Pith. sign in

REVIEW 4 major objections 5 minor 50 references

SLAM: Structured and Localized Analytic Manifold Adaptation for Forgetting-Immune and Domain-Robust Lifelong VPR

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that analytic class-incremental learning can be made robust to seasonal domain shifts by treating its recursive updates as EKF covariance propagation, yielding a memory-free, forgetting-immune lifelong visual place recogni

desk verdict Good core recipe, but the headline domain-robustness mechanism is never tested and the EKF isomorphism is asserted, so the paper needs a major revision and a narrower claim set. read the letter →

arxiv 2607.04764 v2 pith:BEMS2E3L submitted 2026-07-06 cs.RO cs.CV

Kenta Tsukahara , Kanji Tanaka , Rai Hisada This is my paper
classification cs.ROcs.CV
keywords VisualPlaceRecognitionAnalyticClass-IncrementalLearningCatastrophicForgettingDomainAdaptationFeatureDisentanglementExtendedKalmanFilterLifelongRobotics
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 is trying to establish that Analytic Class-Incremental Learning—a closed-form, backpropagation-free way to add classes by updating an autocorrelation matrix—can be made safe for lifelong visual place recognition under changing seasons and lighting. It argues that the recursive ACIL update is algebraically the same as Extended Kalman Filter covariance propagation, and that this isomorphism lets robust control-theoretic mechanisms be imported into the analytic update. On a 10-task benchmark built from the non-stationary NCLT campus dataset, the full SLAM pipeline reaches 27.7% final all-class accuracy (29.0% for a variant) while keeping O(1) memory and, by construction, no forgetting of past classes. A sympathetic reader would care because it points toward lifelong robot perception that respects privacy and memory bounds.

What carries the argument

The load-bearing object is the recursive Woodbury update for the inverse-precision matrix K and cross-covariance Q, with the classifier weight W=KQ. The paper's key move is to read this as an EKF covariance-propagation loop, which justifies adding (1) a dynamic temperature-scaled GMM that partitions the feature space into topological components so non-linearities stay local, (2) an unscented perturbation that creates symmetric soft-logit variations around the teacher targets, and (3) an H-infinity attenuation factor that shrinks the covariance update under a worst-case noise bound. ACIL-D is the canonical invariant feature manifold where the autocorrelation states and decision hyperplanes st

What would settle it

Replace D-DA with a matched data-augmentation/noise-injection scheme that does not separate style from semantics, keeping all other SLAM components fixed; if final all-class accuracy stays at 27.7% or improves, the disentanglement and directional-manifold claims are unsupported. Alternatively, run the full SLAM pipeline with D-DA applied at test time to raw features; if accuracy does not improve over the raw-feature evaluation, the adapter's test-time role is unverified.

Watch

Extended reading notes

Core claim

The paper's central claim is that ACIL's vulnerability to domain shift is the same fragility that EKF-style recursive least squares shows under non-Gaussian noise, and that an algebraic isomorphism between the Woodbury update in ACIL and EKF covariance propagation makes this vulnerability fixable with control-theoretic tools. On that basis, SLAM couples a disentangled domain adapter (invariant semantics vs style) with three analytic regularizers: temperature-scaled GMM for localizing non-linearities, unscented perturbed propagation, and an H-infinity minimax bound. The reported outcome is complete forgetting immunity at O(1) memory: 27.7% final all-class accuracy for full SLAM and 29.0% for

Load-bearing premise

The claim that D-DA provides domain robustness rests on training-time feature conditioning; at test time the raw, unaligned features are used, so the style-truncation mechanism is never actually exercised and could be doing nothing beyond data augmentation.

Editorial extensions

If this is right

  • If the isomorphism and empirical results hold, lifelong VPR can run on a frozen backbone with constant memory, no replay buffer, and no raw image storage, easing privacy constraints in deployed robots.
  • Old decision boundaries are mathematically locked, so the 'forgetting immunity' is exact for the analytic classifier rather than approximate, unlike replay or distillation baselines.
  • The EKF mapping gives a principled way to import established robust-estimation tools (unscented transforms, H-infinity bounds, mixture filtering) into any analytic continual learner, not just place recognition.
  • The reported margins over replay-based, prompt-based, and LoRA baselines indicate the approach may be significantly stronger under severe seasonal shift, if the evaluation protocol holds.

