Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Organizational Regularities in Recurrent Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Dale monopolarity and modularity boost RNN task accuracy, while Hopfield reciprocity lowers it.

desk verdict Useful but under-supported empirical map: modularity's benefit looks solid, but Dale-homogeneity's is confounded with E/I balance. read the letter →

arxiv 2505.22047 v1 pith:D3GENKOK submitted 2025-05-28 q-bio.NC

Claus Metzner , Achim Schilling , Andreas Maier , Patrick Krauss This is my paper
classification q-bio.NC
keywords recurrentneuralnetworksreservoircomputingDale'sprincipleHopfieldreciprocitymodularitynetworkdynamicssequencegenerationphasediagrams
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 asks whether three organizational regularities seen in biological neural circuits—Dale monopolarity (each neuron has one output polarity), reciprocal connections (paired neurons tend to connect both ways), and modular structure (strong blocks embedded in weak background)—change how recurrent neural networks compute, not just how they behave. The authors construct weight matrices in which each regularity's strength is a continuous tuning parameter, and they measure both spontaneous dynamics and performance in a reservoir-computing sequence generation task. Their central claim is that Dale monopolarity and modularity significantly improve task accuracy, while Hopfield reciprocity degrades it by pushing neurons into early saturation and limiting the reservoir's flexibility. If correct, biological connection statistics are design-relevant for reservoir computing performance, not merely a constraint on the dynamical regime. The paper varies each regularity in isolation, so combined effects remain an open question.

What carries the argument

The central objects are tunable weight-matrix ensembles. Dale homogeneity h makes each column (sending neuron) uniformly excitatory or inhibitory with probability h; Hopfield reciprocity r symmetrizes each off-diagonal pair with probability r; modularity m divides the matrix into S×S blocks, giving strong blocks a larger weight width and weak blocks a smaller one, with strong-block width chosen so the total variance stays constant. These matrices drive a reservoir of N tanh neurons with a pseudoinverse-trained linear readout; the paper tracks four signatures: fluctuation F, correlation C at lag one, nonlinearity N, and task accuracy A. The construction matters because it lets the authors vary one regularity while keeping density, balance, width, and task fixed.

What would settle it

A reanalysis that holds the empirical balance B fixed while varying h, or regresses accuracy on h with B as a covariate, would settle it: if accuracy tracks B rather than h, the Dale effect is a balance effect rather than a polarity-consistency effect.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the statistical organization of recurrent weights is a functional knob: raising Dale homogeneity h and modularity m both reduce fluctuation and saturation and improve accuracy in the sequence generation task, whereas raising Hopfield reciprocity r raises nonlinearity and degrades accuracy. The proposed mechanism is that Dale homogeneity stabilizes dynamics by giving each neuron a single output polarity, modularity compartmentalizes the network into strongly coupled blocks linked by weak pathways, and reciprocity acts as a local ordering force that pushes neuron pairs into mutually reinforced digital-like states, fragmenting the reservoir and limiting the flexible trajectories a linear readout can decode.

Load-bearing premise

The Dale-homogeneity result assumes the accuracy gain is caused by consistent output polarity per neuron, not by the correlated rise in excitatory/inhibitory balance B that the same manipulation produces.

Editorial extensions

If this is right

  • If Dale monopolarity is what helps, then reservoirs whose neurons have fixed output polarity can tolerate stronger recurrent coupling in the chaotic regime without losing accuracy.
  • If modularity is what helps, then embedding strong blocks in a weak background makes formerly unproductive oscillatory, fixpoint, and chaotic regimes usable for computation.
  • If Hopfield reciprocity is what hurts, then symmetric bidirectional weights reduce reservoir flexibility by saturating neurons, and performance degrades even in the chaotic regime.
  • Because the regularity parameters were varied continuously, the effects appear as gradual shifts in F, C, N, and A across both a balanced (b=0) and a weakly unbalanced (b=0.2) setting.

