Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Reservoir Computation with Networks of Differentiating Neuron Ring Oscillators

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Networks of differentiating ring oscillators can serve as reservoir computers, reaching 90.65% accuracy on MNIST digit recognition.

desk verdict A plausible new reservoir substrate, but the missing control and test-set-selected 90.65% leave the central claim unproven. read the letter →

arxiv 2507.21377 v1 pith:WU6SUBPM submitted 2025-07-28 cs.NE cs.LG

Alexander Yeung , Peter DelMastro , Arjun Karuvally , Hava Siegelmann , Edward Rietman , Hananel Hazan This is my paper
classification cs.NEcs.LG
keywords differentiatingneuronsringoscillatorsreservoircomputingoscillatoryneuralnetworkssmall-worldneuromorphicenergy-efficientAIMNISTdigitrecognition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a reservoir computer need not be built from integrating neurons, which demand a steady current, and that networks of differentiating neurons—which fire only when their input changes—can do the job. Arranging these neurons in rings creates self-sustaining oscillations, and wiring the rings into a small-world network with weak coupling yields a rich dynamical substrate. On the MNIST digit task the authors report 90.65% test accuracy, comparable to other physical reservoir systems. If true, this opens a path to low-power neuromorphic hardware based on change-sensitive circuits.

What carries the argument

The central object is the ring oscillator built from differentiating (Schmitt-trigger) neurons: each neuron's binary output is determined by whether the time derivative of its capacitor voltage crosses a threshold, so pulses travel around the ring and persist. Rings form the nodes of a Watts-Strogatz small-world graph with coupling strength epsilon, and the input is a Hilbert-curve sliding-window embedding of the image; the readout is a linear transformation plus softmax over the aggregated binary snapshot matrix. The small-world shortcuts plus weak coupling allow rapid mixing of information without saturating the dynamics.

What would settle it

Train the same linear-plus-softmax readout directly on the Hilbert-curve sliding-window encoded MNIST time series, bypassing the ring-oscillator reservoir entirely, and compare accuracy to 90.65%; if the encoding-alone baseline matches or exceeds the reservoir result, the central claim that the ring-oscillator dynamics contribute the classification power is falsified. A second check: disable all inter-ring coupling (epsilon=0) and re-run the experiment; if accuracy does not drop substantially, the small-world coupling is not load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that networks of differentiating ring oscillators function as effective reservoir computers when connected in a small-world topology with weak inter-ring coupling. Each neuron is a binary Schmitt-trigger device whose output depends on the sign of the time derivative of its internal voltage, so it spikes only on changes in input rather than sustained input. In rings, these spikes circulate as pulses; when rings are coupled through Watts-Strogatz connections with coupling strength around 0.2, the reservoir operates near the edge of chaos and separates inputs well. A linear readout trained on snapshots of the full network reaches 90.65% accuracy on MNIST, with accuracy saturating around 300 ring oscillators.

Load-bearing premise

The reservoir's dynamics, and not just the Hilbert-curve sliding-window encoding of the image, are what supply the information that the linear readout uses to reach 90.65% accuracy; the paper does not report a control without the reservoir or with a standard reservoir under the same protocol.

Editorial extensions

If this is right

  • Differentiating-neuron reservoirs can run on event-driven, change-triggered signals, potentially eliminating the constant bias current that integrating-neuron reservoirs require.
  • The finding that accuracy stagnates beyond roughly 300 oscillators suggests architectural efficiency matters more than raw size for this substrate.
  • The reported accuracy places this event-driven reservoir alongside previously demonstrated physical reservoirs such as memristor and skyrmion systems on MNIST.
  • Inter-ring weak coupling near the edge of chaos is sufficient to make ring oscillators computationally useful, indicating that local oscillatory circuits can be harvested without strong global synchronization.
  • The approach extends reservoir computing to neuron models whose dynamics emphasize transients rather than history, broadening the class of physical systems usable as reservoirs.

