Pith. sign in

REVIEW 3 major objections 5 minor 36 references

A task-free online continual learning method, Dual-LS, reduces catastrophic forgetting by up to 74.31% and computational resource demand by up to 94.02% in vehicle motion forecasting.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Dual-LS pairs reservoir and gradient-diversity replay buffers with fast and slow EMA model updates to mitigate catastrophic forgetting in task-free online vehicle motion forecasting.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Dual-LS is a clean integration of known replay and EMA tricks for motion forecasting; the results look plausible, but the headline numbers are undersupported and the self-distillation target needs an ablation. the 3 major comments →

arxiv 2508.19597 v2 pith:ZPTGKCC7 submitted 2025-08-27 cs.LG cs.AI

Complementary Learning System Empowers Online Continual Learning of Vehicle Motion Forecasting in Smart Cities

Zirui Li , Yunlong Lin , Guodong Du , Xiaocong Zhao , Cheng Gong , Chen Lv , Chao Lu , Jianwei Gong This is my paper
classification cs.LG cs.AI
keywords vehicle motion forecastingcontinual learningcatastrophic forgettingcomplementary learning systemmemory replaytask-free online learningautonomous drivingsmart cities
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Vehicle motion forecasting models deployed in smart cities must keep learning from a continuous stream of new traffic scenarios, but updating a deep network on new data makes it forget old scenarios. This paper introduces Dual-LS, a task-free online continual learning method that combines two replay memory buffers—one that samples past data uniformly and one that favors diverse, gradient-representative samples—with a fast/slow pair of model copies acting like short- and long-term memory. On the INTERACTION dataset, the method mitigates catastrophic forgetting by up to 74.31% and cuts computational resource demand by up to 94.02%, while keeping prediction errors closer to joint training than A-GEM, GSS, and DER baselines. A reader should care because it suggests autonomous driving systems can become permanently learnable without storing and reprocessing huge data archives.

Core claim

The paper claims that the complementary learning system theory of human memory—rapid hippocampal learning plus slow neocortical consolidation—can be translated directly into an algorithmic design for online continual learning. The implementation maintains a working model trained on the current stream, plus a fast model and a slow model updated as exponential moving averages. For each replayed sample, the model whose prediction has lower error supplies the target heatmap for a KL-divergence distillation term, letting the working model learn from whichever memory system currently knows the sample better. Two buffers feed this process: reservoir sampling gives equal-probability coverage of the

What carries the argument

The load-bearing machinery is Dual-LS: (1) a dual memory replay buffer combining reservoir sampling (equal probability over the data stream) with gradient-based diversity sampling (cosine-similarity of loss gradients to maximize buffer diversity); and (2) a long-short term model update strategy in which the working model is trained with a replay loss whose KL-divergence target is chosen from a fast EMA model and a slow EMA model based on which predicts each replayed sample better. The two buffers provide coverage and diversity; the fast/slow choice supplies a stable, non-degenerate learning target for replay.

Load-bearing premise

The method assumes that its own fast and slow model predictions provide a trustworthy teaching signal for replay; if those predictions drift together or become uninformative, the distillation loss would anchor the model to itself and the reported forgetting reductions would shrink.

What would settle it

Freeze the fast and slow EMA targets at the model checkpoint before a new task, or replace the lower-error selection in Eq. (17) with always choosing the slow model, and measure FDE-BWT on the INTERACTION stream; if the forgetting reduction persists, the long-short target selection is not the active ingredient.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, vehicle motion forecasting models can be updated continuously on a bounded memory budget without sacrificing performance on previously encountered traffic scenarios.
  • The approach removes the need for task identifiers, so deployment does not require manually defining scenario boundaries.
  • A roughly 94% reduction in computational resource demand could lower storage, energy, and maintenance costs for smart-city AI systems that currently rely on large-scale joint retraining.
  • The dual-buffer, fast/slow design is presented as a general paradigm, extendable to other data-driven prediction systems beyond motion forecasting.
  • Empirically, larger memory buffers improve both GSS and Dual-LS, suggesting replay capacity remains a key lever for forgetting reduction in this setting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same dual-buffer/fast-slow recipe may transfer to other heatmap-based prediction tasks, such as pedestrian trajectory prediction or robot navigation, but the paper only tests one base network and one dataset.
  • Freezing the EMA targets or replacing the lower-error selection with a fixed slow-model target would reveal how much of the gain comes from the CLS-inspired mechanism versus from replay alone.
  • The diversity buffer visibly over-represents later tasks, suggesting Dual-LS may implicitly prioritize recent distributions; in very long streams the two buffers might need dynamic size allocation to stay balanced.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Dual-LS, a task-free online continual learning (CL) algorithm for deep-net vehicle motion forecasting. It combines two replay buffers—one based on reservoir sampling and one based on gradient-diversity sampling—with a 'long-short term' update strategy in which exponential-moving-average (EMA) models (fast and slow) are used to select a target for a KL distillation term during replay. The method is evaluated on the INTERACTION dataset with 8 traffic scenarios, comparing against Vanilla, A-GEM, GSS, and DER, and against joint training as a lower-bound reference. The authors report lower FDE/miss-rate averages and better backward-transfer (BWT) metrics, as well as reductions in memory and compute demand; the abstract headlines 'mitigates catastrophic forgetting by up to 74.31%' and 'reduces computational resource demand by up to 94.02%'.