Reading between the lines

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

  • A testable extension not run in the paper: rebalance the excitatory/inhibitory ratio after setting h=1, or statistically control for B, to see whether the Dale accuracy gain is caused by polarity consistency or by the correlated balance shift.
  • Because only one task (short sequence generation with two classes) was used, the ranking modularity > Dale > reciprocity may not transfer to tasks needing long memory or continuous tracking; the paper itself flags this as open.
  • The S=1 modularity result suggests part of the modularity benefit is statistical—a mixture of strong and weak weights—rather than spatial; comparing block-structured with shuffled or uncorrelated mixtures would separate spatial from distributional effects.
  • Applying the same dynamical and accuracy metrics to an empirical connectome would test whether real circuits actually occupy the favorable regime; the authors list this as a future direction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript introduces three continuous control parameters—Dale homogeneity h, Hopfield reciprocity r, and modularity m—for generating random recurrent weight matrices, and studies their effect on spontaneous RNN dynamics (fluctuation F, correlation C, nonlinearity N) and on task accuracy A in a minimal sequence-generation reservoir task. Using stochastic interpolation between unconstrained random matrices and structured matrices, the authors construct phase diagrams in the balance–width plane and one-dimensional parameter scans. They report that Dale homogeneity and modularity improve task accuracy, while Hopfield reciprocity degrades it by increasing saturation. The central contribution is a controlled simulation framework for isolating biologically motivated structural regularities in RNN reservoirs.

Significance. If the central claim holds, it would provide evidence that biologically motivated connectivity statistics—Dale's principle, modularity, and reciprocity—have functional consequences for reservoir computing, beyond their known role in setting the dynamical regime. The paper's strengths are its transparent generative recipes, the control-parameter validation in Fig. 2 (in particular the demonstration that the modularity mixture preserves the global weight width), and the systematic scans across the regularity parameters. However, the evidence as presented does not yet isolate the mechanisms: the h-scan is confounded with a shift in the empirical excitation/inhibition balance B, no error bars or significance tests accompany the word 'significantly', and a key bias parameter in Eq. (1) is left unspecified. The task is also acknowledged by the authors to be minimal, so the generality of the performance claims is limited. The framework is useful, but the specific causal attributions require additional work.

major comments (4)
  1. [§3.1, Fig. 2(d); §3.5 and §3.7] The claim that Dale homogeneity enhances accuracy through column-wise output polarity is confounded with the empirical balance parameter B. Fig. 2(d) shows that increasing h raises B, and the text in §3.1 explicitly acknowledges this ('along with a slight increase in balance B'). Since Fig. 3 and the b=0 versus b=0.2 comparison in Fig. 4 demonstrate that accuracy is strongly balance-dependent, the accuracy gains in the h-scans of §3.5 and §3.7 could be caused by the correlated B shift rather than by column-sign consistency. The authors do not rebalance the matrices to hold B fixed across h, nor do they statistically control for B (for example, by comparing h-sweeps at matched empirical B, or by a regression that includes both h and B). Without such a control, the abstract's first mechanistic claim—that Dale monopolarity enhances accuracy—is not established.
  2. [§3.7, Fig. 4; §3.4–§3.6] The abstract uses the word 'significantly' and the text repeatedly reports 'marked' or 'slight' effects, but no ensemble sizes, error bars, confidence intervals, or significance tests are reported anywhere. Figures 3 and 4 show single aggregated heat maps or curves, and the Methods do not state how many random matrices N_R, episodes E, or independent runs were averaged. This is particularly problematic for the small performance differences, such as the claimed reciprocity-induced degradation at b=0.2, where sampling variability could change the sign of the effect. Quantitative uncertainty measures and, ideally, finite-size checks (e.g., N=100, 200) are needed to support the comparative claims.
  3. [§2.2, Eq. (1)] The bias term b_{w,n} appears in the reservoir update equation and is essential for the quiescent and fixpoint regimes described in §3.2, but its initialization or distribution is never specified. The Methods only states that each neuron receives a bias term, without giving its range or sampling rule. This omission prevents reproduction of the phase diagrams and makes it difficult to assess whether the reported regularity effects depend on the particular bias setting. Please specify the bias distribution and test sensitivity to its parameters.
  4. [§2.3, §4.2 and Abstract] The accuracy results are obtained on a single, deliberately minimal task (T_I=1, T_O=2, N_DC=2), and in large parts of the phase diagram the baseline accuracy is already close to one. The abstract states the findings as general performance claims ('Dale monopolarity and modularity significantly enhance task accuracy') without this qualification. Either temper the abstract to the specific task, or provide evidence from a second, more demanding task to show that the effects are not an artifact of the minimal setup.
minor comments (5)
  1. [Fig. 3 caption] The caption refers to '50 arctan-neurons', while Methods and Eq. (1) use a tanh activation; please unify the terminology.
  2. [§2.9 and Results] The nonlinearity measure is defined as α in §2.9, but all results and figures use N; the relationship between α and N should be stated explicitly.
  3. [§2.1 and §2.8] The workflow in §2.1 mentions measures C_0 and C_1, but only C^{Δt} is defined in §2.8 and used later; please clarify the notation and whether both lags are actually reported.
  4. [§4.2] The final sentence of §4.2 is incomplete: 'a more fruitful approach would be to characterize' ends without a continuation. Please complete or remove the sentence.
  5. [§5.4] The data availability statement says the data and programs 'will be made available upon reasonable request'; consider depositing the analysis code in a public repository to strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regularity parameters are independently scanned construction inputs, and the accuracy comparisons are empirical simulation outcomes, not fitted predictions.