Reading between the lines

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

  • A direct test that the authors did not run: training the same linear readout on the Hilbert-curve sliding-window input with no reservoir would establish how much of the 90.65% comes from the encoding rather than the ring-oscillator dynamics; the paper does not include this control.
  • The event-driven nature suggests a natural hardware mapping to spiking or threshold-switching circuits where activity and power scale with input change rate rather than with time.
  • Because the readout uses only binary snapshots, the same reservoir could likely be read out with extremely simple, low-precision circuits, which may be an additional energy advantage beyond the neuron model itself.
  • The edge-of-chaos tuning described here could be made quantitative by measuring the reservoir's effective Lyapunov spectrum or information processing capacity across epsilon, connecting the observed accuracy peak to a dynamical quantity rather than to grid-search results.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a reservoir computing architecture built from networks of differentiating neuron ring oscillators. Each ring is a set of Schmitt-trigger neurons whose outputs are binary and respond to changes in input; rings are interconnected through a Watts-Strogatz small-world topology with coupling strength epsilon. Static MNIST images are converted to time series via Hilbert-curve traversal and sliding-window embedding, fed into the reservoir, and the aggregated binary states are classified by a linear softmax layer. The authors report a best test accuracy of 90.65% on MNIST at epsilon=0.2 and p=0.4, and study sensitivity to coupling, rewiring probability, and network size, concluding that weak coupling and small-world structure place the network near the edge of chaos and yield effective reservoir computation.

Significance. If substantiated, this would introduce a new, physically motivated reservoir substrate with potentially low power consumption, since differentiating neurons fire only on input changes. The paper includes an explicit simulation algorithm, a systematic grid over coupling and topology parameters, and a comparison to other in-materio reservoirs, which are useful contributions. However, the central claim that the ring-oscillator dynamics provide the computational power is not yet supported because no baseline without the reservoir is reported; the headline accuracy is also selected from a test-set grid. The manuscript is a promising proof-of-concept, but the evidence in its current form is insufficient for the stated conclusion.

major comments (4)
  1. [Sec. III.B and Sec. IV.A] The central claim that differentiating ring oscillators are 'effective reservoir computers' requires showing that the reservoir dynamics add predictive power beyond the input representation alone. The paper never reports a control: no linear classifier trained directly on the same Hilbert sliding-window embedding without the reservoir, no standard ESN under the identical protocol, and no raw-pixel softmax baseline. Because a linear softmax on raw MNIST pixels already achieves about 92% accuracy, the reported 90.65% is below a known trivial baseline, so the data are consistent with the hypothesis that the encoding alone carries the accuracy and that the reservoir acts as a lossy bottleneck. This is load-bearing for the abstract and conclusion, and the missing ablations should be added.
  2. [Sec. IV.A] The 90.65% test accuracy is the maximum over a grid of epsilon and p values (each varied from 0.1 to 0.9), so it is selected on the test set and is optimistically biased. The corresponding validation accuracy at that configuration is 88.47%, which is a more honest estimate under that selection rule. The paper should report model selection based on validation accuracy, followed by evaluation on a held-out test set, or nested cross-validation, and should provide means and standard deviations across the five trials rather than a single best test number.
  3. [Appendix A, Algorithm 1 vs. Eq. (3)] There is an inconsistency between the mathematical definition of the Schmitt trigger and its algorithmic implementation, which prevents exact reproduction. Equation (3) assigns output 0 when tau*vdot >= v_thh, and 0 when tau*vdot is in [v_thl, v_thh] and y(t-) = 0, with 1 otherwise; with the stated convention that 0 denotes firing, this means high derivative suppresses firing. In contrast, Algorithm 1's SchmittTrigger returns 0 when v <= v_thl and 1 when v >= v_thh, which reverses the roles of the thresholds. The authors should align Eq. (3) and Algorithm 1 and clarify the inverter convention.
  4. [Sec. III.A and Appendix A] The simulation and training parameters are listed by name but their numeric values are never reported: n_in, n_ts, dt, T, v_thh, v_thl, ring-size range, number of rings, noise level, warmup duration, optimizer, epochs, regularization, and data split are all undefined. Without these values, the experiment cannot be reproduced or compared with other reservoirs. A complete table of parameters used for the headline result is needed.
minor comments (5)
  1. [Sec. III, input connection description] The sentence 'A subset of rings is reserved for input and is externally stimulated by a sequence of inputs with frequency controlled by the connectivity parameter p' is misleading: Algorithm 1 uses p as the probability of a Bernoulli trial for input connections, not as a frequency. Please rephrase.
  2. [Sec. IV.B and Fig. 6] The claim that performance 'diminishing returns at approximately 300 ring oscillators' is based on trial means, but the error bars overlap substantially across sizes; without a statistical test, the plateau claim is not established. Please add significance testing or confidence intervals.
  3. [Sec. V] The comparison to Lee and Mochizuki (90%) and other in-materio reservoirs is not apples-to-apples because input encoding, reservoir size, and readout differ; please state the differences explicitly or present the comparison only as qualitative context.
  4. [Data Availability] The data availability statement says raw data will be made available on request but no code is released; for an empirical reservoir-computing paper, releasing the simulation code and hyperparameter configuration would greatly improve reproducibility.
  5. [Eq. (3) and Fig. 2] The unusual convention that 0 denotes firing and 1 denotes dormancy is stated but not visually indicated in Fig. 2; please annotate the figure or caption so that the binary outputs can be interpreted correctly.

