Pith. sign in

REVIEW 3 major objections 4 minor 115 references

NeuMoSync: End-to-End Neuromodulatory Control for Plasticity and Adaptability in Continual Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A central controller that issues four per-neuron modulation signals preserves plasticity and improves forward and backward adaptation in continual learning, with the largest gains on memorization tasks.

desk verdict Solid architecture, credible ablations; the knowledge-transfer attribution is overclaimed because the LCA decomposition in Table 1 doesn't add up. read the letter →

arxiv 2608.04358 v1 pith:XTBEFTCT submitted 2026-08-05 cs.AI cs.LG

Seyed Roozbeh Razavi Rohani , Khashayar Khajavi , Wesley Chung , Mandana Samiei , Mo Chen This is my paper · ORCID
classification cs.AIcs.LG
keywords ContinuallearningPlasticitypreservationNeuromodulationAdaptivelinearityWeightconsolidationKnowledgetransferGlobalcontrollerParametersharing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a deep network can keep its capacity to learn across a long sequence of tasks if a separate controller is allowed to modulate every neuron individually, in real time, based on both the current input and the network's own evolving state. The NeuMoSync architecture pairs a fast-adapting MainNetwork with a slowly moving ConsolidatedNetwork (an exponential moving average of the main weights), and the controller blends the two per neuron, sets each neuron's activation slope, and adds an offset, all trained end-to-end with ordinary gradient descent, with no meta-learning and no weight resets. On benchmarks deliberately stripped of transferable structure (random label assignments, shuffled class mappings) the method keeps learning where plasticity-oriented baselines stagnate, and on class- and domain-incremental benchmarks it adapts faster; the paper attributes this advantage to acquired knowledge transfer rather than a faster base learner. If correct, the work is evidence that continual learning can be improved by coordinating plasticity globally, rather than by protecting parameters or periodically reinitializing parts of the network.

What carries the argument

The object that carries the argument is the per-neuron weight-fusion identity $w_i = \alpha^{\mathrm{SM}}_i \theta_i + \alpha^{\mathrm{WC}}_i \phi_i$, which linearly combines the fast MainNetwork parameters $\theta$ with the slow exponential-moving-average parameters $\phi$ of the ConsolidatedNetwork under controller-generated coefficients, together with the paired modulation of the activation function ($\alpha^{\mathrm{AL}}$ as the PReLU negative slope) and the additive offset ($\alpha^{\mathrm{ARM}}$). The controller is a parameter-sharing sequence model, an encoder-only transformer or a 1D CNN, that maps neuron feature vectors (learnable embeddings, positional information, and an EMA of past activations) plus the current input to the four coefficients. The sharpest mechanism is the cancellation-and-steering regime of Appendix H.5: when $\theta$ and $\phi$ are close in direction but never exactly proportional, oppositely signed $\alpha^{\mathrm{SM}}$ and $\alpha^{\mathrm{WC}}$ nearly cancel the parallel components of the two weight vectors, so the angle $\psi$ of the effective weight vector relative to $\theta$ obeys $\tan \psi = (\alpha^{\mathrm{SM}} \|\theta\| \sin \delta) / (\alpha^{\mathrm{WC}} \|\phi\| + \alpha^{\mathrm{SM}} \|\theta\| \cos \delta)$, and small changes in coefficient magnitudes alone can rotate the InferenceNetwork quickly between tasks. That is what lets the system reuse nearly identical fast and slow networks yet still reorient its effective computation.

What would settle it

Run the Random-Label CIFAR-10 protocol while measuring the cosine similarity between the MainNetwork and ConsolidatedNetwork weights (the angle $\delta$) and the effective steering angle $\psi$; the mechanism is refuted if the performance gains persist while $\delta$ collapses to zero, since the cancellation-and-steering regime requires a strictly nonzero $\delta$. A second direct test is to force $\phi = c\,\theta$ (exactly proportional networks) or to constrain $\alpha^{\mathrm{SM}}$ and $\alpha^{\mathrm{WC}}$ to the same sign and check whether the fast-adaptation advantage disappears.

Watch

Extended reading notes

Core claim

The central claim is that four input-dependent, per-neuron coefficients, namely synaptic modulation $\alpha^{\mathrm{SM}}$, weight consolidation $\alpha^{\mathrm{WC}}$, adaptive linearity $\alpha^{\mathrm{AL}}$, and additive regulation $\alpha^{\mathrm{ARM}}$, are sufficient to preserve plasticity and improve forward and backward adaptation across continual learning benchmarks. Each neuron's effective incoming weights are formed by the linear fusion $w_i = \alpha^{\mathrm{SM}}_i \theta_i + \alpha^{\mathrm{WC}}_i \phi_i$, so the controller constructs a fresh InferenceNetwork for every input without ever changing the underlying parameters; $\alpha^{\mathrm{AL}}$ sets the negative slope of the neuron's PReLU activation and $\alpha^{\mathrm{ARM}}$ shifts its output. The empirical case rests on Random Label CIFAR-10 and MNIST, where NeuMoSync keeps improving for 50 to 100 tasks while reset- and activation-based baselines plateau at far lower accuracy, with additional wins on Shuffle CIFAR-10, Permuted MNIST, Class Split CIFAR-100, and Class Split T-ImageNet. Ablations show that each modulation channel is necessary, that parameter sharing in the controller is the critical inductive bias (a 1D-CNN controller matches the transformer while an MLP controller fails), and that the learned signals are interpretable: the MainNetwork is suppressed at task onset, the consolidated network is gated in later as plasticity stress appears, and neurons split into a plastic subpopulation and a stabilized subpopulation. Against meta-learners MAML and ANML, the advantage holds without any meta-training phase.