full rationale

The paper does not derive its central claims from the same data used to set its control parameters. Dale homogeneity h, Hopfield reciprocity r, and modularity m are construction inputs that are varied independently over their full ranges, while accuracy A is measured by simulating the reservoir on a fixed sequence-generation task and training only the linear readout via pseudoinverse. The regularity parameters are never tuned to maximize accuracy, and no quantity is both fitted and then reported as a prediction. The self-citations to the authors' earlier work on weight statistics, dynamics, and recurrence resonance are used only to motivate the framework and interpret the phase diagrams; the reported effects of h, r, and m on dynamics and accuracy arise from the simulations presented in this paper. The potential confound between Dale homogeneity and the empirical balance B noted in Fig. 2(d) is a question of causal attribution or experimental control, not circularity: it does not make the accuracy measurement equivalent to the input parameter by construction. Accordingly, the circularity burden is zero.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The ledger is short because this is a numerical parameter study. The main unprovided inputs are the bias distribution and the assumption that a minimal two-class sequence task is representative. No new particles, forces, or theoretical objects are introduced.

free parameters (7)
  • bias vector b_w,n
    Appears in Eq. (1) but its distribution and values are never specified in Methods; all measures depend on it.
  • network size N = 50
    All phase diagrams and scans use N=50; finite-size effects are not studied.
  • connection density d = 1
    All phase diagrams and scans use fully connected matrices; sparsity is not studied here despite being a prior focus.
  • weight width w in regularity scans = 0.2
    The h, r, m scans in Sec. 3.7 fix w=0.2; other widths are not scanned for the regularity comparison.
  • modularity block size S and strong-block fraction f_SB = S=10 or S=1; f_SB=0.1 or 0.2
    Modularity results in Secs. 3.6 and 3.7 depend on these chosen values; only S=1 versus S=10 is compared.
  • task geometry = T_I=1, T_O=2, M=K=2, N_DC=2
    All accuracy measurements use this minimal sequence-generation task; generality is acknowledged as open in Sec. 4.2.
  • input amplitude w_I = 0.3
    Input standard deviation for driven simulations; no sensitivity analysis is given.
assumptions (4)
  • standard math The tanh recurrence law y_n(t)=tanh(bias + input + sum W y) is an adequate model of recurrent neural dynamics for this question.
    Standard reservoir-computing/RNN model used throughout the field and in the authors' prior work; it is a modeling choice, not a derived result.
  • domain assumption The chosen measures F, C, N and the sequence-generation accuracy A capture computational usefulness.
    Defined in Secs. 2.7 to 2.10; the paper's claims about task accuracy and reservoir flexibility depend on them, and no validation on other tasks is provided.
  • ad hoc to paper The stochastic interpolation recipes for h, r, and m isolate each structural regularity.
    Sec. 2.5 constructs the matrices this way; Fig. 2 validates the controls, but the h-induced rise in balance B shows the isolation is imperfect.
  • domain assumption The unspecified bias values b_w,n do not qualitatively affect the conclusions.
    Eq. (1) includes b_w,n but the Methods never define its distribution; the phase diagrams in Fig. 3 depend on some implicit choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Organizational Regularities in Recurrent Neural Networks." pith.science (2026). https://pith.science/paper/D3GENKOK