Circularity Check

0 steps flagged · score 2.0 of 10

No equation-level circularity; minor self-citation motivates topology but does not force the headline result.

full rationale

The paper is an empirical reservoir-computing study, not a derivation. Equations (1)-(3) define the neuron and ring dynamics, Algorithm 1 simulates the network, and the readout is a trained linear+softmax layer on the recorded binary states. No predicted quantity is defined in terms of the quantity it is supposed to predict, and no fitted parameter is renamed as a prediction. The reported 90.65% accuracy is a measured test accuracy from a hyperparameter grid search over epsilon and p; selecting the best test configuration is an optimistic-selection/overfitting concern, not circularity. The only circularity-adjacent passage is Section II.C, where the move from grid to small-world topology is attributed to the authors' own prior work (DelMastro et al. [21], whose authors overlap the present paper). Because no grid-topology control is presented here, that design premise is not independently supported. However, the self-citation is not needed to obtain the MNIST accuracy, which comes from the paper's own simulation and hyperparameter sweeps, so it is not load-bearing on the central numerical claim. The absence of a no-reservoir baseline and of an equivalent ESN comparison is a missing-control issue, not a circularity issue. Overall, no significant circularity is present; the score reflects the one minor self-citation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

All numbers in the headline result are empirical outcomes of choices made by the authors. The most consequential free parameters are the coupling strength and rewiring probability, tuned by grid search and reported as the best test point. Several simulation and readout parameters required by Algorithm 1 are not given numeric values, which further lowers reproducibility. No new fundamental entities are postulated; the Schmitt-trigger neuron model is taken from prior work.

free parameters (7)
  • Coupling strength epsilon = 0.2 (best on test grid)
    Grid-searched from 0.1 to 0.9; the headline 90.65% is the best test result at epsilon=0.2 (Section IV.A).
  • Rewiring probability p = 0.4 (best on test grid)
    Grid-searched with epsilon; also controls input Bernoulli connectivity; headline result at p=0.4 (Section IV.A).
  • Schmitt trigger thresholds v_thh and v_thl = not reported
    Appear in Eq. (3) and Algorithm 1 but no values are given; they define firing and dormancy and directly shape reservoir dynamics.
  • Ring size range = 3 to 10
    Ring sizes drawn uniformly over this range (Section III); rings with fewer than 3 neurons are excluded by assumption.
  • Number of rings = 300 for grid search; 450 max mean accuracy
    Network size scan shows saturation near 300 rings; the best single configuration uses 300 rings (Section IV).
  • Core simulation parameters (dt, n_in, n_ts, noise level, warmup duration) = not reported
    Algorithm 1 lists these inputs but the paper does not state values; they set the length, width, and stochasticity of the input time series.
  • Readout training settings (optimizer, epochs, regularization, data split) = not reported
    The readout is described as a linear plus softmax layer, but training details and train/validation/test split are not specified, making the accuracy unverifiable.
assumptions (5)
  • domain assumption Schmitt trigger binary output model corresponds to physical differentiating neuron dynamics.
    Eq. (3) and Section II.B assume the RC-circuit plus inverter model, with no physical hardware validation in this paper.
  • domain assumption Pulses in a ring of at least 3 neurons circulate indefinitely.
    Section III states rings with two or fewer neurons may not circulate pulses; the oscillation sustaining assumption is inherited from Hasslacher and Tilden (Ref. 17) and not re-derived.
  • domain assumption Small-world topology improves mixing relative to grid topology.
    Section II.C argues long-range shortcuts reduce information path length, but no controlled comparison between grid and small-world with identical hyperparameters is provided.
  • domain assumption Optimal reservoir performance occurs near the edge of chaos.
    Section IV.A interprets the best accuracy at intermediate epsilon as edge-of-chaos behavior, but no quantitative measure of chaos or information capacity is computed.
  • domain assumption Hilbert curve sliding-window embedding preserves MNIST spatial information.
    Section III.B uses this encoding; no baseline of a linear classifier on the encoding alone is reported, so the reservoir's added value is not isolated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reservoir Computation with Networks of Differentiating Neuron Ring Oscillators." pith.science (2026). https://pith.science/paper/WU6SUBPM