Reading between the lines

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

  • The D-DA adapter is trained with gradient-based Adam and never used at test time; its claimed 'style truncation' is therefore not directly exercised. A plausible reading is that D-DA acts as data augmentation for the analytic update, and the disentanglement itself is unverified.
  • Because the adapter is gradient-trained, 'complete forgetting immunity' applies only to the analytic classifier state; the adapter's parameters can drift across tasks, and its effect on future tasks is not controlled by the Woodbury update.
  • The EKF isomorphism opens a two-way street: adaptive noise-covariance estimation, outlier gating, or measurement validation from state estimation could be imported into ACIL; the paper does not explore these.
  • One testable extension: apply SLAM updates to other frozen-backbone continual classification streams (e.g., object or scene recognition across domains) to see whether the observed robustness transfers beyond GPS-quantized place labels.
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 / 5 minor

Summary. The paper proposes SLAM, a framework for lifelong visual place recognition built on Analytic Class-Incremental Learning (ACIL). The authors introduce disentangled domain alignment (D-DA) with a claimed inference-time style-truncation mechanism, and augment the ACIL recursion with GMM-based temperature scaling, unscented perturbation, and H-infinity robust terms, motivated by an alleged algebraic isomorphism between ACIL updates and EKF covariance propagation. On a 10-task NCLT benchmark, the paper reports 27.7% final all-class accuracy for the full SLAM framework and 29.0% for a U+H variant, while claiming O(1) memory and complete forgetting immunity.

Significance. If substantiated, the paper would offer a memory-free, privacy-preserving lifelong VPR classifier that is both forgetting-immune and robust to domain shift. The closed-form update structure and the extensive ablation over components are useful; the idea of linking ACIL to recursive least squares / Kalman-type updates is suggestive. However, the central domain-robustness mechanism is not evaluated at inference time, the ACIL-EKF mapping is asserted rather than proven, and the modified covariance update is not shown to preserve the ACIL guarantee. As presented, the empirical and theoretical support for the headline claims is incomplete.

major comments (4)
  1. [§IV.C and Table II] The abstract and Section I claim that D-DA performs inference-time style truncation and directional projection into ACIL-D. However, §IV.C states that 'the DA module is employed exclusively during the training phase... all evaluations are conducted on raw projected representations without adapter inference.' Therefore the reported 27.7% (SLAM full) and 29.0% (U+H) results measure a training-time feature-conditioning effect, not the proposed inference-time purification mechanism. Moreover, Analytic A-DDA outperforms Vanilla DA only marginally (27.7 vs. 27.4; U+H 29.0 vs. 28.9), so even as a training-time augmentation its benefit over simple moment matching is small. This is a load-bearing gap: without test-time adapter evaluation, the paper's central domain-robustness claim is untested. Please add experiments that pass raw test features through the style-truncated P_geom and compare, or r
  2. [§III.D, Eq. (16)] Eq. (16) inserts η_combined into the Woodbury K update while Eq. (15) leaves the Q update unchanged. ACIL's 'absolute memorization / forgetting immunity' property from [5] relies on the exact recursive least-squares solution. No proof is given that this scaled update preserves that property or that it minimizes a well-defined objective. With W=KQ, scaling K differently from Q means the resulting weight matrix is not the standard ACIL solution, so the claimed forgetting-immunity guarantee is not automatic. Please provide a theorem showing the modified recursion is still exact, or remove/modify the guarantee.
  3. [§III.A–III.D] The paper repeatedly invokes an 'algebraic isomorphism between recursive ACIL updates and EKF covariance propagation,' but the isomorphism is never stated formally. The equations shown are the standard Woodbury identity for regularized least squares; no mapping of state vector, measurement model, process/measurement noise, or Jacobians is given. Without a precise theorem and proof, the control-theoretic robustification is heuristic, not derived. This is a central claim of the paper and needs a formal treatment.
  4. [§IV.B and Eqs. (5), (10), (12), (13)] Many hyperparameters (α, λreg, κ, T, αukf, γhinf, GMM scaling constants, JL dimension, bottleneck size, loss weights, learning rate) are tuned on the same NCLT benchmark, and no sensitivity analysis or held-out domain is reported. Given that the headline contribution is domain robustness, the lack of validation on data not used for hyperparameter selection weakens the 'robustness' claim. Please include a sensitivity study or a second dataset/domain split.
minor comments (5)
  1. [Abstract and §I] The abstract reports 27.7% / 29.0% final accuracy, but the introduction says 'up to 28.0% all-class accuracy.' The numbers should be reconciled.
  2. [References] References [15], [26], and [35] contain the placeholder 'To be updated' and are incomplete; please replace with full citations.
  3. [§III.A–III.B] The D-DA description in the text mentions f_style and f_static and 'style truncation,' but the equations (1)–(6) only present moment matching and P_geom. A formal definition of f_style, f_static, and the truncation operation is missing.
  4. [Tables I and II] The 'Seen' and 'All' metrics are not defined in the captions. Also, baseline names are inconsistent: DER/ER++ in Table I vs. ER-BoT in the appendix.
  5. [Eq. (11)] The notation σ(S) is used for softmax while σ is also used for variance elsewhere; please disambiguate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation chain: no equation reduces by construction to a fitted input or a self-citation; the main weakness is an evaluation/validation gap, not circularity.