@misc{pith2026250522047,
  author       = {Pith},
  title        = {Pith review of: Organizational Regularities in Recurrent Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3GENKOK}},
  note         = {Machine review of arXiv:2505.22047}
}
read the original abstract

Previous work has shown that the dynamical regime of Recurrent Neural Networks (RNNs) - ranging from oscillatory to chaotic and fixpoint behavior - can be controlled by the global distribution of weights in connection matrices with statistically independent elements. However, it remains unclear how network dynamics respond to organizational regularities in the weight matrix, as often observed in biological neural networks. Here, we investigate three such regularities: (1) monopolar output weights per neuron, in accordance with Dale's principle, (2) reciprocal symmetry between neuron pairs, as in Hopfield networks, and (3) modular structure, where strongly connected blocks are embedded in a background of weaker connectivity. We construct weight matrices in which the strength of each regularity can be continuously tuned via control parameters, and analyze how key dynamical signatures of the RNN evolve as a function of these parameters. Moreover, using the RNN for actual information processing in a reservoir computing framework, we study how each regularity affects performance. We find that Dale monopolarity and modularity significantly enhance task accuracy, while Hopfield reciprocity tends to reduce it by promoting early saturation, limiting reservoir flexibility.

Figures

Figures reproduced from arXiv: 2505.22047 by the authors.

Figure 1
Figure 1. Example weight matrices. • (a) A standard weight matrix of size 50 × 50, with density d = 1, balance b = 0, width w = 0.5, and all organizational regularity parameters set to zero: r = h = m = 0. • (b) With maximal Hopfield reciprocity r = 1 and all other parameters unchanged, the matrix becomes symmet￾ric about the diagonal, while preserving density, balance, and width. • (c) With maximal Dale homogeneity h = 1, ea… view at source ↗
Figure 2
Figure 2. Prescribed and empirical control parameters. We use a weight matrix of size 50 × 50 with parameters initially set to standard values d = 1 and b = r = h = m = 0. In each of the plots (a-e), one control parameter x is scanned through its full permissible range, while all others remain at their standard values. The empirical measures D, B, R, H are evaluated as a function of the scanned parameter x. • (a) Scan of the … view at source ↗
Figure 3
Figure 3. Phase diagrams of RNN dynamics and computational perfor￾mance, as functions of the balance b and the width w. In all cases, the RNN consists of 50 arctan-neurons, fully connected (density d = 1). • The four columns of phase diagrams, from left to right, correspond to the nonlinearity N, the fluc￾tuation F, the correlation C1, and the accuracy A in a sequence generation task (For details see Methods and Results). • T… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Effect of increasing regularity on RNN dynamics and compu￾tational performance. A RNN with 50 neurons and width w = 0.2 is used in the sequence generation task. We compute the fluctuation F, the nonlinearity N, the correlation C and the accuracy A as one of the regular…
Figure 5
Figure 5. Figure 5: Example neuron activations over time. • (a) Standard, free-running RNN of 50 neurons, connected with density d = 1, balance b = 0, width w = 0.5, and all organizational regularity parameters set to zero: r = h = m = 0. • (b) RNN with maximal Hopfield reciprocity r = 1 …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 33 canonical work pages