@misc{pith2026250721377,
  author       = {Pith},
  title        = {Pith review of: Reservoir Computation with Networks of Differentiating Neuron Ring Oscillators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WU6SUBPM}},
  note         = {Machine review of arXiv:2507.21377}
}
read the original abstract

Reservoir Computing is a machine learning approach that uses the rich repertoire of complex system dynamics for function approximation. Current approaches to reservoir computing use a network of coupled integrating neurons that require a steady current to maintain activity. Here, we introduce a small world graph of differentiating neurons that are active only when there are changes in input as an alternative to integrating neurons as a reservoir computing substrate. We find the coupling strength and network topology that enable these small world networks to function as an effective reservoir. We demonstrate the efficacy of these networks in the MNIST digit recognition task, achieving comparable performance of 90.65% to existing reservoir computing approaches. The findings suggest that differentiating neurons can be a potential alternative to integrating neurons and can provide a sustainable future alternative for power-hungry AI applications.

Figures

Figures reproduced from arXiv: 2507.21377 by the authors.

Figure 1
Figure 1. FIG. 1. Reservoir architecture for computing with networks of differentiating ring oscillators. The [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Dynamics of a ring of 8 differeniating neurons. The overall [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Example coupling of two oscillators, in this case one with [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: FIG. 5. MNIST digit resized and read along Hilbert curve. The origi [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Test accuracy of ring oscillator networks ranging from 50 to [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 22 canonical work pages

  1. [1]

    author author K. M. \ Stiefel \ and\ author J. S. \ Coggan ,\ title title The energy challenges of artificial superintelligence , \ https://api.semanticscholar.org/CorpusID:264497615 journal journal Frontiers in Artificial Intelligence \ volume 6 ( year 2023 ) NoStop

    work page 2023

  2. [2]

    author author D. J. \ Watts \ and\ author S. H. \ Strogatz ,\ title title Collective dynamics of ‘small-world’ networks , \ 10.1038/30918 journal journal Nature \ volume 393 ,\ pages 440--442 ( year 1998 ) NoStop

    doi:10.1038/30918 1998

  3. [3]

    Still \ and\ author M

    author author S. Still \ and\ author M. W. \ Tilden ,\ title Controller for a four-legged walking machine , \ in\ 10.1142/9789812816535\_0012 booktitle Neuromorphic Systems \ ( year 1998 )\ pp.\ pages 138--148 ,\ http://arxiv.org/abs/https://www.worldscientific.com/doi/pdf/10.1142/9789812816535\_0012 https://www.worldscientific.com/doi/pdf/10.1142/9789812...

    work page doi:10.1142/9789812816535 1998

  4. [4]

    author author J. R. \ Frigo \ and\ author M. W. \ Tilden ,\ title title Biologically inspired neural network controller for an infrared tracking system , \ in\ 10.1117/12.335726 booktitle Mobile Robots XIII and Intelligent Transportation Systems, Boston, MA, USA, November 1, 1998 ,\ series SPIE Proceedings , Vol.\ volume 3525 ,\ editor edited by\ editor H...

    work page doi:10.1117/12.335726

  5. [5]

    author author H. Jaeger ,\ title title Towards a generalized theory comprising digital, neuromorphic and unconventional computing , \ 10.1088/2634-4386/abf151 journal journal Neuromorphic Computing and Engineering \ volume 1 ,\ pages 012002 ( year 2021 ) NoStop

    work page doi:10.1088/2634-4386/abf151

  6. [6]

    author author L. Deng ,\ title title The mnist database of handwritten digit images for machine learning research [best of the web] , \ 10.1109/MSP.2012.2211477 journal journal IEEE Signal Processing Magazine \ volume 29 ,\ pages 141--142 ( year 2012 ) NoStop

    arXiv 2012

  7. [7]

