Pith. sign in

REVIEW 3 major objections 5 minor 34 references

A gradient-trained policy network can grow and prune its own architecture during learning, using only local weight-update statistics, and match or beat fixed-width MLPs on control tasks.

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 ↗

A gradient-trained control network whose size adjusts online through a local structural plasticity module matches or beats fixed-size MLPs on three control benchmarks.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A modest but genuine new local rule for growing/pruning network capacity during policy training; the cyclic-graph backprop ambiguity is real and needs fixing before the empirical claims are fully reproducible. the 3 major comments →

arxiv 2512.12713 v2 pith:4KTMQLCZ submitted 2025-12-14 cs.NE cs.LG

Self-Motivated Growing Neural Network for Adaptive Architecture via Local Structural Plasticity

Yiyang Jia , Chengxu Zhou This is my paper
classification cs.NE cs.LG
keywords structural plasticitygrowing neural networkspolicy distillationreinforcement learning controladaptive network capacitylocal learning rulesnetwork pruning
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

The paper tries to show that a neural controller can adjust its own capacity during learning without manual architecture selection. It introduces SMGrNN, a graph network whose topology is edited online by a Structural Plasticity Module (SPM) that uses only short-window statistics of each edge's weight updates. When an edge's updates fluctuate around zero with high variance, the SPM inserts a relay node; it also adds random exploratory edges and prunes weak or orphaned structure. Trained by policy distillation on three control benchmarks, SMGrNN matches or exceeds fixed-width MLPs' returns with lower variance and task-appropriate parameter counts. If true, this would make a case for local structural plasticity as a complement to gradient-based weight learning.

Core claim

The paper presents SMGrNN, a directed graph in which hidden nodes and edges are inserted or removed during training by a Structural Plasticity Module (SPM), while synaptic weights are optimized by ordinary backpropagation. The SPM flags an edge as unstable when the mean of its weight updates over a rolling window is within half a standard deviation of zero and its variance is large relative to that mean; it then inserts a relay node, creating a parallel two-step path alongside the original edge. Random exploratory growth adds extra edges from a candidate pool, and periodic pruning deletes small-magnitude, stalled edges and orphan nodes. In policy-distillation experiments across CartPole, Acr

What carries the argument

The load-bearing mechanism is the local instability criterion in Eq. (2): an edge i->j is deemed unstable when |mean(Δw)| < 0.5·std(Δw) and variance(Δw) > λ_edge·|mean(Δw)| over a sliding window of T steps. This condition detects edges whose gradient updates 'flicker' around zero without settling in a direction, and it triggers insertion of a relay node — locally increasing representational capacity exactly where the optimizer is uncertain. Random exploratory growth (Bernoulli-triggered addition of edges from the candidate pool) diversifies topology, and pruning (weight-magnitude and update-stall thresholds plus orphan-node removal) enforces compactness. The SPM uses only these local signals

Load-bearing premise

The load-bearing premise is that the training procedure (backpropagation through the current graph) is well-defined for the graphs the SPM produces, but the paper explicitly permits cycles while never specifying how a cyclic graph is evaluated or differentiated; if cycles arise, the learning curves rest on an unspecified computation.

What would settle it

Inspect the learned SMGrNN graphs after training: count directed cycles. If any exist and the reported training used single-pass backpropagation without unrolling or iterative fixed-point solving, then either the graphs must in fact be acyclic (contradicting the stated absence of constraints) or the results are not reproducible as described; additionally, re-running the growth-only condition with the same 2000-episode budget as the full model would test whether the pruning advantage is partly a training-length artifact.

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

If this is right

  • On CartPole and Acrobot, SMGrNN reaches higher asymptotic rewards with faster learning and lower variance than static MLPs at similar or smaller parameter counts; on LunarLander it reaches strong performance earlier than the widest MLP while using roughly half the parameters.
  • The growth-only ablation shows that pruning is load-bearing for compactness: without it, the network expands to thousands of parameters while keeping similar late-training rewards, so the growth rule alone does not self-terminate.
  • Final hidden-node count scales with task difficulty (XOR < CartPole < Acrobot < LunarLander) under a single fixed set of SPM hyperparameters, indicating the topology self-matches representational demand.
  • Initial hidden-node counts of 0, 5, or 10 converge to similar final sizes, so the mechanism compensates for over- or under-parameterized starting graphs without manual reconfiguration.