  1. [1]

    Deep learning.nature, 521(7553):436–444, 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.nature, 521(7553):436–444, 2015

    2015

  2. [2]

    Review of deep learning: Concepts, cnn architec- tures, challenges, applications, future directions.Journal of big Data, 8(1):1–74, 2021

    Laith Alzubaidi, Jinglan Zhang, Amjad J Humaidi, Ayad Al-Dujaili, Ye Duan, Omran Al-Shamma, Jos´ e Santamar ´ ıa, Mohammed A Fadhel, Muthana Al- Amidie, and Laith Farhan. Review of deep learning: Concepts, cnn architec- tures, challenges, applications, future directions.Journal of big Data, 8(1):1–74, 2021

    work page 2021

  3. [3]

    Recent ad- vances in natural language processing via large pre-trained language models: A survey.ACM Computing Surveys, 56(2):1–40, 2023

    Bonan Min, Hayley Ross, Elior Sulem, Amir Pouran Ben Veyseh, Thien Huu Nguyen, Oscar Sainz, Eneko Agirre, Ilana Heintz, and Dan Roth. Recent ad- vances in natural language processing via large pre-trained language models: A survey.ACM Computing Surveys, 56(2):1–40, 2023

    work page 2023

  4. [4]

    Universality and individuality in neural dynamics across large populations of recurrent networks.Advances in neural information pro- cessing systems, 2019:15629, 2019

    Niru Maheswaranathan, Alex H Williams, Matthew D Golub, Surya Ganguli, and David Sussillo. Universality and individuality in neural dynamics across large populations of recurrent networks.Advances in neural information pro- cessing systems, 2019:15629, 2019

    work page 2019

  5. [5]

    Recurrent neural networks are universal approximators

    Anton Maximilian Sch¨ afer and Hans Georg Zimmermann. Recurrent neural networks are universal approximators. InInternational Conference on Artificial Neural Networks, pages 632–640. Springer, 2006

    work page 2006

  6. [6]

    Universal approximation of flows of control systems by recurrent neural networks

    Miguel Aguiar, Amritam Das, and Karl H Johansson. Universal approximation of flows of control systems by recurrent neural networks. In2023 62nd IEEE Conference on Decision and Control (CDC), pages 2320–2327. IEEE, 2023

    2023

  7. [7]

    echo state

    Herbert Jaeger. The “echo state” approach to analysing and training recurrent neural networks-with an erratum note.Bonn, Germany: German National Re- search Center for Information Technology GMD Technical Report, 148(34):13, 2001

    work page 2001

  8. [8]

    Optimal sequence memory in driven random networks.Physical Review X, 8(4):041029, 2018

    Jannis Schuecker, Sven Goedeke, and Moritz Helias. Optimal sequence memory in driven random networks.Physical Review X, 8(4):041029, 2018

    2018

Show all 47 references
  1. [9]

    Connectivity, dy- namics, and memory in reservoir computing with binary and analog neurons

    Lars B¨ using, Benjamin Schrauwen, and Robert Legenstein. Connectivity, dy- namics, and memory in reservoir computing with binary and analog neurons. Neural computation, 22(5):1272–1311, 2010

    2010

  2. [10]

    In- formation processing capacity of dynamical systems.Scientific reports, 2(1):1–7, 2012

    Joni Dambre, David Verstraeten, Benjamin Schrauwen, and Serge Massar. In- formation processing capacity of dynamical systems.Scientific reports, 2(1):1–7, 2012

    2012

  3. [11]

    Randomly connected networks have short temporal memory.Neural computation, 25(6):1408–1439, 2013

    Edward Wallace, Hamid Reza Maei, and Peter E Latham. Randomly connected networks have short temporal memory.Neural computation, 25(6):1408–1439, 2013. 21