    Wolfram ,\ https://www.wolframscience.com title A New Kind of Science \ ( publisher Wolfram Media ,\ year 2002 ) NoStop

    author author S. Wolfram ,\ https://www.wolframscience.com title A New Kind of Science \ ( publisher Wolfram Media ,\ year 2002 ) NoStop

    work page 2002

  8. [8]

    author author S. Stepney ,\ title title Programming unconventional computers: Dynamics, development, self-reference , \ 10.3390/e14101939 journal journal Entropy \ volume 14 ,\ pages 1939--1952 ( year 2012 ) NoStop

    work page doi:10.3390/e14101939 1939

Show all 30 references
  1. [9]

    Maass , author T

    author author W. Maass , author T. Natschl\" a ger , \ and\ author H. Markram ,\ title title Real-time computing without stable states: a new framework for neural computation based on perturbations , \ 10.1162/089976602760407955 journal journal Neural Comput. \ volume 14 ,\ pa...

    doi

  2. [10]

    Schaetti , author M

    author author N. Schaetti , author M. Salomon , \ and\ author R. Couturier ,\ title title Echo state networks-based reservoir computing for mnist handwritten digits recognition , \ in\ 10.1109/CSE-EUC-DCABES.2016.229 booktitle 2016 IEEE Intl Conference on Computational Science...

    2016 doi

  3. [11]

    Rietman , author L

    author author E. Rietman , author L. Schuum , author A. Salik , author M. Askenazi , \ and\ author H. Siegelmann ,\ title title Machine learning with quantum matter: An example using lead zirconate titanate , \ 10.3390/quantum4040030 journal journal Quantum Reports \ volume 4 ...

    doi

  4. [12]

    \ Lee \ and\ author M

    author author M.-K. \ Lee \ and\ author M. Mochizuki ,\ title title Handwritten digit recognition by spin waves in a skyrmion reservoir , \ 10.1038/s41598-023-46677-w journal journal Scientific Reports \ volume 13 ,\ pages 19423 ( year 2023 ) NoStop

    doi

  5. [13]

    echo state

    author author H. Jaeger ,\ title title The" echo state" approach to analysing and training recurrent neural networks-with an erratum note' , \ @noop journal journal Bonn, Germany: German National Research Center for Information Technology GMD Technical Report \ volume 148 ( ye...

    2001

  6. [14]

    author author C. G. \ Langton ,\ title title Computation at the edge of chaos: Phase transitions and emergent computation , \ https://doi.org/10.1016/0167-2789(90)90064-V journal journal Physica D: Nonlinear Phenomena \ volume 42 ,\ pages 12--37 ( year 1990 ) NoStop

    doi

  7. [15]

    Dambre , author D

    author author J. Dambre , author D. Verstraeten , author B. Schrauwen , \ and\ author S. Massar ,\ title title Information processing capacity of dynamical systems , \ 10.1038/srep00514 journal journal Scientific reports \ volume 2 ,\ pages 514 ( year 2012 ) NoStop

    doi

  8. [16]

    Choi \ and\ author P

    author author J. Choi \ and\ author P. Kim ,\ title title Critical neuromorphic computing based on explosive synchronization , \ http://arxiv.org/abs/1810.10944 journal journal CoRR \ volume abs/1810.10944 ( year 2018 ) ,\ http://arxiv.org/abs/1810.10944 1810.10944 NoStop

    2018 arXiv

  9. [17]

    Hasslacher \ and\ author M

    author author B. Hasslacher \ and\ author M. W. \ Tilden ,\ title title Theoretical foundations for nervous networks , \ 10.1063/1.54209 journal journal AIP Conference Proceedings \ volume 411 ,\ pages 179--184 ( year 1997 ) NoStop

    1997 doi

  10. [18]

    Hochreiter \ and\ author J

    author author S. Hochreiter \ and\ author J. Schmidhuber ,\ title title Long short-term memory , \ 10.1162/neco.1997.9.8.1735 journal journal Neural Comput. \ volume 9 ,\ pages 1735–1780 ( year 1997 ) NoStop

