:10.1109/CVPR4...

    2020

  84. [92]

    B., Ravfogel, S., and Goldberg, Y

    Zaken, E. B., Ravfogel, S., and Goldberg, Y. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021

    2021

  85. [93]

    Continual learning through synaptic intelligence

    Zenke, F., Poole, B., and Ganguli, S. Continual learning through synaptic intelligence. In International Conference on Machine Learning, pp.\ 3987--3995. PMLR, 2017

    2017

  86. [94]

    Fine-grained knowledge selection and restoration for non-exemplar class incremental learning

    Zhai, J.-T., Liu, X., Yu, L., and Cheng, M.-M. Fine-grained knowledge selection and restoration for non-exemplar class incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 6971--6978, 2024

    2024

  87. [95]

    Maintaining discrimination and fairness in class incremental learning

    Zhao, B., Xiao, X., Gan, G., Zhang, B., and Xia, S.-T. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 13208--13217, 2020

    2020

  88. [96]

    Expandable subspace ensemble for pre-trained model-based class-incremental learning

    Zhou, D.-W., Sun, H.-L., Ye, H.-J., and Zhan, D.-C. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 23554--23564, 2024

    2024

  89. [97]

    Class-incremental learning via dual augmentation

    Zhu, F., Cheng, Z., Zhang, X.-Y., and Liu, C.-l. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34: 0 14306--14318, 2021 a

    2021

  90. [98]

    Prototype augmentation and self-supervision for incremental learning

    Zhu, F., Zhang, X.-Y., Wang, C., Yin, F., and Liu, C.-L. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 5871--5880, 2021 b

    2021

  91. [99]

    Self-sustaining representation expansion for non-exemplar class-incremental learning

    Zhu, K., Zhai, W., Cao, Y., Luo, J., and Zha, Z.-J. Self-sustaining representation expansion for non-exemplar class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9296--9305, 2022

    2022

Pith tools

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