    2013

  4. [12]

    Fading memory echo state networks are universal.Neural Networks, 138:10–13, 2021

    Lukas Gonon and Juan-Pablo Ortega. Fading memory echo state networks are universal.Neural Networks, 138:10–13, 2021

    2021

  5. [13]

    Gradient-based learning drives robust representations in re- current neural networks by balancing compression and expansion.Nature Ma- chine Intelligence, 4(6):564–573, 2022

    Matthew Farrell, Stefano Recanatesi, Timothy Moore, Guillaume Lajoie, and Eric Shea-Brown. Gradient-based learning drives robust representations in re- current neural networks by balancing compression and expansion.Nature Ma- chine Intelligence, 4(6):564–573, 2022

    2022

  6. [14]

    Stimulus-dependent sup- pression of chaos in recurrent neural networks.Physical Review E, 82(1):011903, 2010

    Kanaka Rajan, LF Abbott, and Haim Sompolinsky. Stimulus-dependent sup- pression of chaos in recurrent neural networks.Physical Review E, 82(1):011903, 2010

    2010

  7. [15]

    Controlling recurrent neural networks by conceptors.arXiv preprint arXiv:1403.3369, 2014

    Herbert Jaeger. Controlling recurrent neural networks by conceptors.arXiv preprint arXiv:1403.3369, 2014

    2014 arXiv

  8. [16]

    Understanding and con- trolling memory in recurrent neural networks

    Doron Haviv, Alexander Rivkind, and Omri Barak. Understanding and con- trolling memory in recurrent neural networks. InInternational Conference on Machine Learning, pages 2663–2671. PMLR, 2019

    2019

  9. [17]

    Suppressing chaos in neural networks by noise.Physical review letters, 69(26):3717, 1992

    Lutz Molgedey, J Schuchhardt, and Heinz G Schuster. Suppressing chaos in neural networks by noise.Physical review letters, 69(26):3717, 1992

    1992

  10. [18]

    Noise-modulated neural networks as an application of stochastic resonance.Neurocomputing, 277:29–37, 2018

    Shuhei Ikemoto, Fabio DallaLibera, and Koh Hosoda. Noise-modulated neural networks as an application of stochastic resonance.Neurocomputing, 277:29–37, 2018

    2018

  11. [19]

    Recur- rence resonance” in three-neuron motifs.Frontiers in computational neuro- science, 13, 2019

    Patrick Krauss, Karin Prebeck, Achim Schilling, and Claus Metzner. Recur- rence resonance” in three-neuron motifs.Frontiers in computational neuro- science, 13, 2019

    2019

  12. [20]

    Con- trol of noise-induced coherent oscillations in time-delayed neural motifs.arXiv preprint arXiv:2106.11361, 2021

    Florian B¨ onsel, Patrick Krauss, Claus Metzner, and Marius E Yamakou. Con- trol of noise-induced coherent oscillations in time-delayed neural motifs.arXiv preprint arXiv:2106.11361, 2021

    2021 arXiv

  13. [21]

    Dynamics and information import in re- current neural networks.Frontiers in Computational Neuroscience, 16, 2022

    Claus Metzner and Patrick Krauss. Dynamics and information import in re- current neural networks.Frontiers in Computational Neuroscience, 16, 2022

    2022

  14. [22]

    Recurrent neural networks as versatile tools of neuroscience re- search.Current opinion in neurobiology, 46:1–6, 2017

    Omri Barak. Recurrent neural networks as versatile tools of neuroscience re- search.Current opinion in neurobiology, 46:1–6, 2017

    2017

  15. [23]

    Highly nonrandom features of synaptic connectivity in local cortical circuits.PLoS biology, 3(3):e68, 2005

    Sen Song, Per Jesper Sj¨ ostr¨ om, Markus Reigl, Sacha Nelson, and Dmitri B Chklovskii. Highly nonrandom features of synaptic connectivity in local cortical circuits.PLoS biology, 3(3):e68, 2005