Significance. If the results hold, Dual-LS would provide a practical rehearsal-based approach for keeping motion-forecasting models up to date in smart-city deployments without task labels, using only a bounded memory buffer. The paper uses a realistic multi-country dataset, reports 10 repeated runs with standard deviations, and makes code available. However, the methodological novelty over existing replay-plus-distillation baselines (e.g., DER) is incremental, and the central mechanism (self-distillation via EMA copies) is not ablated. The headline quantitative claims are not tied to specific tables or derivations in the main text, which weakens the immediate credibility of the abstract's claims.

major comments (3)
  1. [Abstract; §Results (Mitigating memory and computational resource demand)] The headline percentages are not traceable. The 74.31% forgetting-mitigation figure appears only in the abstract; no equation or table in the main text defines it. The 94.02% resource reduction appears in the text but is computed against joint training (JT), which trivially uses all data and is not a CL baseline. Neither number is accompanied by a protocol (e.g., relative BWT reduction vs Vanilla? memory measured as peak GPU memory?). Since these are the paper's central quantitative claims, they must be tied to a specific table/derivation and reported across tasks, not only Task 1 (Fig. 3a,b) and Task 3 (Fig. 3c,d).
  2. [§Methods, Eqs. (14)–(17); Algorithm 3] The KL-distillation target is self-referential. Eq. (17) selects f_theta*(X_M) as the output of fast/slow EMA models f^F/f^S, which are defined by Eqs. (15)–(16) as exponential moving averages of the same working model theta_W being updated. If theta_W drifts on past tasks, the EMAs inherit that drift, so the D_KL term in Eq. (14) can reward matching the model's own moving average rather than recovering the old-task predictive distribution. The true-label terms beta1, beta2 may anchor learning, but the manuscript provides no ablation that removes or replaces the Eq. (17) target (e.g., using a fixed task-boundary snapshot, or the working model itself). Without such an ablation, the contribution of the proposed CLS-inspired long-short term mechanism is not isolated, and the BWT improvements cannot be attributed specifically to it.
  3. [§Results, Fig. 3; §Methods (Implementation)] The computational-resource claim is not properly measured. Fig. 3a/c plots 'cumulative volume of data processed,' which is proportional to memory footprint only under a specific batch-processing assumption, not to actual computation (FLOPs, GPU time, peak memory). The 34.08% compute reduction vs Vanilla is stated in the text without any definition of the measurement. No resource benchmark is described in Methods. As one of the two headline claims is about computational resource demand, the paper should report a standard resource metric (e.g., training time, peak GPU memory, FLOPs) under the same experimental protocol.
minor comments (5)
  1. [Eq. (14)] There is a missing '+' before the alpha_2 term; as printed, Eq. (14) appears to have two separate expressions rather than a single sum.
  2. [Algorithm 3, lines 12–13] 'Update M_R as in (15)' and 'Update M_D as in (16)' are incorrect: Eqs. (15)–(16) define EMA updates, not memory-buffer updates. The text should refer to Algorithms 1 and 2.
  3. [Fig. 4b caption] 'The error bars donate the standard deviation' should be 'denote'.
  4. [Eq. (21) and surrounding text] The miss-rate definition uses N'_t and W but W is not defined. It should be specified (apparently the number of goal samples per scenario).
  5. [§Results (Mitigating catastrophic forgetting)] In the BWT formulas (Eqs. (23)–(24)), 'c' is used without definition; it should be the current task index. Clarify the notation.