Load-bearing premise

The load-bearing premise is that the fast MainNetwork and the slow ConsolidatedNetwork remain nearly aligned in direction but never exactly proportional during training; if the exponential averaging ever made them exact scalar multiples of each other, the fusion $w_i = \alpha^{\mathrm{SM}}_i \theta_i + \alpha^{\mathrm{WC}}_i \phi_i$ would reduce to a per-neuron rescaling and the controller would lose its ability to reorient the network between tasks.

Editorial extensions

If this is right

  • On memorization benchmarks where input structure carries no transferable signal, NeuMoSync still learns a transferable strategy for forming rapid associations, an emergent, meta-objective-free analogue of meta-learning.
  • Because the paper derives and uses the decomposition $\mathrm{LCA} = \mathrm{LCA}_0 + \text{average knowledge transfer}$, and NeuMoSync's adaptation advantage is almost entirely in FKT and BKT, fast adaptation in this architecture is an acquired modulation policy rather than faster gradient descent.
  • Parameter sharing is the critical inductive bias: one universal modulation function applied across all neurons succeeds, while giving each neuron its own controller parameters fails, so scaling the idea means scaling a shared attention or pooling pass, not per-neuron capacity.
  • The learned modulation policy is interpretable and layer-dependent: deeper layers are modulated for rapid repurposing and consolidation while the first layer carries more variable modulation, suggesting per-layer or sparse modulation may retain most of the benefit at scale.
  • Combined with a standard experience-replay buffer, the method matches or exceeds stability-oriented baselines such as EWC, A-GEM, and HAT on average forgetting on several benchmarks, so plasticity preservation and forgetting prevention are complementary rather than competing.

Reading between the lines

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

  • If the cancellation-and-steering mechanism is the true driver, the controller is effectively a gain field over a two-network basis, and the resource is the separation between a fast and a slow copy in weight space; that design principle is generic enough to test in other non-stationary settings, such as test-time adaptation or continual reinforcement learning, which the paper does not explore.
  • The paper's metric decomposition suggests a cheap diagnostic for any plasticity method: reinitialize per task and subtract; a method whose gains vanish under reinitialization is a transfer method. Adopting this as a standard reporting protocol would make the source of adaptation gains visible across the field.
  • A concrete extension left implicit: since the direction of the modulation vectors stays nearly constant across tasks in the deeper layer, replacing per-neuron coefficients with per-layer or per-population coefficients there should retain most of the benefit at a fraction of the cost; the paper's coarse-grained ablation hints at exactly this trade-off.
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

3 major / 4 minor

Summary. The paper introduces NeuMoSync, a continual-learning architecture in which a global NeuroSync controller produces four per-neuron modulation coefficients (αSM, αWC, αAL, αARM) that combine a fast MainNetwork with a slow EMA ConsolidatedNetwork, adapt the PReLU slope, and add an activation offset. The resulting InferenceNetwork is trained end-to-end by standard backpropagation. The authors evaluate the method on memorization, concept drift, class-incremental, and domain-incremental benchmarks, report strong plasticity preservation and fast forward/backward adaptation, and use an LCA-based decomposition to argue that the gains come from knowledge transfer rather than intrinsic learning speed. Ablations, a parameter-sharing analysis, scaling experiments, and forgetting/stability experiments are also included.

Significance. If the main claims hold, NeuMoSync is a useful practical contribution: it preserves plasticity without resets or meta-learning, it scales to ResNet-sized backbones with a sparse controller, and it is backed by a public code release, extensive baselines, ablations, and sensitivity analyses. The decomposition of adaptation speed into intrinsic speed and knowledge transfer is methodologically valuable. However, the central attribution claim is undermined by problems in the decomposition: the LCA0 identity is not internally consistent, and the reported numbers violate the stated equations. The empirical comparisons themselves remain informative and largely support the architecture's effectiveness, but the paper's interpretation of why it works needs correction before the claims can be accepted.

major comments (3)
  1. [Appendix B.1, Eqs. (4)-(10)] The decomposition identity is not valid as stated. LCAF sums over tasks k=1..T-1, while LCAB sums over tasks k=0..T-2, so a single LCA0 value cannot satisfy both LCAB = LCA0 + mean(BKT) and LCAF = LCA0 + mean(FKT) unless the scratch-learning curves are symmetric over tasks. The note that 'the definition of LCA0 is slightly distinct between the two proofs' does not resolve this: the proofs use the same symbol LCA0 for two different averages. This issue is load-bearing because Section 3.3 uses the decomposition to conclude that NeuMoSync's advantage is 'almost entirely attributable to superior knowledge transfer'.
  2. [Table 1] The reported numbers violate the stated identities. For NeuMoSync on Random Label MNIST, LCA0=10.13, FKTmean=7.17, and LCAF=18.81, so LCA0+FKTmean=17.30, not 18.81; the implied LCA0 from the forward identity is 11.64 and from the backward identity is 11.34, neither equal to 10.13. Similar discrepancies appear in other rows. If FKTmean and BKTmean are not the averages over b that appear in Eqs. (7) and (10), the table should define them explicitly; as presented, the decomposition is not empirically verifiable and cannot support the knowledge-transfer attribution.
  3. [Section 3.3] The text claims NeuMoSync's LCA0 is 'consistently outperformed by other baselines,' but Table 1 contradicts this. For example, on Random Label MNIST NeuMoSync's LCA0 is 10.13 versus 10.03 for L2Init+EWC; on Shuffle CIFAR-10 it is 32.02 versus 31.12 for CReLU; on Class Split CIFAR-100 it is 51.69 versus 51.09 for CBP; and on Permuted MNIST it is higher than CBP, ReDo, and L2Init+EWC. This misstatement weakens the paper's claim that NeuMoSync is not intrinsically faster; the corrected comparison may change the interpretation.
minor comments (4)
  1. [Title page] The author affiliation contains a typo: 'Univeristy' should be 'University'.
  2. [Table 28] In the 'With NeuroSync on Shuffle CIFAR10' table, the second 'αSM-Only' row appears to be mislabeled; from context it should likely be 'αWC-Only'. This also affects the corresponding ablation narrative in Appendix G.1.
  3. [Appendix J] The phrase 'metathetical formulation' appears to be a typo; it should probably be 'mathematical formulation'.
  4. [Appendix H.5] The mechanistic 'cancellation-and-steering' explanation is explicitly speculative, which is appropriate, but it currently reads as a demonstrated mechanism. It would help to label Eq. (11) and the surrounding analysis as a hypothesis supported by correlational evidence, not a proven account.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: NeuMoSync's central claims are empirical comparisons against external baselines; the Appendix B.1 LCA decomposition is an algebraic identity rather than a self-deriving prediction, and the self-citations are motivational and independently ablated.