    2005

  16. [24]

    Is cortical connectivity optimized for storing information?Na- ture neuroscience, 19(5):749–755, 2016

    Nicolas Brunel. Is cortical connectivity optimized for storing information?Na- ture neuroscience, 19(5):749–755, 2016

    2016

  17. [25]

    Exploring sparsity in recurrent neural networks.arXiv preprint arXiv:1704.05119, 2017

    Sharan Narang, Erich Elsen, Gregory Diamos, and Shubho Sengupta. Exploring sparsity in recurrent neural networks.arXiv preprint arXiv:1704.05119, 2017. 22

    2017 arXiv

  18. [26]

    Sparsity through evolutionary pruning prevents neuronal networks from over- fitting.Neural Networks, 128:305–312, 2020

    Richard C Gerum, Andr´ e Erpenbeck, Patrick Krauss, and Achim Schilling. Sparsity through evolutionary pruning prevents neuronal networks from over- fitting.Neural Networks, 128:305–312, 2020

    2020

  19. [27]

    Effect of dilution in asymmetric recurrent neural networks.Neural Networks, 104:50–59, 2018

    Viola Folli, Giorgio Gosti, Marco Leonetti, and Giancarlo Ruocco. Effect of dilution in asymmetric recurrent neural networks.Neural Networks, 104:50–59, 2018

    2018

  20. [28]

    Analysis of structure and dynamics in three-neuron motifs.Frontiers in Computational Neuroscience, 13:5, 2019

    Patrick Krauss, Alexandra Zankl, Achim Schilling, Holger Schulze, and Claus Metzner. Analysis of structure and dynamics in three-neuron motifs.Frontiers in Computational Neuroscience, 13:5, 2019

    2019

  21. [29]

    Weight statistics controls dynamics in recurrent neural networks.PloS one, 14(4):e0214541, 2019

    Patrick Krauss, Marc Schuster, Verena Dietrich, Achim Schilling, Holger Schulze, and Claus Metzner. Weight statistics controls dynamics in recurrent neural networks.PloS one, 14(4):e0214541, 2019

    2019

  22. [30]

    Intrinsic noise improves speech recognition in a computational model of the auditory pathway.Frontiers in Neuroscience, 16:908330, 2022

    Achim Schilling, Richard Gerum, Claus Metzner, Andreas Maier, and Patrick Krauss. Intrinsic noise improves speech recognition in a computational model of the auditory pathway.Frontiers in Neuroscience, 16:908330, 2022

    2022

  23. [31]

    Patrick Krauss, Konstantin Tziridis, Claus Metzner, Achim Schilling, Ulrich Hoppe, and Holger Schulze. Stochastic resonance controlled upregulation of internal noise after hearing loss as a putative cause of tinnitus-related neuronal hyperactivity.Frontiers in neuroscience, 10...

    2016

  24. [32]

    Achim Schilling, Konstantin Tziridis, Holger Schulze, and Patrick Krauss. The stochastic resonance model of auditory perception: A unified explanation of tinnitus development, zwicker tone illusion, and residual inhibition.Progress in Brain Research, 262:139–157, 2021

    2021

  25. [33]

    Predictive coding and stochastic resonance as fundamental principles of auditory phantom perception.Brain, 146(12):4809–4825, 2023

    Achim Schilling, William Sedley, Richard Gerum, Claus Metzner, Konstantin Tziridis, Andreas Maier, Holger Schulze, Fan-Gang Zeng, Karl J Friston, and Patrick Krauss. Predictive coding and stochastic resonance as fundamental principles of auditory phantom perception.Brain, 146(...