full rationale

Walking the derivation chain: ACIL updates (Eqs. 7-17) are closed-form Woodbury recursions whose targets Y_all are soft labels generated by an analytic teacher (Eqs. 7-9) from the same task data; this is a standard training loop, not a case where the predicted accuracy is an input. The hyperparameters (alpha, T, gamma, lambda, etc.) are chosen constants, not fitted values that are then reported as predictions, so no fitted-input-called-prediction step is exhibited. The ACIL-D/EKF 'isomorphism' is asserted and used as a naming/organizing analogy; no equation is shown to be identical to an input by construction, and the robustified update rules are heuristics rather than a re-derivation of the claimed output. There is no self-citation chain: ACIL [5], CL-SLAM [30], and baselines are external works, and forgetting immunity is imported from the external ACIL theorem, which is independent support even if the modifications may exceed its assumptions. The only significant issue is Section IV.C: 'the DA module is employed exclusively during the training phase to condition the feature spaces, whereas all evaluations are conducted on raw projected representations without adapter inference.' This means the inference-time style-truncation/directional projection mechanism is not validated by Table II, and A-DDA results measure a training-time conditioning effect. That is a validation gap and a correctness/rigor risk, but it is not circularity that makes the result equal to its input, so it does not raise the circularity score.

Assumptions & free parameters 12 free parameters · 6 assumptions · 2 invented entities

The central performance claims rest on a large set of hand-set hyperparameters and on the unproven transfer of a training-only DA module to test-time raw features. The core ACIL least-squares recursion is modified without a consistent objective, and the new conceptual entities (ACIL-D, style vector) have no independent falsifiable handle.