Circularity Check

0 steps flagged

No significant circularity: the central forgetting-mitigation claim is measured against held-out ground truth and joint training, and the EMA self-distillation target is a training heuristic rather than a forced prediction.

full rationale

The paper's main claims—mitigating catastrophic forgetting by up to 74.31% and reducing computational demand by up to 94.02%—are empirical results computed on held-out scenario test sets (FDE, MR, FDE-BWT, MR-BWT; Eqs. 18–24) relative to joint training and standard baselines. No fitted parameter is renamed as a prediction, and no external benchmark result is asserted from the authors' own prior work. The only self-referential element is Eq. (17), where the KL-distillation target f_{theta*}(X_M) is selected from the fast and slow models, which are themselves exponential moving averages of the working model (Eqs. 15–16). This makes the distillation term a self-distillation / temporal-consistency regularizer rather than an independent teacher. However, the paper does not derive the reported forgetting reductions from this term alone; the true-label replay losses (beta_1, beta_2 terms in Eq. 14) also contribute, and the evaluation is against ground truth, not against the model's own output. The lack of an ablation isolating Eq. (17) weakens causal attribution but is an experimental-design limitation, not a logical circularity. The use of the authors' prior UQnet as the base network is a minor self-citation, but it is applied identically across all methods and does not carry the load of the CL claim. No uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result as a new derivation are present. The paper is self-contained against external INTERACTION-data benchmarks, so no significant circularity is found.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities; the fast and slow models are algorithmic components. The central claim rests on several empirical hyperparameters and domain assumptions about the INTERACTION dataset and the UQnet backbone.

free parameters (6)
  • Replay loss weights α1, α2, β1, β2 = not reported in main text
    Introduced in Eq. (14) to balance the KL and label losses from the two memory buffers; tuned empirically.
  • EMA decay rates αF, αS = not reported in main text
    Control the update speed of the fast and slow models in Eqs. (15) and (16); set manually.
  • Update rates pF, pS = not reported in main text
    Stochastic update probabilities for the fast and slow models in Algorithm 3, with pF < pS.
  • Memory buffer sizes |MR| and |MD| = total buffer size varied: 500, 1000, 2000, 4000
    The total buffer capacity is a controlled variable; the split between reservoir and diversity buffers is not detailed.
  • Number of samples B for similarity score = not reported in main text
    Used in Algorithm 2 to estimate the gradient similarity score; affects buffer update behavior.
  • Learning rate αW = not reported in main text
    Learning rate for the working model update in Algorithm 3.
axioms (5)
  • standard math Reservoir sampling provides uniform random selection from an infinite stream.
    Used to justify the equal-probability property of buffer MR (Algorithm 1, Fig. 6a).
  • domain assumption The eight INTERACTION scenarios define meaningful tasks and the fixed stream order is representative of real smart-city data streams.
    The experiments use scenario order as the data stream; conclusions about task-free CL depend on this segmentation.
  • domain assumption The UQnet heatmap output supports KL divergence between model predictions and is a valid representation for distillation.
    Eq. (14) uses DKL on heatmaps; the paper assumes this is a meaningful training signal.
  • domain assumption The fast/slow model selection in Eq. (17) provides a valid, non-degenerate target for replay.
    The KL target is the better of two EMA models; if both EMA models are poor, this target could reinforce errors.
  • domain assumption Backward transfer metrics computed with known task boundaries accurately measure forgetting in a task-free setting.
    Evaluation uses task indices to compute BWT, even though the algorithm is claimed to be task-free.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Complementary Learning System Empowers Online Continual Learning of Vehicle Motion Forecasting in Smart Cities." pith.science (2026). https://pith.science/paper/ZPTGKCC7