full rationale

The paper's main architecture (Eq. 1: w_i = alpha_SM theta_i + alpha_WC phi_i; Eq. 2: EMA update of phi) is a constructive definition of the model, not a result derived from its own outputs. The central empirical claims are benchmark comparisons against external baselines (CBP, ReDo, EWC, MAML, ANML, etc.), so no fitted parameter is renamed as a prediction. The Appendix B.1 identity LCAF = LCA0 + mean(FKT) and LCAB = LCA0 + mean(BKT) is a definitional algebraic decomposition: FKT and BKT are defined as accuracy gains over a from-scratch baseline, and LCA0 is defined as the from-scratch LCA. The paper does measure LCA0 independently by reinitializing before each task, so the decomposition is not circular in the sense of fitting LCA0 to force the identity. However, the paper's claim that 'The relationship stated in Equation B.1 ... is empirically verifiable based on the results' is not supported by Table 1: for NeuMoSync on Random Label MNIST, LCA0=10.13, FKTmean=7.17, and LCAF=18.81, giving LCA0+FKTmean=17.30, not 18.81; similar discrepancies appear elsewhere. This is a numerical/consistency problem in the reported decomposition and weakens the 'knowledge transfer' attribution, but it is not a circular derivation. The self-citation to Razavi Rohani et al. (2025) motivates the alpha_AL mechanism and the PReLU baseline, but the paper independently ablates alpha_AL (Figure 3 and Appendix H.4), so the citation is not load-bearing. There is no imported uniqueness theorem and no ansatz smuggled in only via self-citation. Overall, no claim in the paper reduces by construction to its own inputs; the main caveats are the LCA table inconsistency and a minor, non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The architecture's central contribution is empirical; the ledger shows that the main hand-chosen parameters (β, feature dim, embedding size, sampling ratio) affect the reported results, and the mechanism depends on an unproven but empirically observed optimization regime. No external falsifiable handle for the new components is provided beyond the paper's own benchmarks.

free parameters (4)
  • consolidation rate β = 0.999
    Set by hand across all tasks; sensitivity analysis in Appendix G.3 shows performance degrades strongly at 0.9, 0.99, and 0.9999, so the main results hinge on this chosen value.
  • per-neuron feature vector dimensionality k = 4
    Fixed to 4 across all tasks; not ablated, but it defines the context available to the NeuroSync module.
  • NeuroSync embedding size e = 64, 128, or 512 per benchmark
    Selected via grid search per benchmark to keep controller overhead at 5-8% of parameters; affects representation capacity of the controller.
  • sparse sampling ratio K (ResNet experiments) = 0.1 (Shuffle Mini-ImageNet), 0.01 (Split ImageNet)
    Chosen for the encoder-decoder controller; ablation in Figure 9 shows performance degrades as K decreases, so the reported large-scale results depend on this choice.
assumptions (4)
  • domain assumption Backpropagation through the modulation coefficients and EMA-updated consolidated network yields usable gradients.
    The method assumes end-to-end training is stable; Appendix H.2 provides empirical support but no guarantee.
  • domain assumption Average online task accuracy on training batches is a valid measure of plasticity.
    The main plasticity metric is training accuracy on current-task minibatches (Section 3.2); if this does not reflect generalization, the adaptability claim is weakened.
  • domain assumption The MainNetwork and ConsolidatedNetwork remain similar but not exactly proportional under the β=0.999 EMA dynamics.
    The cancellation-and-steering mechanism in Appendix H.5 requires near-cancellation without exact proportionality; if they became exactly proportional, modulation would reduce to scalar per-neuron scaling.
  • standard math Standard gradient descent and the Gram-Schmidt lemma hold.
    Used in the LCA decomposition proof (Appendix B.1) and Lemma 1 (Appendix H.5); elementary and correct.
invented entities (2)
  • NeuroSync global controller module
    purpose: Generates the four per-neuron modulation coefficients conditioned on current input and neuron feature vectors.
    New architectural component introduced by this paper; its effectiveness is demonstrated only on the paper's own benchmarks and ablations. No external replication or formal guarantee is provided.
  • Four per-neuron α coefficients (αSM, αWC, αAL, αARM)
    purpose: Dynamically modulate effective weights, consolidation weights, activation slope, and post-activation offset for each neuron.
    The specific combination of four modulation channels is new to this paper; αAL alone relates to prior PReLU work, but the integrated scheme is only evaluated here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeuMoSync: End-to-End Neuromodulatory Control for Plasticity and Adaptability in Continual Learning." pith.science (2026). https://pith.science/paper/XTBEFTCT