    1997 doi

  11. [19]

    Hasslacher \ and\ author M

    author author B. Hasslacher \ and\ author M. W. \ Tilden ,\ title title Living machines , \ 10.1016/0921-8890(95)00019-C journal journal Robotics Auton. Syst. \ volume 15 ,\ pages 143--169 ( year 1995 ) NoStop

    1995 doi

  12. [20]

    Zhang , author M

    author author T. Zhang , author M. R. \ Haider , author I. D. \ Alexander , \ and\ author Y. Massoud ,\ title title A coupled schmitt trigger oscillator neural network for pattern recognition applications , \ in\ 10.1109/MWSCAS.2018.8624010 booktitle 2018 IEEE 61st Internation...

  13. [21]

    DelMastro , author A

    author author P. DelMastro , author A. Karuvally , author H. Hazan , author H. T. \ Siegelmann , \ and\ author E. A. \ Rietman ,\ title title Transient dynamics in lattices of differentiating ring oscillators , \ @noop journal journal ArXiv \ volume abs/2506.07253 ( year 2025 ) NoStop

    2025 arXiv

  14. [22]

    Midya , author Z

    author author R. Midya , author Z. Wang , author S. Asapu , author X. Zhang , author M. Rao , author W. Song , author Y. Zhuo , author N. Upadhyay , author Q. Xia , \ and\ author J. J. \ Yang ,\ title title Reservoir computing using diffusive memristors , \ https://doi.org/10....

    doi

  15. [23]

    author author J. A. \ Acebr\'on , author L. L. \ Bonilla , author C. J. \ P\'erez Vicente , author F. Ritort , \ and\ author R. Spigler ,\ title title The kuramoto model: A simple paradigm for synchronization phenomena , \ 10.1103/RevModPhys.77.137 journal journal Rev. Mod. Ph...

    doi

  16. [24]

    author author R. C. \ Budzinski , author T. T. \ Nguyen , author G. B. \ Benigno , author J. Đoàn , author J. Mináč , author T. J. \ Sejnowski , \ and\ author L. E. \ Muller ,\ title title Analytical prediction of specific spatiotemporal patterns in nonlinear oscillator networ...

    2022

  17. [25]

    author author G. B. \ Benigno , author R. C. \ Budzinski , author Z. W. \ Davis , author J. H. \ Reynolds , \ and\ author L. E. \ Muller ,\ title title Waves traveling over a map of visual space can ignite short-term predictions of sensory input , \ https://api.semanticscholar...

    2023

  18. [26]

    author author T. A. \ Keller , author L. E. \ Muller , author T. J. \ Sejnowski , \ and\ author M. Welling ,\ title title Traveling waves encode the recent past and enhance sequence learning , \ https://api.semanticscholar.org/CorpusID:262013982 journal journal ArXiv \ volume ...

    2023 arXiv

  19. [27]

    Karuvally , author T

    author author A. Karuvally , author T. Sejnowski , \ and\ author H. T. \ Siegelmann ,\ title title Hidden traveling waves bind working memory variables in recurrent neural networks , \ in\ https://proceedings.mlr.press/v235/karuvally24a.html booktitle Proceedings of the 41st I...

    2024

  20. [28]

    Hazan \ and\ author L

    author author H. Hazan \ and\ author L. M. \ Manevitz ,\ title title Topological constraints and robustness in liquid state machines , \ https://doi.org/10.1016/j.eswa.2011.06.052 journal journal Expert Systems with Applications \ volume 39 ,\ pages 1597--1606 ( year 2012 ) NoStop

    2011 doi

  21. [29]

    author author E. A. \ Rietman , author M. W. \ Tilden , \ and\ author M. Askenazi ,\ title title Analog computation with rings of quasiperiodic oscillators: the microdynamics of cognition in living machines , \ @noop journal journal Robotics Auton. Syst. \ volume 45 ,\ pages 2...

    2003

  22. [30]

    author author T. W. \ Hughes , author I. A. D. \ Williamson , author M. Minkov , \ and\ author S. Fan ,\ title title Wave physics as an analog recurrent neural network , \ https://api.semanticscholar.org/CorpusID:139104334 journal journal Science Advances \ volume 5 ( year 201...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.