Where Pith is reading between the lines

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

  • The instability criterion is a local, threshold-driven measure of 'flickering' gradient updates; one could test whether growth events occur near a self-organized critical point of the weight dynamics, and whether λ_edge controls a phase transition between no growth and runaway growth.
  • Because the SPM uses only local statistics, replacing gradient weight updates with a Hebbian or spike-timing-dependent rule should preserve the structural dynamics; a direct test would be to swap the optimizer for a local rule and see if task-appropriate sizes still emerge.
  • The paper's allowance of cycles alongside ordinary backpropagation is underspecified; checking whether learned graphs actually contain cycles, and if so how gradients are computed, would clarify whether the reported behavior is an artifact of an implicit acyclic implementation.
  • The growth-only and full SMGrNN were trained for different episode budgets (1000 vs 2000); re-running growth-only for 2000 episodes would isolate how much of pruning's benefit is due to training length rather than structural removal.
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 SMGrNN, a policy network whose directed graph topology is modified online by a Structural Plasticity Module (SPM) that uses local, short-term statistics of activations and weight updates to insert and prune edges and hidden neurons, while weights are trained by backpropagation on a policy-distillation objective. The method is evaluated on CartPole-v1, Acrobot-v1, and LunarLander-v3 against fixed-width MLP baselines, and through ablations that disable or restrict growth and pruning. The central claim is that the SPM enables competitive or superior control performance with similar or smaller parameter budgets, lower across-seed variance, and task-appropriate final network sizes, without task-specific architectural tuning. The paper also reports a scaling result across XOR and the three control tasks, and it explicitly limits its scope to within-task capacity adaptation, acknowledging that synaptic learning remains gradient-based and the SPM hyperparameters are hand-tuned.

Significance. If the central claim holds, the paper makes a useful contribution to growing-neural-network and local-plasticity research by showing that a simple, local structural rule can adjust capacity during gradient-based training and yield favorable reward/variance properties relative to fixed architectures. The study has several strengths: ten-seed runs for the main comparisons, explicit hyperparameter reporting (Table A1), ablation designs that separate growth and pruning, and an unusually honest limitation section. However, the manuscript's central correctness is currently at risk because the training procedure is underspecified for the cyclic graphs that the SPM is explicitly allowed to generate. The paper also lacks inferential statistics and relies on a single-run growth-only ablation. These issues are fixable within the manuscript's scope, but they are load-bearing and require a major revision.

major comments (3)
  1. [III-B / III-E] The training algorithm is underspecified for the graphs the SPM is allowed to produce. Section III-B states 'No feedforward or acyclicity constraint is imposed on E_t, so cycles and skip connections can emerge among all node groups during learning,' while Section III-E says weights are updated by 'backpropagation through the current graph G_t' and forward propagation is described only as 'standard affine–nonlinear transformations along the edges.' For a directed graph containing a cycle, neither forward evaluation nor backpropagation is well-defined without unrolling the graph over time, solving a fixed point, or imposing an ordering that breaks the cycle. Cycles are not merely hypothetical: Eq. (4) samples uniformly from V_t × V_t without excluding reverse edges, so a 2-cycle can appear whenever (i,j) and (j,i) both exist; relay insertion (Section III-C1) can likewise create a 2-cycle w
  2. [IV-C / Table A2 / Table II] Claims of lower variance, higher rewards, and more consistent convergence are not supported by inferential statistics. Table A2 reports means ± std over 10 seeds; with n=10, none of the differences (e.g., Acrobot -94.0±1.9 vs -95.9±5.5; LunarLander 207.6±33.3 vs 145.7±130.5) are assessed for significance or effect size, and the shaded regions in Fig. 3 overlap for several baselines. More seriously, the pruning ablation (Section IV-C2, Table II) is based on a single run per environment, so the conclusion that pruning 'is critical for preventing uncontrolled network expansion' (Section V-A) rests on one trajectory. I request a proper statistical treatment, such as paired tests or bootstrap confidence intervals across the ten seeds, and multi-seed growth-only runs, or a proportionate rephrasing of the claims.
  3. [IV-A2 / IV-B] The experimental protocol for hyperparameters weakens the clean evaluation on CartPole. Section IV-A2 states 'All SPM hyperparameters are chosen once based on preliminary tuning on CartPole-v1 and kept fixed for all environments,' and Section IV-B then reports CartPole as evidence that SMGrNN 'reaches higher asymptotic rewards with faster learning and lower across-seed variance than any static MLP.' Since CartPole was used for tuning, those CartPole results are not a prediction from an untuned configuration; the 'no manual tuning' claim is strongest for Acrobot and LunarLander and weaker for CartPole. Please either show that the qualitative conclusions are insensitive to reasonable variations of the SPM hyperparameters (e.g., a small random search grid) or explicitly separate tuning and evaluation conditions.
minor comments (5)
  1. [IV-A2] The sentence 'Random exploratory growth ... with ... growth ratio ρ_rand as in Eq. (2)' should refer to Eq. (3)–(4); Eq. (2) is the instability criterion, not the growth ratio formula.
  2. [III-D] The reset of history buffers after each structural edit is not analyzed. Because buffers are reset whenever node or edge indexing changes, the temporal statistics in Eq. (1) may be systematically truncated by the very structural changes they trigger. Please clarify the practical effect and, if possible, provide a sensitivity check on the window length T=100.
  3. [Figs. 3–7] The shaded regions are described as spanning 'all runs' or 'across-run variability,' but the exact statistic (min–max, standard deviation, or percentile interval) is not defined. Please state this explicitly in the captions or experimental setup.
  4. [Appendix Table A4] The XOR task is included in the scaling analysis, but its initial hidden count, number of episodes, and training procedure are not specified in the table or the main text. Please state these details.
  5. [Reproducibility] No code or data repository is mentioned. Given that the paper's central reproducibility depends on the unspecified handling of possibly cyclic graphs, a code release or a precise algorithmic description of the forward/backward pass would be especially valuable.