@misc{pith2026260804358,
  author       = {Pith},
  title        = {Pith review of: NeuMoSync: End-to-End Neuromodulatory Control for Plasticity and Adaptability in Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTBEFTCT}},
  note         = {Machine review of arXiv:2608.04358}
}
read the original abstract

Continual learning (CL) requires models to learn tasks sequentially, yet deep neural networks often suffer from plasticity loss and poor knowledge transfer, which can impede their long-term adaptability. Drawing high-level inspiration from global neuromodulatory mechanisms in the brain, we introduce Neuromodulation and Synchronization (NeuMoSync), a novel architecture that integrates dynamic, neuron-specific modulation into deep neural networks to enhance their adaptability and plasticity. NeuMoSync extends standard neural network architectures with learnable feature vectors for each neuron that track network-wide historical context and with a module operating at a higher level of abstraction. This module synthesizes neuron-specific signals, conditioned on both current inputs and the network's evolving state, to adaptively regulate activation dynamics and synaptic plasticity. Evaluated on diverse CL benchmarks, including memorization (Random Label CIFAR-10 and Random Label MNIST), concept drift (Shuffle CIFAR-10 and Shuffle Mini-ImageNet), class-incremental learning (Class Split ImageNet and Class Split CIFAR-100), and domain-incremental learning (Permuted MNIST), NeuMoSync demonstrates strong performance in retaining plasticity and achieves improvements in both forward and backward adaptation compared with existing methods. Ablation studies validate the necessity of each component, while analysis of the learned modulatory signals reveals interpretable coordination patterns across tasks. Our work underscores the potential of integrating global coordination mechanisms into deep learning systems to advance robust, adaptive continual learning. The code is publicly available at https://github.com/RoozbehRazavi/NeuMoSync.

Figures

Figures reproduced from arXiv: 2608.04358 by the authors.

Figure 1
Figure 1. Overview of the architecture and information flow in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Learning curves for three plasticity-evaluation tasks and different baseline algorithms. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Learning curves for ablations of NeuMoSync. Each curve corresponds to removing one type of α-parameters. The black curve corresponds to making all α-parameters learnable (i.e., not input-dependent). All ablations but one show degraded performance compared to full NeuMoSync. 3.4.2 Inductive Biases of NeuroSync Our previous analysis highlights the importance of the NeuroSync module’s architecture, but a key question r… view at source ↗
Figures from the paper (27 more)
Figure 4
Figure 4. Figure 4: These plots show the analyses of the α-parameters and the NeuMoSync module on the Random Label CIFAR-10 tasks. (A) shows that the absolute value of αSM decreases at task changes as losses spike up (B) shows that an increase in αWC co-occurs with an increase in accuracy…
Figure 5
Figure 5. Figure 5: Average Online Task Accuracy on Shuffle Mini-ImageNet and Split ImageNet for our method, NeuMoSync, and the baselines, using a ResNet-18 and ResNet-50 classifier, respectively. plasticity rather than prevent forgetting. We compare against representative forgetting base…
Figure 6
Figure 6. Figure 6: Average Task Accuracy Across Continual Learning Benchmarks. This figure presents the average [PITH_FULL_IMAGE:figures/full_fig_p038_6.png]
Figure 7
Figure 7. Figure 7: Test Accuracy for Generalization Assessment. This figure illustrates the test accuracy for [PITH_FULL_IMAGE:figures/full_fig_p039_7.png]
Figure 8
Figure 8. Figure 8: Learning curves on three families of tasks, measuring knowledge accumulation over repeated [PITH_FULL_IMAGE:figures/full_fig_p039_8.png]
Figure 9
Figure 9. Figure 9: Comparison of training accuracy on Shuffle Mini-ImageNet under different sampeling ratios using [PITH_FULL_IMAGE:figures/full_fig_p040_9.png]
Figure 10
Figure 10. Figure 10: (a) Performance comparison on Random-Label CIFAR10. While the Vanilla ViT exhibits a sharp loss of plasticity, all NeuMoSync-enhanced variants preserve adaptability. The combined approach (NeuMoSync-Both), which modulates both attention heads and MLP blocks, yields th…
Figure 11
Figure 11. Figure 11: The sparse encoder–decoder architecture for the [PITH_FULL_IMAGE:figures/full_fig_p042_11.png]
Figure 12
Figure 12. Figure 12: Effect of global (Global) versus per-neuron (Ours) synaptic modulation signals, αSM and αWC. together with the similarity in the modulation trends of neurons within a layer shown in Figures 14 and 15, this ablation suggests a potential path to improve the scalability …
Figure 13
Figure 13. Figure 13: Effect of conditioning the controller on the input image, evaluated on [PITH_FULL_IMAGE:figures/full_fig_p049_13.png]
Figure 14
Figure 14. Figure 14: Heatmaps illustrating the evolution of α parameters in a continual learning setting on the Shuffle CIFAR10 benchmark. The x-axis represents training steps, and the y-axis denotes the values of the α parameters. • The fraction of dormant neurons in the MainNetwork acco…
Figure 15
Figure 15. Figure 15: Heatmaps illustrating the evolution of α parameters in a continual learning setting on the Random-label CIFAR10 benchmark. The x-axis represents training steps, and the y-axis denotes the values of the α parameters. H.3 A Discussion of the Relationship Between αARM an…
Figure 16
Figure 16. Figure 16: The ℓ2 norm of the weights across two layers and two benchmarks is shown. As training progresses, the weight norm of MainNetwork continues to grow, whereas the baseline remains on a plateau, suggesting a loss of adaptability. effective downscaling of the learning rate…
Figure 17
Figure 17. Figure 17: Spectral rank of weights across two fully connected layers and two benchmarks is shown. The [PITH_FULL_IMAGE:figures/full_fig_p054_17.png]
Figure 18
Figure 18. Figure 18: Feature rank across two fully connected layers and two benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p055_18.png]
Figure 19
Figure 19. Figure 19: Dead neuron ratio across two fully connected layers and two benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p056_19.png]
Figure 20
Figure 20. Figure 20: αARM and activation of InferenceNetwork in both first and second layer of MLP in Random-label CIFAR10 benchmark. 57 [PITH_FULL_IMAGE:figures/full_fig_p057_20.png]
Figure 21
Figure 21. Figure 21: αARM and activation of InferenceNetwork in both first and second layer of MLP in Shuffle CIFAR10 benchmark. 58 [PITH_FULL_IMAGE:figures/full_fig_p058_21.png]
Figure 22
Figure 22. Figure 22: Effect of αAL modulation on gradients in each layer for two benchmarks, Random-label CIFAR10 and Shuffle CIFAR10. Both the gradient magnitude and the fraction of zero gradients are shown [PITH_FULL_IMAGE:figures/full_fig_p059_22.png]
Figure 23
Figure 23. Figure 23: Cosine similarity between weights of the same layers across consecutive tasks, shown separately for [PITH_FULL_IMAGE:figures/full_fig_p059_23.png]
Figure 24
Figure 24. Figure 24: A similar change in parameter direction to that observed in [PITH_FULL_IMAGE:figures/full_fig_p060_24.png]
Figure 25
Figure 25. Figure 25: Both layers of MainNetwork become increasingly similar in direction over time, suggesting that the weight directions remain effectively constant across tasks and that the gradients are largely aligned throughout training. reflected in a lower ℓ2 norm for the second la…
Figure 26
Figure 26. Figure 26: In terms of cosine similarity, the MainNetwork and the ConsolidatedNetwork converge to nearly the same direction. In terms of magnitude, their difference is more pronounced in the first layer, whereas for the second layer (the one we focus on due to its changing direc…
Figure 27
Figure 27. Figure 27: The direction of the vector of αWC values across neurons in each layer remains almost unchanged (cosine similarity close to 1), indicating that the modulation direction does not vary significantly from one task to the next. This effect is particularly pronounced in th…
Figure 28
Figure 28. Figure 28: The ℓ2 norms of the modulation vectors and MainNetwork change from one task to the next for the second layer. This contrasts with their directions, which remain almost unchanged across tasks. results are illustrated in [PITH_FULL_IMAGE:figures/full_fig_p061_28.png]
Figure 29
Figure 29. Figure 29: For the second layer, InferenceNetwork exhibits a smaller ℓ2 norm when αWC is enabled than when it is disabled. A lower magnitude implies that the denominator of Equation 11 is closer to the near-cancellation regime, where small changes in the numerator’s magnitude ca…
Figure 30
Figure 30. Figure 30: Evolution of the angle between InferenceNetwork and MainNetwork across two layers, measured at the end of each task. While the angle in the first layer remains relatively constant, it shifts significantly between tasks in the second layer, even as MainNetwork and Cons…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