    2023

  26. [34]

    Recur- rence resonance-noise-enhanced dynamics in recurrent neural networks.Fron- tiers in Complex Systems, 2:1479417, 2024

    Claus Metzner, Achim Schilling, Andreas Maier, and Patrick Krauss. Recur- rence resonance-noise-enhanced dynamics in recurrent neural networks.Fron- tiers in Complex Systems, 2:1479417, 2024

    2024

  27. [35]

    Dynamic routing between capsules.Advances in Neural Information Processing Systems, 30, 2017

    Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules.Advances in Neural Information Processing Systems, 30, 2017

    2017

  28. [36]

    Matrix capsules with em routing.International Conference on Learning Representations (ICLR), 2018

    Geoffrey E Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with em routing.International Conference on Learning Representations (ICLR), 2018

    2018

  29. [37]

    Routing networks: Adap- tive selection of non-linear functions for multi-task learning

    Chandler Rosenbaum, Tim Klinger, and Irina Rish. Routing networks: Adap- tive selection of non-linear functions for multi-task learning. InInternational Conference on Learning Representations (ICLR), 2018. 23

    2018

  30. [38]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Mateusz Maziarz, et al. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  31. [39]

    Routing networks and the challenges of modular and compositional computation.arXiv preprint arXiv:1904.12574, 2019

    Chandler Rosenbaum, Irina Rish, and Tim Klinger. Routing networks and the challenges of modular and compositional computation.arXiv preprint arXiv:1904.12574, 2019

    1904 arXiv

  32. [40]

    A critique of pure learning and what artificial neural net- works can learn from animal brains.Nature Communications, 10(1):3770, 2019

    Anthony M Zador. A critique of pure learning and what artificial neural net- works can learn from animal brains.Nature Communications, 10(1):3770, 2019

    2019

  33. [41]

    Neural net- work models with biologically plausible connectivity patterns.Frontiers in Computational Neuroscience, 15:737577, 2021

    Keivan Kim, Apostolos P Georgopoulos, and Andras J Pellionisz. Neural net- work models with biologically plausible connectivity patterns.Frontiers in Computational Neuroscience, 15:737577, 2021

    2021

  34. [42]

    Modularity and dynamics of biological neural networks: A review.Journal of Computa- tional Neuroscience, 50(1):95–117, 2022

    Grace Chang, Shamim Nemati, Matt Wytock, and Michael Collins. Modularity and dynamics of biological neural networks: A review.Journal of Computa- tional Neuroscience, 50(1):95–117, 2022

    2022

  35. [43]

    Dale’s principle.Brain Research Bul- letin, 50(5-6):349–350, 1999

    Piergiorgio Strata and Robert Harvey. Dale’s principle.Brain Research Bul- letin, 50(5-6):349–350, 1999

    1999

  36. [44]

    Salient features of synaptic organisation in the cerebral cortex.Brain Research Reviews, 26(2-3):113–135, 1998

    Peter Somogyi, G´ abor Tam´ as, Rafael Luj´ an, and Eberhard H Buhl. Salient features of synaptic organisation in the cerebral cortex.Brain Research Reviews, 26(2-3):113–135, 1998

    1998

  37. [45]

    A synaptic organizing principle for cortical neuronal groups.Proceedings of the National Academy of Sciences, 108(13):5419–5424, 2011

    Rodrigo Perin, Thomas K Berger, and Henry Markram. A synaptic organizing principle for cortical neuronal groups.Proceedings of the National Academy of Sciences, 108(13):5419–5424, 2011

    2011

  38. [46]

    Modular brain networks.Annual Review of Psychology, 67:613–640, 2016

    Olaf Sporns and Richard F Betzel. Modular brain networks.Annual Review of Psychology, 67:613–640, 2016

    2016

  39. [47]

    Modular and hi- erarchically modular organization of brain networks.Frontiers in Neuroscience, 4:200, 2010

    David Meunier, Renaud Lambiotte, and Edward T Bullmore. Modular and hi- erarchically modular organization of brain networks.Frontiers in Neuroscience, 4:200, 2010. 24 (a) (b) (d)(c) Figure 1:Example weight matrices.•(a)A standard weight matrix of size 50×50, with densityd= 1, ...

    2010

Pith tools

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