Circularity Check

0 steps flagged

No circular derivation: SPM growth/pruning is empirical and not a reduction to its own inputs.

full rationale

The paper's central claim is that the SPM's local instability trigger (Eq. 2) plus random growth and pruning produces adaptive architectures with competitive returns and task-appropriate sizes. None of the equations define the reported outcomes in terms of the fitted constants. Eq. 2 uses local mean/variance of weight updates to trigger relay insertion, Eqs. 3–4 define random growth, Eqs. 5–6 define pruning, and Eq. 7 is the schedule. The final network size and reward are emergent empirical quantities, not algebraic consequences of the SPM hyperparameters. No self-citation is load-bearing: the cited dynamic-expansion works ([22], [23]) are not by the authors and are used only as related work. The only tuning concern is explicit and limited: the paper states in Section IV-A2 that "All SPM hyperparameters are chosen once based on preliminary tuning on CartPole-v1 and kept fixed for all environments; their numerical values are summarized in Table A1." This means CartPole is not fully out-of-sample, and Section V-C acknowledges the SPM is "hand-designed: thresholds, sampling ratios, and update schedules are selected by preliminary tuning on a small set of environments." However, this is model selection, not circular derivation: the tuned hyperparameters do not by construction force the reported reward curves or final sizes, which still depend on gradient optimization and the stochastic structural dynamics. The unresolved cyclic-graph issue (Section III-B permits cycles while Section III-E assumes backpropagation through the current graph) is a reproducibility/correctness concern, not a circularity one. The derivation is therefore self-contained with respect to its empirical claims; no fitted quantity is renamed as a prediction.

Axiom & Free-Parameter Ledger

10 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical entities. The load-bearing assumptions are the instability heuristic, the well-definedness of gradient flow on cyclic graphs, the validity of policy distillation, and the compatibility of random exploration with the 'self-motivated' claim. All SPM hyperparameters are hand-chosen or tuned on CartPole and then frozen.