115 extracted references · 42 canonical work pages

  1. [1]

    arXiv preprint arXiv:1607.06450 , year=

    Layer normalization , author=. arXiv preprint arXiv:1607.06450 , year=

    arXiv

  2. [2]

    international conference on machine learning , pages=

    Understanding and improving convolutional neural networks via concatenated rectified linear units , author=. international conference on machine learning , pages=. 2016 , organization=

    2016

  3. [3]

    arXiv preprint arXiv:2002.09571 , year=

    Learning to continually learn , author=. arXiv preprint arXiv:2002.09571 , year=

    arXiv 2002

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Afec: Active forgetting of negative transfer in continual learning , author=. Advances in Neural Information Processing Systems , volume=

  5. [5]

    Brain research reviews , volume=

    The locus coeruleus--noradrenergic system: modulation of behavioral state and state-dependent cognitive processes , author=. Brain research reviews , volume=. 2003 , publisher=

    2003

  6. [6]

    2013 , publisher=

    Matrix computations , author=. 2013 , publisher=

    2013

  7. [7]

    APL Machine Learning , volume=

    Brain-inspired learning in artificial neural networks: a review , author=. APL Machine Learning , volume=. 2024 , publisher=

    2024

  8. [8]

    arXiv preprint arXiv:2505.14125 , year=

    Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning , author=. arXiv preprint arXiv:2505.14125 , year=

    work page arXiv

Show all 115 references
  1. [9]

    Trends in cognitive sciences , volume=

    What learning systems do intelligent agents need? Complementary learning systems theory updated , author=. Trends in cognitive sciences , volume=. 2016 , publisher=

    2016

  2. [10]

    PloS one , volume=

    Norepinephrine drives persistent activity in prefrontal cortex via synergistic 1 and 2 adrenoceptors , author=. PloS one , volume=. 2013 , publisher=

    2013

  3. [11]

    Frontiers in neuroanatomy , volume=

    Layer-and cell type-specific modulation of excitatory neuronal activity in the neocortex , author=. Frontiers in neuroanatomy , volume=. 2018 , publisher=

    2018

  4. [12]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Film: Visual reasoning with a general conditioning layer , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  5. [13]

    Advances in neural information processing systems , volume=

    Matching networks for one shot learning , author=. Advances in neural information processing systems , volume=

  6. [14]

    Nature reviews neuroscience , volume=

    The locus coeruleus and noradrenergic modulation of cognition , author=. Nature reviews neuroscience , volume=. 2009 , publisher=