@misc{pith2026250819597,
  author       = {Pith},
  title        = {Pith review of: Complementary Learning System Empowers Online Continual Learning of Vehicle Motion Forecasting in Smart Cities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZPTGKCC7}},
  note         = {Machine review of arXiv:2508.19597}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Artificial intelligence underpins most smart city services, yet deep neural network (DNN) that forecasts vehicle motion still struggle with catastrophic forgetting, the loss of earlier knowledge when models are updated. Conventional fixes enlarge the training set or replay past data, but these strategies incur high data collection costs, sample inefficiently and fail to balance long- and short-term experience, leaving them short of human-like continual learning. Here we introduce Dual-LS, a task-free, online continual learning paradigm for DNN-based motion forecasting that is inspired by the complementary learning system of the human brain. Dual-LS pairs two synergistic memory rehearsal replay mechanisms to accelerate experience retrieval while dynamically coordinating long-term and short-term knowledge representations. Tests on naturalistic data spanning three countries, over 772,000 vehicles and cumulative testing mileage of 11,187 km show that Dual-LS mitigates catastrophic forgetting by up to 74.31\% and reduces computational resource demand by up to 94.02\%, markedly boosting predictive stability in vehicle motion forecasting without inflating data requirements. Meanwhile, it endows DNN-based vehicle motion forecasting with computation efficient and human-like continual learning adaptability fit for smart cities.

Figures

Figures reproduced from arXiv: 2508.19597 by Chao Lu, Cheng Gong, Chen Lv, Guodong Du, Jianwei Gong, Xiaocong Zhao, Yunlong Lin, Zirui Li.

Figure 1
Figure 1. Figure 1: The proposed CLS-inspired Dual-LS algorithm. a, The DNN model is sequentially trained by the samples from each scenario and trying to adapt to the newly encounter one. It may suffer from the catastrophic forgetting when meets the previously learnt scenario. Left: the traffic monitor system in smart cities; Middle: the DNN-based system continual adaptation to various traffic scenarios; Right: the catastroph… view at source ↗
Figure 2
Figure 2. Figure 2: The illustration of the used dataset30. Three scenarios (roundabout, merging and intersection) are selected for comparison. a, The visualization of map and all trajectories in the training set. b, The distributions of distance between the target vehicle and surrounding vehicles. c, The distributions of distance between the vehicle speed. The difference in trajectories, relative distance and speed indicates… view at source ↗
Figure 3
Figure 3. Figure 3: The overall performance of Dual-LS, Vanilla and joint training in mitigating catastrophic forgetting, memory and computational resource demand. a and c, the cumulative volume of data processed over the entire training procedure for Task 1 Task 3, respectively. b and d (from left to right), the relative memory used between JT and Dual-LS, the missing rate error (%) between JT and Dual-LS, the relative compu… view at source ↗
Figure 4
Figure 4. Figure 4: The overall results of Dual-LS and other baseline methods in mitigating the catastrophic forgetting. a, The quantification of catastrophic forgetting in FDE and MR (Total memory buffer size: 1000). The x-axis represents that each model is trained from Task 1 to Task 8. The y-axis is the metrics. The transparent area is the standard deviation of 10 repeated experiments. b, The comparison of all methods in d… view at source ↗
Figure 5
Figure 5. Figure 5: The CLS-inspired learning mechanism of Dual-LS in the CL setting. a, The proportion of samples from each past tasks in the memory buffer. It changes after each update of memory buffer. The comparison of two kinds of sampling strategies highlights the significant of CLS in memory buffer design. b, The constitution of two memory buffers from each past task. The color represents the number of samples from cor… view at source ↗
Figure 6
Figure 6. Figure 6: Two memory sampling strategies in Dual-LS. a, Reservoir sampling strategy selects each sample from an unlimited online stream with equal probability: First, the buffer is initially empty. As new samples are observed, they are stored in the buffer until its capacity is reached. Once full, newly encountered samples may replace one of the existing samples. The probability of a sample being stored is designed … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 35 canonical work pages

  1. [2]

    3.Jumper, J.et al.Highly accurate protein structure prediction with alphafold.nature596, 583–589 (2021)

    Silver, D.et al.A general reinforcement learning algorithm that masters chess, shogi, and go through self-play.Science 362, 1140–1144 (2018). 3.Jumper, J.et al.Highly accurate protein structure prediction with alphafold.nature596, 583–589 (2021). 4.Feng, S.et al.Dense reinforcement learning for safety validation of autonomous vehicles.Nature615, 620–627 (...

    work page 2018

  2. [6]

    & Scaramuzza, D

    Song, Y ., Romero, A., Müller, M., Koltun, V . & Scaramuzza, D. Reaching the limit in autonomous racing: Optimal control versus reinforcement learning.Sci. Robotics8, eadg1462 (2023). 7.Gehrig, D. & Scaramuzza, D. Low-latency automotive vision with event cameras.Nature629, 1034–1040 (2024)

    work page 2023

  3. [8]