free parameters (10)
  • T (temporal window length) = 100
    Chosen by preliminary tuning on CartPole; used in Eq. (1) to compute update statistics.
  • p_rand (random growth probability) = 0.25
    Hand-selected SPM hyperparameter controlling Bernoulli trigger for exploratory growth.
  • rho_rand (random growth ratio) = 0.01
    Hand-selected; controls number of random edges added per growth event.
  • eta_prune (pruned weak-edge fraction) = 0.8
    Hand-selected; controls fraction of weak edges deleted each pruning step.
  • s (pruning period) = 30
    Hand-selected global schedule step determining pruning frequency.
  • tau_w (weight threshold) = 0.1
    Hand-selected; defines which weights count as small for pruning.
  • tau_delta (derivative threshold) = 1e-4
    Hand-selected; defines when weight updates are considered stalled.
  • lambda_edge (variance scaling) = 0.1
    Hand-selected; scales the instability criterion in Eq. (2).
  • initial connection density = 0.8
    Chosen by hand as the initial edge density over allowed node pairs.
  • optimizer hyperparameters (e.g., learning rate) = not reported
    The paper says 'standard gradient-based optimizer' but gives no learning rate or optimizer-specific settings, making exact replication impossible.
axioms (4)
  • domain assumption Weight-update fluctuation signals representational instability.
    Eq. (2) treats high-variance, near-zero-mean weight updates as a trigger for relay insertion; the paper does not prove that inserting a relay is the right response to this signal.
  • domain assumption Backpropagation through the current graph is well-defined even when the graph contains cycles.
    Section III-B allows cycles, and Section III-E says weights are updated by backpropagation through the current graph, but no unrolling, ordering, or recurrent training procedure is specified.
  • domain assumption Supervised distillation from a frozen expert under the student's own state distribution yields the expert's control performance.
    Training uses only a supervised loss between student and expert outputs (Section III-E); no reward signal is used, so the method assumes behavior cloning does not diverge on the student's visited states.
  • ad hoc to paper Random exploratory growth is consistent with the 'self-motivated' framing.
    Section III-C2 adds edges via a Bernoulli random trigger, which is not driven by local activity statistics, despite the paper defining 'self-motivated' as changes driven solely by internally generated signals.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Motivated Growing Neural Network for Adaptive Architecture via Local Structural Plasticity." pith.science (2026). https://pith.science/paper/4KTMQLCZ

@misc{pith2026251212713,
  author       = {Pith},
  title        = {Pith review of: Self-Motivated Growing Neural Network for Adaptive Architecture via Local Structural Plasticity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4KTMQLCZ}},
  note         = {Machine review of arXiv:2512.12713}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Control policies are often implemented with fixed-capacity multilayer perceptrons trained by backpropagation, which require architecture selection in advance and cannot adapt their capacity during learning. This paper introduces the Self-Motivated Growing Neural Network (SMGrNN), a gradient-trained controller whose topology evolves online through a local Structural Plasticity Module (SPM). The SPM monitors edge-wise weight update statistics over short temporal windows and uses these local signals to trigger neuron insertion and pruning, while synaptic weights are optimized by a standard gradient-based optimizer. This allows network capacity to be adjusted during learning without manual architectural tuning. SMGrNN is evaluated on control benchmarks via policy distillation. Compared with multilayer perceptron baselines, it achieves similar or higher returns, lower variance, and task-appropriate network sizes. Ablation studies with growth disabled and growth-only variants isolate the role of structural plasticity, showing that adaptive growth improves reward stability while pruning prevents uncontrolled expansion and supports compact network formation. These results establish the independent value of local structural plasticity within gradient-trained networks and motivate future investigation of whether similar structural rules can be extended to more local or spike-based learning settings.

Figures

Figures reproduced from arXiv: 2512.12713 by Chengxu Zhou, Yiyang Jia.

Figure 1
Figure 1. Figure 1: Overview of the SMGrNN. The current graph [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Relay-node insertion on an unstable edge in a minimal [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Reward learning curves for SMGrNN and single-layer MLP baselines on three control tasks. Shaded regions indicate [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation on structural growth. Each panel compares SMGrNN (growth enabled) with a static graph in which the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Growth-only variant vs. SMGrNN. Panel (a) shows [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity to initial topology. Each panel averages over 10 runs with initial hidden-node counts [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scaling of SMGrNN capacity with task difficulty. Left: [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] 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

34 extracted references · 12 linked inside Pith

  1. [1]

    Biologically plausible local synaptic learning rules robustly implement deep supervised learning,

    M. Konishi, K. M. Igarashi, and K. Miura, “Biologically plausible local synaptic learning rules robustly implement deep supervised learning,” Frontiers in Neuroscience, vol. 17, p. 1160899, 2023. ARXIV PREPRINT 10

    2023

  2. [2]

    Towards biologically plausible deep learning,

    Y . Bengio, D.-H. Lee, J. Bornschein, T. Mesnard, and Z. Lin, “Towards biologically plausible deep learning,”arXiv preprint arXiv:1502.04156, 2015

    Pith/arXiv arXiv 2015

  3. [3]