free parameters (12)
  • residual smoothing factor α = 0.01
    Eq. (5) controls the geometric manifold correction strength; chosen by hand.
  • Tikhonov regularization λreg = 0.1
    Eqs. (5), (7); controls matrix inversion conditioning.
  • teacher logit scale κ = 16.0
    Eq. (9); set manually for the specialist teacher output.
  • distillation temperature T = 2.0
    Eq. (10); scaling for logit perturbation and distillation.
  • unscented perturbation strength αukf = 0.6
    Eq. (10); amplitude of symmetric soft-logit variations.
  • H∞ attenuation factor γhinf = 5.0
    Eq. (13); defines η_hinf = 1 − γ−2 = 0.96.
  • GMM scaling constants 0.25 and 0.5 = 0.25, 0.5
    Eq. (12); constants in the heuristic η_gmm formula with no derivation.
  • JL compression dimension d = 384
    §IV-B; reduces D=1024 features to 384 via Johnson-Lindenstrauss projection.
  • bottleneck dimension dbottleneck = 256
    §IV-B; width of the domain adapter bottleneck.
  • GMM cluster count K = 8
    §IV-B; topological partition components initialized via K-Means.
  • DA loss weights λdis and λid = 1.0, 0.1
    §IV-B; disentanglement/identity loss weights, though no loss equations are given.
  • Adam learning rate = 1e-3
    §IV-B; used for the adapter optimizer.
assumptions (6)
  • standard math Woodbury matrix identity / recursive least squares is valid for the K update (Eq. 16).
    Used without proof to update the inverse covariance; standard linear algebra.
  • domain assumption DINOv2 frozen features are fixed and sufficiently discriminative for 100-class VPR under the analytically learned classifier.
    §IV-A: features are pre-extracted with a frozen DINOv2 backbone; the entire pipeline depends on this external representation.
  • ad hoc to paper ACIL's 'absolute memorization / forgetting immunity' property from [5] remains valid after the η_combined-modified K update (Eq. 16).
    The paper modifies K but not Q, and no optimization objective is given for the modified recursion; forgetting is never measured.
  • ad hoc to paper A DA module used only during training is sufficient to align test-time raw features into ACIL-D.
    §IV-C states evaluations are on raw projected representations, contradicting the test-time style-truncation description in the abstract.
  • ad hoc to paper The GMM responsibility / temperature scaling formula (Eq. 12) captures topological nonlinearity.
    No derivation is given; constants 0.25 and 0.5 are chosen to improve benchmark performance.
  • domain assumption GPS-discretized pseudo-labels define a valid class-incremental VPR task.
    §I and §IV-A: the 100-class label space is derived from spatial quantization of GPS groundtruth.
invented entities (2)
  • ACIL-D (Analytic Class-Incremental Learning Domain)
    purpose: Canonical invariant feature manifold where autocorrelation states and decision hyperplanes remain locked; used as the conceptual target of D-DA.
    Defined only verbally; no metric verifies that features lie in ACIL-D, and test-time evaluation does not use the proposed projection.
  • style vector fstyle / static semantic stream fstatic
    purpose: Disentangle environment-specific variance from invariant semantics in D-DA.
    No disentanglement loss, style-truncation equation, or direct measurement of the style vector is provided; the paper uses the DA module only as training-time augmentation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SLAM: Structured and Localized Analytic Manifold Adaptation for Forgetting-Immune and Domain-Robust Lifelong VPR." pith.science (2026). https://pith.science/paper/BEMS2E3L

@misc{pith2026260704764,
  author       = {Pith},
  title        = {Pith review of: SLAM: Structured and Localized Analytic Manifold Adaptation for Forgetting-Immune and Domain-Robust Lifelong VPR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEMS2E3L}},
  note         = {Machine review of arXiv:2607.04764}
}
abstract

Visual Place Recognition (VPR) under long-term operation is essential for autonomous mobile robots. While Analytic Class-Incremental Learning (ACIL) provides memory-free ($O(1)$) task adaptation with exact forgetting immunity, applying it to lifelong VPR suffers from extreme vulnerability to non-linear domain shifts induced by environmental variations. In this work, we introduce the concept of the \textbf{ACIL-Domain (ACIL-D)}---a canonical invariant feature manifold where autocorrelation states remain locked. We resolve the domain vulnerability via Disentangled Domain Alignment (D-DA), which decouples latent features into invariant semantics within ACIL-D and variant style vectors for directional projection. Furthermore, by uncovering an algebraic isomorphism between recursive ACIL updates and Extended Kalman Filter (EKF) covariance propagation, we establish a control-theoretic framework designated as \textbf{SLAM} (\textbf{S}tructured and \textbf{L}ocalized \textbf{A}nalytic \textbf{M}anifold adaptation). SLAM integrates dynamic temperature-scaled Gaussian Mixture Models (GMM) to isolate topological non-linearities, Unscented perturbed propagation to dampen feature variations, and minimax $H_{\infty}$-robust criteria to bound worst-case noise accumulation. Empirical evaluations on the non-stationary NCLT dataset demonstrate that our proposed framework substantially outperforms existing baselines, achieving a final all-class accuracy of 27.7\% with the full SLAM framework (and up to 29.0\% with the U+H variant) while guaranteeing complete forgetting immunity.