    2009

  7. [15]

    arXiv preprint arXiv:2107.14795 , year=

    Perceiver io: A general architecture for structured inputs & outputs , author=. arXiv preprint arXiv:2107.14795 , year=

    arXiv

  8. [16]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    arXiv

  9. [17]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    2010 arXiv

  10. [18]

    , author=

    The perceptron: a probabilistic model for information storage and organization in the brain. , author=. Psychological review , volume=. 1958 , publisher=

    1958

  11. [19]

    Frontiers in human neuroscience , volume=

    Effects of repetition learning on associative recognition over time: Role of the hippocampus and prefrontal cortex , author=. Frontiers in human neuroscience , volume=. 2018 , publisher=

    2018

  12. [20]

    The bulletin of mathematical biophysics , volume=

    A logical calculus of the ideas immanent in nervous activity , author=. The bulletin of mathematical biophysics , volume=. 1943 , publisher=

    1943

  13. [21]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    arXiv

  14. [22]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Emerging properties in self-supervised vision transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  15. [23]

    Advances in Neural Information Processing Systems , volume=

    Yolov10: Real-time end-to-end object detection , author=. Advances in Neural Information Processing Systems , volume=

  16. [24]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [25]

    arXiv preprint arXiv:2307.15818 , year=

    Rt-2: Vision-language-action models transfer web knowledge to robotic control , author=. arXiv preprint arXiv:2307.15818 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2205.09991 , year=

    Planning with diffusion for flexible behavior synthesis , author=. arXiv preprint arXiv:2205.09991 , year=

    arXiv

  19. [27]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    A survey on multimodal large language models for autonomous driving , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  20. [28]

    Future healthcare journal , volume=

    Artificial intelligence in healthcare: transforming the practice of medicine , author=. Future healthcare journal , volume=. 2021 , publisher=

    2021

  21. [29]

    Journal of Artificial Intelligence Research , volume=

    Towards continual reinforcement learning: A review and perspectives , author=. Journal of Artificial Intelligence Research , volume=

  22. [30]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    A comprehensive survey of continual learning: Theory, method and application , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  23. [31]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Riemannian walk for incremental learning: Understanding forgetting and intransigence , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  24. [32]

    Nature , volume=

    Loss of plasticity in deep continual learning , author=. Nature , volume=. 2024 , publisher=

    2024

  25. [33]

    International conference on machine learning , pages=

    Overcoming catastrophic forgetting with hard attention to the task , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  26. [34]

    Proceedings of the national academy of sciences , volume=

    Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the national academy of sciences , volume=. 2017 , publisher=

    2017

  27. [35]

    IEEE/CAA Journal of Automatica Sinica , volume=

    A survey on negative transfer , author=. IEEE/CAA Journal of Automatica Sinica , volume=. 2022 , publisher=

    2022

  28. [36]

    Advances in Neural Information Processing Systems , volume=

    Beyond not-forgetting: Continual learning with backward knowledge transfer , author=. Advances in Neural Information Processing Systems , volume=

  29. [37]

    Current opinion in neurobiology , volume=

    Neuromodulation of neurons and synapses , author=. Current opinion in neurobiology , volume=. 2014 , publisher=

    2014

  30. [38]

    Neuron , volume=

    Single cortical neurons as deep artificial neural networks , author=. Neuron , volume=. 2021 , publisher=

    2021

  31. [39]

    Nature Communications , volume=

    Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics , author=. Nature Communications , volume=. 2024 , publisher=

    2024

  32. [40]

    Nature , volume=

    Dendritic spikes enhance stimulus selectivity in cortical neurons in vivo , author=. Nature , volume=. 2013 , publisher=

    2013

  33. [41]

    Nature , volume=

    Dendritic spikes as a mechanism for cooperative long-term potentiation , author=. Nature , volume=. 2002 , publisher=

    2002

  34. [42]

    Science , volume=

    Glia as architects of central nervous system formation and function , author=. Science , volume=. 2018 , publisher=

    2018

  35. [43]

    Cold Spring Harbor perspectives in biology , volume=

    Homeostatic synaptic plasticity: local and global mechanisms for stabilizing neuronal function , author=. Cold Spring Harbor perspectives in biology , volume=. 2012 , publisher=

    2012

  36. [44]

    International Conference on Machine Learning , pages=

    The dormant neuron phenomenon in deep reinforcement learning , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  37. [45]

    Cold Spring Harbor perspectives in biology , volume=

    Memory consolidation , author=. Cold Spring Harbor perspectives in biology , volume=. 2015 , publisher=

    2015

  38. [46]

    Trends in cognitive sciences , volume=

    ‘What’is happening in the dorsal visual pathway , author=. Trends in cognitive sciences , volume=. 2016 , publisher=

    2016

  39. [47]

    Brain Connectivity , volume=

    Default mode network engagement beyond self-referential internal mentation , author=. Brain Connectivity , volume=. 2018 , publisher=

    2018

  40. [48]

    neuron , volume=

    Neuromodulation of brain states , author=. neuron , volume=. 2012 , publisher=

    2012

  41. [49]

    International Conference on Machine Learning , pages=

    Understanding plasticity in neural networks , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  42. [50]

    arXiv preprint arXiv:2402.18762 , year=

    Disentangling the causes of plasticity loss in neural networks , author=. arXiv preprint arXiv:2402.18762 , year=

    arXiv

  43. [51]

    Plastic Learning with Deep Fourier Features , author=

  44. [52]

    Conference on lifelong learning agents , pages=

    Loss of plasticity in continual deep reinforcement learning , author=. Conference on lifelong learning agents , pages=. 2023 , organization=