    Biologically plausible reinforcement learning,

    T. Stegmaier, “Biologically plausible reinforcement learning,” 2023

    2023

  4. [4]

    Reinforcement learning: Computational theory and biological mechanisms,

    K. Doya, “Reinforcement learning: Computational theory and biological mechanisms,”HFSP journal, vol. 1, no. 1, p. 30, 2007

    2007

  5. [5]

    D. O. Hebb,The organization of behavior: A neuropsychological theory. Psychology press, 2005

    2005

  6. [6]

    Hebbian deep learning without feedback,

    A. Journ ´e, H. G. Rodriguez, Q. Guo, and T. Moraitis, “Hebbian deep learning without feedback,”arXiv preprint arXiv:2209.11883, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Pytorch-Hebbian: fa- cilitating local learning in a deep learning framework,

    J. Talloen, J. Dambre, and A. Vandesompele, “Pytorch-Hebbian: fa- cilitating local learning in a deep learning framework,”arXiv preprint arXiv:2102.00428, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,”Neural networks, vol. 113, pp. 54–71, 2019

    2019

  9. [9]

    Continual learning and catastrophic forgetting,

    G. M. van de Ven, N. Soures, and D. Kudithipudi, “Continual learning and catastrophic forgetting,”arXiv preprint arXiv:2403.05175, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  11. [11]

    A comprehensive survey of forgetting in deep learning beyond continual learning,

    Z. Wang, E. Yang, L. Shen, and H. Huang, “A comprehensive survey of forgetting in deep learning beyond continual learning,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024

    2024

  12. [12]

    Exploring the role of neuroplasticity in development, aging, and neu- rodegeneration,

    P. Marzola, T. Melzer, E. Pavesi, J. Gil-Mohapel, and P. S. Brocardo, “Exploring the role of neuroplasticity in development, aging, and neu- rodegeneration,”Brain sciences, vol. 13, no. 12, p. 1610, 2023

    2023

  13. [13]

    A history of spike-timing- dependent plasticity,

    H. Markram, W. Gerstner, and P. J. Sj ¨ostr¨om, “A history of spike-timing- dependent plasticity,”Frontiers in synaptic neuroscience, vol. 3, p. 4, 2011

    2011

  14. [14]

    Toward training recurrent neural networks for lifelong learning,

    S. Sodhani, S. Chandar, and Y . Bengio, “Toward training recurrent neural networks for lifelong learning,”Neural computation, vol. 32, no. 1, pp. 1–35, 2020

    2020

  15. [15]

    Chen and B

    Z. Chen and B. Liu,Lifelong machine learning. Morgan & Claypool Publishers, 2018

    2018

  16. [16]

    Understanding capac- ity saturation in incremental learning

    S. Huang, V . Francois-Lavet, and G. Rabusseau, “Understanding capac- ity saturation in incremental learning.” inCanadian AI, 2021

    2021

  17. [17]

    Complementary learning for overcoming catastrophic forgetting using experience replay,

    M. Rostami, S. Kolouri, and P. K. Pilly, “Complementary learning for overcoming catastrophic forgetting using experience replay,”arXiv preprint arXiv:1903.04566, 2019

    Pith/arXiv arXiv 1903

  18. [18]

    Expe- rience replay for continual learning,

    D. Rolnick, A. Ahuja, J. Schwarz, T. Lillicrap, and G. Wayne, “Expe- rience replay for continual learning,”Advances in neural information processing systems, vol. 32, 2019

    2019

  19. [19]

    Progressive neural networks,

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

    Pith/arXiv arXiv 2016

  20. [20]

    Lifelong learning with dynamically expandable networks,

    J. Yoon, E. Yang, J. Lee, and S. J. Hwang, “Lifelong learning with dynamically expandable networks,”arXiv preprint arXiv:1708.01547, 2017

    Pith/arXiv arXiv 2017

  21. [21]

    Neurogenesis deep learning: Extending deep networks to accommodate new classes,

    T. J. Draelos, N. E. Miner, C. C. Lamb, J. A. Cox, C. M. Vineyard, K. D. Carlson, W. M. Severa, C. D. James, and J. B. Aimone, “Neurogenesis deep learning: Extending deep networks to accommodate new classes,” in2017 international joint conference on neural networks (IJCNN). IEEE, 2017, pp. 526–533

    2017

  22. [22]