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 9 linked inside Pith

  1. [5]

    Acil: Analytic class-incremental learning with absolute memorization and privacy protection

    Huiping Zhuang, Zhenyu Weng, Hongxin Wei, Lihua Xie, Kar-Ann Toh, and Zhiping Lin. Acil: Analytic class-incremental learning with absolute memorization and privacy protection. InProceedings of the European Conference on Computer Vision (ECCV), pages 105–121, 2022

    2022

  2. [1]

    Leonard, David Cox, Peter Corke, and Michael Milford

    Stephanie Lowry, Niko S ¨underhauf, Paul Newman, John J. Leonard, David Cox, Peter Corke, and Michael Milford. Visual place recognition: A survey.IEEE Transactions on Robotics, 32(1):1–19, 2016

    2016

  3. [2]

    Planet - photo geolocation with combinatorial partitioning

    Tobias Weyand, Ilya Kostrikov, and James Philbin. Planet - photo geolocation with combinatorial partitioning. InProceedings of the European Conference on Computer Vision (ECCV), pages 37–55, 2016

    2016

  4. [3]

    Catastrophic interference in connectionist networks: The sequential learning problem.Psychology of Learning and Motivation, 24:109–165, 1989

    Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem.Psychology of Learning and Motivation, 24:109–165, 1989

    1989

  5. [4]

    Lampinen, Ekin D

    David Rolnick, Arun Mallya, Andrew K. Lampinen, Ekin D. Cubuk, Eileen T. Zhang, Stefano Soatto, and Roger Grosse. Experience replay for continual learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  6. [6]

    Ushani, and Ryan M

    Nicholas Carlevaris-Bianco, Arash K. Ushani, and Ryan M. Eustice. University of Michigan North Campus long-term vision and lidar dataset.International Journal of Robotics Research, 35(9):1023–1035, 2015

    2015

  7. [7]

    X. Li, H. Wang, Z. Zhang, Z. Zhang, and Z. Chen. Continual Learning: A Review of Techniques, Challenges, and Future Directions.arXiv preprint arXiv:2403.06501, 2024

    arXiv 2024

  8. [8]

    Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwi ´nska, et al

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwi ´nska, et al. Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017

    2017

Show all 50 references
  1. [9]

    Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwi ´nska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwi ´nska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting...

    2017

  2. [10]

    Mitigating the Di- minishing Effect of Elastic Weight Consolidation.arXiv preprint arXiv:2210.03816, 2022

    Canasai Kruengkrai and Junichi Yamagishi. Mitigating the Di- minishing Effect of Elastic Weight Consolidation.arXiv preprint arXiv:2210.03816, 2022

    2022 arXiv

  3. [11]

    Learning without Forgetting

    Zhizhong Li and Derek Hoiem. Learning without Forgetting. In European Conference on Computer Vision (ECCV), 2016

    2016

  4. [13]

    Distilling the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  5. [14]

    Continual Learning Through Synaptic Intelligence.arXiv preprint arXiv:1703.04200, 2017

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual Learning Through Synaptic Intelligence.arXiv preprint arXiv:1703.04200, 2017

    2017 arXiv

  6. [15]

    Prototype-sample relation distillation: Towards replay-free continual learning

    Xudong Wang et al. Prototype-sample relation distillation: Towards replay-free continual learning. InTo be updated, 2023

    2023

  7. [16]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2001–2010, 2017

    2001

  8. [17]

    S. A. Rebuffi, A. Kolesnikov, and C. H. Lampert. iCaRL: Incremental Classifier and Representation Learning.CVPR, 2017

    2017

  9. [18]

    Gradient Episodic Memory for Continual Learning.arXiv preprint arXiv:1706.08850, 2017

    David Lopez-Paz and Marc’Aurelio Ranzato. Gradient Episodic Memory for Continual Learning.arXiv preprint arXiv:1706.08850, 2017

    2017 arXiv

  10. [19]