    2023

  45. [53]

    arXiv preprint arXiv:2308.11958 , year=

    Maintaining plasticity in continual learning via regenerative regularization , author=. arXiv preprint arXiv:2308.11958 , year=

    arXiv

  46. [54]

    Advances in neural information processing systems , volume=

    On warm-starting neural network training , author=. Advances in neural information processing systems , volume=

  47. [55]

    Advances in Neural Information Processing Systems , volume=

    Non-stationary learning of neural networks with automatic soft parameter reset , author=. Advances in Neural Information Processing Systems , volume=

  48. [56]

    arXiv preprint arXiv:2410.07994 , year=

    Neuroplastic Expansion in Deep Reinforcement Learning , author=. arXiv preprint arXiv:2410.07994 , year=

    arXiv

  49. [57]

    International conference on machine learning , pages=

    Continual learning through synaptic intelligence , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  50. [58]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Measuring catastrophic forgetting in neural networks , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  51. [59]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Memory aware synapses: Learning what (not) to forget , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  52. [60]

    Advances in neural information processing systems , volume=

    Continual learning with deep generative replay , author=. Advances in neural information processing systems , volume=

  53. [61]

    arXiv preprint arXiv:1606.04671 , year=

    Progressive neural networks , author=. arXiv preprint arXiv:1606.04671 , year=

    arXiv

  54. [62]

    Proceedings of the IEEE , volume=

    Gradient-based learning applied to document recognition , author=. Proceedings of the IEEE , volume=. 1998 , publisher=

    1998

  55. [63]

    2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    BIMRL: Brain Inspired Meta Reinforcement Learning , author=. 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2022 , organization=

    2022

  56. [64]

    Machine Learning , volume=

    Inductive biases for zero-shot systematic generalization in language-informed reinforcement learning , author=. Machine Learning , volume=. 2025 , publisher=

    2025

  57. [65]

    Nature communications , volume=

    Brain-inspired replay for continual learning with artificial neural networks , author=. Nature communications , volume=. 2020 , publisher=

    2020

  58. [66]

    arXiv preprint arXiv:2408.14811 , year=

    Brain-inspired Artificial Intelligence: A Comprehensive Review , author=. arXiv preprint arXiv:2408.14811 , year=

    arXiv

  59. [67]

    Nature communications , volume=

    Random synaptic feedback weights support error backpropagation for deep learning , author=. Nature communications , volume=. 2016 , publisher=

    2016

  60. [68]

    Advances in Neural Information Processing Systems , volume=

    Structured flexibility in recurrent neural networks via neuromodulation , author=. Advances in Neural Information Processing Systems , volume=

  61. [69]

    PloS one , volume=

    Introducing neuromodulation in deep neural networks to learn adaptive behaviours , author=. PloS one , volume=. 2020 , publisher=

    2020

  62. [70]

    arXiv preprint arXiv:2002.10585 , year=

    Backpropamine: training self-modifying neural networks with differentiable neuromodulated plasticity , author=. arXiv preprint arXiv:2002.10585 , year=

    2002 arXiv

  63. [71]

    International Conference on Machine Learning , pages=

    Differentiable plasticity: training plastic neural networks with backpropagation , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  64. [72]

    Duan, Yan and Schulman, John and Chen, Xi and Bartlett, Peter L and Sutskever, Ilya and Abbeel, Pieter , journal=. Rl \^

  65. [73]

    Advances in neural information processing systems , volume=

    Experience replay for continual learning , author=. Advances in neural information processing systems , volume=

  66. [74]

    bioRxiv , pages=

    Neuromodulators generate multiple context-relevant behaviors in a recurrent neural network by shifting activity flows in hyperchannels , author=. bioRxiv , pages=. 2021 , publisher=

    2021

  67. [75]

    Advances in Neural Information Processing Systems , volume=

    Biologically-plausible backpropagation through arbitrary timespans via local neuromodulators , author=. Advances in Neural Information Processing Systems , volume=

  68. [76]

    Frontiers in neural circuits , volume=

    Neuromodulated spike-timing-dependent plasticity, and theory of three-factor learning rules , author=. Frontiers in neural circuits , volume=. 2016 , publisher=

    2016

  69. [77]

    Scientific reports , volume=

    Homeostatic plasticity and burst activity are mediated by hyperpolarization-activated cation currents and T-type calcium channels in neuronal cultures , author=. Scientific reports , volume=. 2021 , publisher=

    2021

  70. [78]

    1 potassium channels contribute to the slow afterhyperpolarization in L5 neocortical pyramidal neurons , author=

    Ca2+-activated KCa3. 1 potassium channels contribute to the slow afterhyperpolarization in L5 neocortical pyramidal neurons , author=. Scientific Reports , volume=. 2020 , publisher=

    2020

  71. [79]

    Current opinion in neurobiology , volume=

    Plasticity of intrinsic neuronal excitability , author=. Current opinion in neurobiology , volume=. 2019 , publisher=

    2019

  72. [80]

    Elife , volume=

    Minimal requirements for a neuron to coregulate many properties and the implications for ion channel correlations and robustness , author=. Elife , volume=. 2022 , publisher=

    2022

  73. [81]

    Neuroscience research , volume=

    Synaptic plasticity during systems memory consolidation , author=. Neuroscience research , volume=. 2022 , publisher=

    2022

  74. [82]

    Cell reports , volume=

    Brain state dependence of hippocampal subthreshold activity in awake mice , author=. Cell reports , volume=. 2017 , publisher=

    2017

  75. [83]

    arXiv preprint arXiv:2007.06712 , year=

    Embedded encoder-decoder in convolutional networks towards explainable AI , author=. arXiv preprint arXiv:2007.06712 , year=