    Y ., Dianati, M., Jennings, P

    Mozaffari, S., Al-Jarrah, O. Y ., Dianati, M., Jennings, P. & Mouzakitis, A. Deep learning-based vehicle behavior prediction for autonomous driving applications: A review.IEEE Transactions on Intell. Transp. Syst.23, 33–47 (2020)

    work page 2020

  4. [9]

    communications15, 1306 (2024)

    Wang, X.et al.Traffic light optimization with low penetration rate vehicle trajectory data.Nat. communications15, 1306 (2024)

    work page 2024

  5. [10]

    & Liu, Y .-C

    Jafari, A. & Liu, Y .-C. Pedestrians’ safety using projected time-to-collision to electric scooters.Nat. communications15, 5701 (2024)

    work page 2024

  6. [11]

    Chen, Z.et al.Predicting driving comfort in autonomous vehicles using road information and multi-head attention models. Nat. Commun.16, 2709 (2025)

    work page 2025

  7. [12]

    Cao, Z.et al.Continuous improvement of self-driving cars using dynamic confidence-aware reinforcement learning.Nat. Mach. Intell.5, 145–158 (2023)

    work page 2023

  8. [13]

    Lesort, T.et al.Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges. Inf. fusion58, 52–68 (2020)

    work page 2020

  9. [14]

    Tajbakhsh, N.et al.Convolutional neural networks for medical image analysis: Full training or fine tuning?IEEE transactions on medical imaging35, 1299–1312 (2016)

    work page 2016

  10. [15]

    communications12, 5678 (2021)

    Perkonigg, M.et al.Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging.Nat. communications12, 5678 (2021)

    work page 2021

  11. [16]

    M., Tuytelaars, T

    Van de Ven, G. M., Tuytelaars, T. & Tolias, A. S. Three types of incremental learning.Nat. Mach. Intell.4, 1185–1197 (2022)

    work page 2022

  12. [17]

    national academy sciences114, 3521–3526 (2017)

    Kirkpatrick, J.et al.Overcoming catastrophic forgetting in neural networks.Proc. national academy sciences114, 3521–3526 (2017)

    work page 2017

  13. [18]

    & Zhu, J

    Wang, L., Zhang, X., Su, H. & Zhu, J. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis Mach. Intell.46, 5362–5383 (2024)

    work page 2024

  14. [19]

    De Lange, M.et al.A continual learning survey: Defying forgetting in classification tasks.IEEE transactions on pattern analysis machine intelligence44, 3366–3385 (2021)

    work page 2021

  15. [20]

    & Calderara, S

    Buzzega, P., Boschini, M., Porrello, A., Abati, D. & Calderara, S. Dark experience for general continual learning: a strong, simple baseline.Adv. neural information processing systems33, 15920–15930 (2020). 21.Kudithipudi, D.et al.Biological underpinnings for lifelong learning machines.Nat. Mach. Intell.4, 196–210 (2022)

    work page 2020

  16. [22]

    & McClelland, J

    Kumaran, D., Hassabis, D. & McClelland, J. L. What learning systems do intelligent agents need? complementary learning systems theory updated.Trends cognitive sciences20, 512–534 (2016)

    work page 2016

  17. [23]

    & Fitzgerald, J

    Sun, W., Advani, M., Spruston, N., Saxe, A. & Fitzgerald, J. E. Organizing memories for generalization in complementary learning systems.Nat. neuroscience26, 1438–1448 (2023)

    work page 2023

  18. [24]

    & Bengio, Y

    Aljundi, R., Lin, M., Goujaud, B. & Bengio, Y . Gradient based sample selection for online continual learning.Adv. neural information processing systems32(2019). 25.Yan, X.et al.Learning naturalistic driving environment with statistical realism.Nat. communications14, 2037 (2023)

    work page 2019

  19. [26]

    & Liu, H

    Feng, S., Yan, X., Sun, H., Feng, Y . & Liu, H. X. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment.Nat. communications12, 748 (2021). 27.Yu, W.et al.Online legal driving behavior monitoring for self-driving vehicles.Nat. communications15, 408 (2024). 15/19

    work page 2021

  20. [28]

    communica- tions14, 8002 (2023)

    Duan, J.et al.Spatiotemporal dynamics of traffic bottlenecks yields an early signal of heavy congestions.Nat. communica- tions14, 8002 (2023)

    work page 2023

  21. [29]