    Dark experience for general continual learning

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  11. [20]

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

    Zheda Mai, Ruiqi Gao, Feng Liu, Sheng Jie, Diyi Yang, and Zhi- wei Steven Wu. Dark experience for general continual learning: A strong, simple baseline. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  12. [21]

    Packnet: Adding multiple tasks to a single network by iterative pruning

    Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  13. [22]

    Overcoming catastrophic forgetting with hard attention to the task

    Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning (ICML), 2018

    2018

  14. [23]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  15. [24]

    No Forgetting Learning: A Memory-free Continual Learning Framework.arXiv preprint arXiv:2503.04638, 2025

    Xiaowei Ma, Weiran Huang, Jiahua Dong, and Yining Qi. No Forgetting Learning: A Memory-free Continual Learning Framework.arXiv preprint arXiv:2503.04638, 2025

    2025 arXiv

  16. [25]

    Solutions of Ill-posed Problems

    Andrey Nikolayevich Tikhonov and Vasiliy Yakovlevich Arsenin. Solutions of Ill-posed Problems. Winston, 1977

    1977

  17. [26]

    Al-gnn: Privacy-preserving and replay-free continual graph learning via analytic learning.To be updated, 2025

    First Author et al. Al-gnn: Privacy-preserving and replay-free continual graph learning via analytic learning.To be updated, 2025

    2025

  18. [27]

    Catastrophic Forgetting in Deep Graph Networks: A Graph Classifica- tion Benchmark.Frontiers in Artificial Intelligence, 5, 2022

    Antonio Carta, Andrea Cossu, Federico Errica, and Davide Bacciu. Catastrophic Forgetting in Deep Graph Networks: A Graph Classifica- tion Benchmark.Frontiers in Artificial Intelligence, 5, 2022

    2022

  19. [28]

    1 year, 1000 km: The oxford robotcar dataset.The International Journal of Robotics Research, 36(1):3–15, 2017

    Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, 1000 km: The oxford robotcar dataset.The International Journal of Robotics Research, 36(1):3–15, 2017

    2017

  20. [29]

    Mapillary Street-Level Sequences: A Dataset for Lifelong Place Recognition

    Frederik Warburg, Soren Hauberg, Manuel Lopez-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary Street-Level Sequences: A Dataset for Lifelong Place Recognition. InIEEE Conference on Computer Vision and Pattern Recognition, 2020

    2020

  21. [30]

    Continual slam: Beyond lifelong simultaneous localization and mapping through continual learning.IEEE Transactions on Robotics, 2023

    Giulio Baldini, Nicholas Kromer, and Luca Carlone. Continual slam: Beyond lifelong simultaneous localization and mapping through continual learning.IEEE Transactions on Robotics, 2023

    2023

  22. [31]

    How SLAM (Simultaneous Localization and Mapping) Enhances AR Experiences, 2025

    Zilliz. How SLAM (Simultaneous Localization and Mapping) Enhances AR Experiences, 2025

    2025

  23. [32]

    Continuous learning route map for robot navigation using a growing-on-demand self-organizing neural network.arXiv preprint arXiv:2403.02256, 2024

    Chaoliang Zhong, Shirong Liu, Qiang Lu, and Botao Zhang. Continuous learning route map for robot navigation using a growing-on-demand self-organizing neural network.arXiv preprint arXiv:2403.02256, 2024

    2024 arXiv

  24. [33]

    Zhong, S

    C. Zhong, S. Liu, Q. Lu, and B. Zhang. Lifelong navigation and mapping for mobile robots with continuous learning.arXiv preprint arXiv:2407.21416, 2024

    2024 arXiv

  25. [34]

    S. Zeng, Y . Liu, Y . Wang, and Y . Li. Continual Learning for Mobile Robot Navigation.MDPI, 2024

    2024

  26. [35]

    Malossi et al

    Gianluca I. Malossi et al. Vascular fluid-structure interaction: unified continuum formulation, image-based mesh generation pipeline, and scalable fully implicit solver technology.To be updated, 2022

    2022

  27. [36]