    2007 arXiv

  76. [84]

    Frontiers in Physiology , volume=

    Tonic neuromodulation of the inspiratory rhythm generator , author=. Frontiers in Physiology , volume=. 2012 , publisher=

    2012

  77. [85]

    Trends in cognitive sciences , volume=

    Embracing change: Continual learning in deep neural networks , author=. Trends in cognitive sciences , volume=. 2020 , publisher=

    2020

  78. [86]

    Neural Networks , volume=

    Continual learning in the presence of repetition , author=. Neural Networks , volume=. 2025 , publisher=

    2025

  79. [87]

    Proceedings of the IEEE international conference on computer vision , pages=

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification , author=. Proceedings of the IEEE international conference on computer vision , pages=

  80. [88]

    arXiv preprint arXiv:2505.09486 , year=

    Preserving Plasticity in Continual Learning with Adaptive Linearity Injection , author=. arXiv preprint arXiv:2505.09486 , year=

    arXiv

  81. [89]

    arXiv preprint arXiv:1812.00420 , year=

    Efficient lifelong learning with a-gem , author=. arXiv preprint arXiv:1812.00420 , year=

    arXiv

  82. [90]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    DKT: Diverse knowledge transfer transformer for class incremental learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  83. [91]

    Psychological Review , volume=

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory , author=. Psychological Review , volume=. 1995 , publisher=

    1995

  84. [92]

    Neural Networks , volume=

    Continual lifelong learning with neural networks: A review , author=. Neural Networks , volume=. 2019 , publisher=

    2019

  85. [93]

    Brain Sciences , volume=

    Neuromodulators and Long-Term Synaptic Plasticity in Learning and Memory: A Steered-Glutamatergic Perspective , author=. Brain Sciences , volume=. 2019 , publisher=. doi:10.3390/brainsci9110300 , url=

    2019 doi

  86. [94]

    Nature , volume=

    A distributional code for value in dopamine-based reinforcement learning , author=. Nature , volume=. 2020 , publisher=

    2020

  87. [95]

    Proceedings of the 34th International Conference on Machine Learning , pages=

    The shattered gradients problem: If resnets are the answer, then what is the question? , author=. Proceedings of the 34th International Conference on Machine Learning , pages=

  88. [96]

    International conference on machine learning , pages=

    Model-agnostic meta-learning for fast adaptation of deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  89. [97]

    Current Opinion in Neurobiology , volume=

    Plasticity of intrinsic neuronal excitability , author=. Current Opinion in Neurobiology , volume=

  90. [98]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Overcoming generic knowledge loss with selective parameter update , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  91. [99]

    2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Rapid adaptation in online continual learning: Are we evaluating it right? , author=. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages=. 2023 , organization=

    2023

  92. [100]

    2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ ICML 2024) , year=

    Dash: Warm-starting neural network training without loss of plasticity under stationarity , author=. 2nd Workshop on Advancing Neural Network Training: Computational Efficiency, Scalability, and Resource Optimization (WANT@ ICML 2024) , year=

    2024

  93. [101]

    Advances in Neural Information Processing Systems , volume=

    On the stability-plasticity dilemma in continual meta-learning: Theory and algorithm , author=. Advances in Neural Information Processing Systems , volume=

  94. [102]

    Advances in Neural Information Processing Systems , volume=

    Optimizing reusable knowledge for continual learning via metalearning , author=. Advances in Neural Information Processing Systems , volume=

  95. [103]

    Advances in Neural Information Processing Systems , volume=

    Learning where to learn: Gradient sparsity in meta and continual learning , author=. Advances in Neural Information Processing Systems , volume=

  96. [104]

    Advances in Neural Information Processing Systems , volume=

    Online fast adaptation and knowledge accumulation (osaka): a new approach to continual learning , author=. Advances in Neural Information Processing Systems , volume=

  97. [105]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    When meta-learning meets online and continual learning: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  98. [106]

    Nature neuroscience , volume=

    Organizing memories for generalization in complementary learning systems , author=. Nature neuroscience , volume=. 2023 , publisher=

    2023

  99. [107]

    arXiv preprint arXiv:2406.02596 , year=

    Slow and steady wins the race: Maintaining plasticity with hare and tortoise networks , author=. arXiv preprint arXiv:2406.02596 , year=

    arXiv

  100. [108]

    Conference on lifelong learning agents , pages=

    Synergy between synaptic consolidation and experience replay for general continual learning , author=. Conference on lifelong learning agents , pages=. 2022 , organization=

    2022

  101. [109]

    arXiv preprint arXiv:2201.12604 , year=

    Learning fast, learning slow: A general continual learning method based on complementary learning system , author=. arXiv preprint arXiv:2201.12604 , year=

    arXiv

  102. [110]

    arXiv preprint arXiv:1609.09106 , year=

    Hypernetworks , author=. arXiv preprint arXiv:1609.09106 , year=

    arXiv

  103. [111]

    arXiv preprint arXiv:1710.04759 , year=

    Bayesian hypernetworks , author=. arXiv preprint arXiv:1710.04759 , year=

    arXiv

  104. [112]

    arXiv preprint arXiv:1711.01791 , year=

    HyperNetworks with statistical filtering for defending adversarial examples , author=. arXiv preprint arXiv:1711.01791 , year=

    arXiv

  105. [113]

    Artificial Intelligence Review , volume=

    A brief review of hypernetworks in deep learning , author=. Artificial Intelligence Review , volume=. 2024 , publisher=

    2024

  106. [114]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  107. [115]

    arXiv preprint arXiv:2505.14511 , year=

    ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains , author=. arXiv preprint arXiv:2505.14511 , year=

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.