    Self-controlled dynamic ex- pansion model for continual learning,

    R. Wu, K. Huang, H. Zhang, and F. Ye, “Self-controlled dynamic ex- pansion model for continual learning,”arXiv preprint arXiv:2504.10561, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Self-evolved dynamic expansion model for task- free continual learning,

    F. Ye and A. G. Bors, “Self-evolved dynamic expansion model for task- free continual learning,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 22 102–22 112

    2023

  24. [24]

    Modular dynamic neural network: A continual learning architecture,

    D. Turner, P. J. Cardoso, and J. M. Rodrigues, “Modular dynamic neural network: A continual learning architecture,”Applied Sciences, vol. 11, no. 24, p. 12078, 2021

    2021

  25. [25]

    Progress & compress: A scalable framework for continual learning,

    J. Schwarz, W. Czarnecki, J. Luketina, A. Grabska-Barwinska, Y . W. Teh, R. Pascanu, and R. Hadsell, “Progress & compress: A scalable framework for continual learning,” inInternational conference on ma- chine learning. PMLR, 2018, pp. 4528–4537

    2018

  26. [26]

    A neural dirichlet pro- cess mixture model for task-free continual learning,

    S. Lee, J. Ha, D. Zhang, and G. Kim, “A neural dirichlet pro- cess mixture model for task-free continual learning,”arXiv preprint arXiv:2001.00689, 2020

    Pith/arXiv arXiv 2001

  27. [27]

    Enhancing efficient continual learning with dynamic structure development of spiking neural networks,

    B. Han, F. Zhao, Y . Zeng, W. Pan, and G. Shen, “Enhancing efficient continual learning with dynamic structure development of spiking neural networks,”arXiv preprint arXiv:2308.04749, 2023. TABLE A1: Hyperparameters of the SPM used in all experi- ments. Symbol Name Value TTemporal window length (steps)100 prand Random growth probability0.25 ρrand Random g...

    Pith/arXiv arXiv 2023

  28. [28]

    Evolving neural networks through augmenting topologies,

    K. O. Stanley and R. Miikkulainen, “Evolving neural networks through augmenting topologies,”Evolutionary computation, vol. 10, no. 2, pp. 99–127, 2002

    2002

  29. [29]

    Hypernca: Growing developmental networks with neural cellular automata,

    E. Najarro, S. Sudhakaran, C. Glanois, and S. Risi, “Hypernca: Growing developmental networks with neural cellular automata,”arXiv preprint arXiv:2204.11674, 2022

    Pith/arXiv arXiv 2022

  30. [30]

    Towards self-assembling artificial neural networks through neural developmental programs,

    E. Najarro, S. Sudhakaran, and S. Risi, “Towards self-assembling artificial neural networks through neural developmental programs,” in Artificial Life Conference Proceedings 35, vol. 2023, no. 1. MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info . . . , 2023, p. 80

    2023

  31. [31]

    Evolving self-assembling neural networks: From spontaneous activity to experience-dependent learning,

    E. Plantec, J. W. Pedersen, M. L. Montero, E. Nisioti, and S. Risi, “Evolving self-assembling neural networks: From spontaneous activity to experience-dependent learning,” inALIFE 2024: Proceedings of the 2024 Artificial Life Conference. MIT Press, 2024

    2024

  32. [32]

    Growing neural cellular automata,

    A. Mordvintsev, E. Randazzo, E. Niklasson, and M. Levin, “Growing neural cellular automata,”Distill, vol. 5, no. 2, p. e23, 2020

    2020

  33. [33]

    The forward-forward algorithm: Some preliminary investi- gations,

    G. Hinton, “The forward-forward algorithm: Some preliminary investi- gations,”arXiv preprint arXiv:2212.13345, vol. 2, no. 3, p. 5, 2022

    Pith/arXiv arXiv 2022

  34. [34]

    Heavy-tailed neuronal connectivity arises from Hebbian self-organization,

    C. W. Lynn, C. M. Holmes, and S. E. Palmer, “Heavy-tailed neuronal connectivity arises from Hebbian self-organization,”Nature Physics, vol. 20, no. 3, pp. 484–491, 2024. APPENDIXA STRUCTURALPLASTICITYHYPERPARAMETERS This section summarizes the hyperparameters of the Struc- tural Plasticity Module (SPM) used in all experiments, shown in Table A1. The same ...

    2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.