    H ¨ausler, N

    S. H ¨ausler, N. Sunderhauf, L. v.d. Hengel, F. T. K. L. Hausler, and M. Milford. Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition. InCVPR 2021, 2021

    2021

  28. [37]

    VIPeR: Visual Incremental Place Recognition With Adaptive Mining and Continual Learning.IEEE Conference on Robotics and Automation, 2025

    Yuhang Ming, Minyang Xu, Xingrui Yang, Wanzeng Kong, W Wang, Y Peng, W Dai, and W Kong. VIPeR: Visual Incremental Place Recognition With Adaptive Mining and Continual Learning.IEEE Conference on Robotics and Automation, 2025

    2025

  29. [38]

    Anyloc: Towards universal visual place recognition.arXiv preprint arXiv:2308.00688, 2023

    Nikhil Keetha, Abhishek Tyagi, Akshay Agrawal, Niko Sunderhauf, Michael Milford, and Sourav Garg. Anyloc: Towards universal visual place recognition.arXiv preprint arXiv:2308.00688, 2023

    2023 arXiv

  30. [39]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy P Lillicrap, and Greg Wayne. Experience replay for continual learning. InAdvances in Neural Information Processing Systems, volume 32, pages 348–358, 2019

    2019

  31. [40]

    Rethinking experience replay: a bag of tricks for continual learning

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking experience replay: a bag of tricks for continual learning. InProceedings of the 25th International Conference on Pattern Recognition (ICPR), pages 2180–2187, 2020

    2020

  32. [41]

    Overcoming catastrophic forgetting with hard attention to the task

    Joan Serr `a, D´ıdac Sur´ıs, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In Proceedings of the 35th International Conference on Machine Learning (ICML), volume 80 ofProceedings of Machine Learning Research, page...

    2018

  33. [42]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935–2947, 2018

    2018

  34. [43]

    Prol: Rehearsal free continual learning in streaming data via prompt online learning

    M Anwar Ma’sum, Mahardhika Pratama, Savitha Ramasamy, Lin Liu, Habibullah Habibullah, and Ryszard Kowalczyk. Prol: Rehearsal free continual learning in streaming data via prompt online learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages ...

    2025

  35. [44]

    Replay-free continual low-rank adaptation with dynamic memory

    Huancheng Chen, Jingtao Li, Weiming Zhuang, Chen Chen, and Lingjuan Lyu. Replay-free continual low-rank adaptation with dynamic memory. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026

    2026

  36. [45]

    Extensions of lipschitz mappings into a hilbert space

    William B Johnson and Joram Lindenstrauss. Extensions of lipschitz mappings into a hilbert space. InConference in Modern Analysis and Probability, volume 26 ofContemporary Mathematics, pages 189–206. American Mathematical Society, 1984. APPENDIX A. Adaptation of CLEAR for Life...

    1984

  37. [46]

    Feature Extraction: The original raw sensor/image backbone is replaced with pre-computed, frozen 1D DINOv2 feature embeddings

  38. [47]

    Output Head: The pose-regression heads are re- placed with a 100-class spatial classification head trained on spatial pseudo-labels derived from global GPS trajectories

  39. [48]

    c) Implementation and Adaptation of PROL Baseline: To benchmark against state-of-the-art rehearsal-free online continual learning, we incorporate Prompt Online Learning (PROL) [43]

    Training Pipeline: Model updates are performed on offline DINOv2 feature caches introduced se- quentially by geographic domain, rather than from live streaming sensor data. c) Implementation and Adaptation of PROL Baseline: To benchmark against state-of-the-art rehearsal-free ...

  40. [49]

    A dual-adapter architecture consisting of an orthogonal adapter (O) to ensure parameter stability and a residual adapter (R) to enhance task plasticity

  41. [50]

    Singular Value Decomposition (SVD)-based gradient projection that constrains weight updates to subspaces orthogonal to previously learned tasks

  42. [51]

    A Dynamic Memory (DM) module combined with inference-time logit calibration, which adaptively weights residual features based on input descriptor similarities. b) What is Modified:To adapt DualLoRA from ViT- based image classification to sequential robotic VPR using the NCLT d...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.