    & Yilmaz, Y

    Doshi, K. & Yilmaz, Y . Continual learning for anomaly detection in surveillance videos. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 254–255 (2020)

    work page 2020

  22. [30]

    Zhan, W.et al.Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps.arXiv preprint arXiv:1910.03088(2019)

    Pith/arXiv arXiv 1910

  23. [31]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9710–9719 (2021)

    Ettinger, S.et al.Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9710–9719 (2021)

    work page 2021

  24. [32]

    & Elhoseiny, M

    Chaudhry, A., Ranzato, M., Rohrbach, M. & Elhoseiny, M. Efficient lifelong learning with a-gem. InInternational Conference on Learning Representations(2018)

    work page 2018

  25. [33]

    Syst.36, 7786–7805 (2025)

    Xing, J.et al.A comprehensive study on self-learning methods and implications to autonomous driving.IEEE Transactions on Neural Networks Learn. Syst.36, 7786–7805 (2025). 34.Wang, W.et al.Social interactions for autonomous driving: A review and perspectives.F oundations Trends Robotics10, 198–376 (2022). 35.Dai, T.et al.Autonomous mobile robots for explor...

    work page 2025

  26. [36]

    Zhou, C.et al.Human-behaviour-based social locomotion model improves the humanization of social robots.Nat. Mach. Intell.4, 1040–1052 (2022)

    work page 2022

  27. [37]

    Meng, Y .et al.Preserving and combining knowledge in robotic lifelong reinforcement learning.Nat. Mach. Intell.7, 256–269 (2025)

    work page 2025

  28. [38]

    Commun.15, 6793 (2024)

    Stanojevic, A.et al.High-performance deep spiking neural networks with 0.3 spikes per neuron.Nat. Commun.15, 6793 (2024)

    work page 2024

  29. [39]

    & Bohté, S

    Yin, B., Corradi, F. & Bohté, S. M. Accurate online training of dynamical spiking neural networks through forward propagation through time.Nat. Mach. Intell.5, 518–527 (2023)

    work page 2023

  30. [40]

    Syst.36, 5921–5941 (2025)

    Jiao, L.et al.Brain-inspired learning, perception, and cognition: A comprehensive review.IEEE Transactions on Neural Networks Learn. Syst.36, 5921–5941 (2025)

    work page 2025

  31. [41]

    Commun.13, 65 (2022)

    Wu, Y .et al.Brain-inspired global-local learning incorporated with neuromorphic computing.Nat. Commun.13, 65 (2022)

    work page 2022

  32. [42]

    M., Siegelmann, H

    Van de Ven, G. M., Siegelmann, H. T. & Tolias, A. S. Brain-inspired replay for continual learning with artificial neural networks.Nat. communications11, 4069 (2020). 43.Williamson, M. R.et al.Learning-associated astrocyte ensembles regulate memory recall.Nature637, 478–486 (2024)

    work page 2020

  33. [44]

    P.et al.Organizing the coactivity structure of the hippocampus from robust to flexible memory.Science385, 1120–1127 (2024)

    Gava, G. P.et al.Organizing the coactivity structure of the hippocampus from robust to flexible memory.Science385, 1120–1127 (2024). 45.Zaki, Y .et al.Offline ensemble co-reactivation links memories across days.Nature637, 145–155 (2024). 46.Shuai, Y .et al.Forgetting is regulated through rac activity in drosophila.Cell140, 579–589 (2010). 47.Soltoggio, A....

    work page 2024

  34. [48]

    & Botvinick, M

    Hassabis, D., Kumaran, D., Summerfield, C. & Botvinick, M. Neuroscience-inspired artificial intelligence.Neuron95, 245–258 (2017)

    work page 2017

  35. [49]

    Li, G., Li, Z., Knoop, V . L. & van Lint, H. Unravelling uncertainty in trajectory prediction using a non-parametric approach. Transp. Res. Part C: Emerg. Technol.163, 104659 (2024)

    work page 2024

  36. [50]

    & Ranzato, M

    Lopez-Paz, D. & Ranzato, M. Gradient episodic memory for continual learning.Adv. neural information processing systems30(2017). Acknowledgements This research is supported by the National Natural Science Foundation of China under Grant 52372405. Data availability The dataset used in this study is publicly available on https://interaction-dataset.com/. 16/...

    work page 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.