Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Asymmetric firing-rate networks can be read as games in which each neuron minimizes its own energy, and stacked lateral-inhibition columns sharpen subthreshold input differences with a closed-form depth estimate.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 18:26 UTC pith:GCR5L2HT

load-bearing objection A clean game-theoretic reformulation of asymmetric rate networks, with an oversold contrast-enhancement claim and a main-text error about the simulation parameters. the 3 major comments →

arxiv 2512.05252 v2 pith:GCR5L2HT submitted 2025-12-04 q-bio.NC cond-mat.dis-nnmath.OC

Competition, stability, and functionality in excitatory-inhibitory neural circuits

Simone Betteti , William Retnaraj , Alexander Davydov , Jorge Cort\'es , Francesco Bullo This is my paper
classification q-bio.NC cond-mat.dis-nnmath.OC MSC 92C2091A1093D30
keywords asymmetric firing-rate networksenergy-based modelsNash equilibriumexcitatory-inhibitory circuitslateral inhibitionwinner-take-allcortical columnscontrast enhancement
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.

This paper sets out to restore the interpretive power of energy-based models for asymmetric excitatory-inhibitory neural circuits, where the classic symmetric energy landscape does not exist. Its central move is to assign each neuron its own scalar energy, so that the collective dynamics become a game and the stable outcomes are Nash equilibria. On top of this, Lyapunov diagonal stability supplies conditions under which lateral-inhibition circuits perform winner-take-all selection with finite precision. When such circuits are stacked into a cortical column with per-layer inhibitory compensation, the paper derives a closed-form expression for the number of layers needed to turn subthreshold input differences into categorical choices. A sympathetic reader would care because the framework turns a design question—how deep must the hierarchy be to achieve a given discrimination—into a one-line formula.

Core claim

The paper's central assertion is that a firing-rate network with asymmetric synaptic matrix W admits a family of per-neuron energies E_i(x,u_i) = -x_i Σ_j(1-1/2 δ_ij)W_ij x_j - x_i u_i + ∫_0^{x_i} Φ_i^{-1}(s) ds, and that the fixed points of the dynamics coincide exactly with the Nash equilibria of the game in which neuron i minimizes E_i. In the E2I lateral-inhibition circuit, the unique globally stable winner-take-all equilibrium exists precisely when the WTA conditions (19) and dissipation-excitation dominance (20) hold; the E_kI generalization inherits the same conditions. Stacked E2I layers, each with inhibitory compensation that normalizes the output mean, amplify an input half-separat

What carries the argument

The carrying object is the neuron-specific energy E_i(x,u_i), built from the proximal-gradient decomposition of firing-rate dynamics: each neuron's interaction cost is chosen so that its partial derivative equals the synaptic drive, and the activation cost encodes the saturation nonlinearity via the inverse activation function. The dynamics become proximal pseudo-gradient play, and equilibria are Nash equilibria. Stability is certified by Lyapunov diagonal stability (LDS) of D-W, which for the E2I circuit reduces to w_EE<d_E. The closed-form depth estimate follows from multiplying the equilibrium difference recursion x^{l+1}_{E1}-x^{l+1}_{E2} = [w_EE/(d_E-w_EE)](x^l_{E1}-x^l_{E2}) layer afte

Load-bearing premise

The quantitative depth formula assumes that all synapses of a given type have identical strengths and that at every layer the neurons stay in the linear, unsaturated regime with the inhibitory bias adjusted by exactly the compensation law (148); if any of these fail, the per-layer gain is no longer constant and the closed-form layer count collapses.

What would settle it

Simulate the E2I column with the prescribed parameters but choose inputs near the saturation boundary (e.g., where one excitatory neuron's drive exceeds the activation ceiling) and measure the equilibrium half-separation at each layer; if the amplification factor differs from w_EE/(d_E-w_EE) layer to layer, or if the required depth deviates from the formula, the recursion underlying L fails. Alternatively, add random multiplicative perturbations of ±10% to the homogeneous weights and check whether the layer count predicted by the formula still yields WTA; a sharp increase in required depth wou

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • The full dynamical repertoire of the Wilson-Cowan model—unique fixed point, multiple equilibria, spiral convergence, and limit cycles—is classified as consensual and antagonistic game regimes, with the self-excitatory weight w_EE as the main switch.
  • E2I and E_kI lateral-inhibition circuits implement finite-precision winner-take-all, and the conditions (19)-(20) are necessary and sufficient for a unique, globally stable categorical outcome.
  • Cortical columns built from these units with per-layer inhibitory compensation act as contrast enhancers: subthreshold input differences are amplified by a fixed per-layer gain, and the required depth grows logarithmically as the input difference shrinks.
  • The single-layer E2I model is recovered in the limit ε→δ, where the depth estimate collapses to L=1, unifying the layered and single-stage descriptions.
  • The game-energetic formulation indicates a route to mechanism design: choosing energies that induce desired regimes of competition, oscillation, or selective amplification.

Where Pith is reading between the lines

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

  • The log-linear relation between required depth and input separation is a quantitative prediction that could be tested in simulations with heterogeneous synaptic weights; departures from the homogeneous weights would reveal how fragile the closed-form depth estimate is.
  • The per-neuron energies suggest a natural local learning rule: if each neuron descends its own energy, then Hebbian-style updates on W might be derived as gradient play, linking the framework to online learning in asymmetric circuits.
  • The inhibitory compensation law (148) is an exact prescription; real circuits that only approximate it would be expected to show degraded amplification, making the robustness of the depth formula an empirical question.
  • Because LDS is sufficient but possibly not necessary for global stability in these networks, a weaker stability certificate (e.g., exploiting the sign structure of E-I matrices) might extend the WTA conditions to a larger parameter region.

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 a per-neuron, game-theoretic energy for asymmetric firing-rate networks. Starting from the proximal-gradient reformulation of firing-rate dynamics, it defines neuron-specific interaction and activation costs whose Nash equilibria coincide with the fixed points of the original dynamics. The framework is then applied to three settings: a Wilson-Cowan E-I pair, where it reproduces and reinterprets the known regimes (consensual, antagonistic weak decision, limit cycles); an E2I lateral-inhibition circuit, for which the authors state WTA and global-stability conditions; and a layered cortical column of E2I units, for which they derive a closed-form depth estimate L = ceil(1 + ln(epsilon/delta)/ln((1-wEE)/wEE)) (Eq. 25). The central algebraic equivalence is internally consistent, and the Wilson-Cowan classification is a useful reinterpretation. However, the quantitative WTA and cortical-column results are based on an infeasible or incomplete parameter condition, as detailed below.

Significance. The per-neuron energy construction is a clean formal observation: for asymmetric W, the equilibrium condition can be written as the first-order condition of N coupled scalar minimization problems, with the non-uniqueness of interaction costs explicitly acknowledged. The Wilson-Cowan regime analysis and the connection to zero-sum games are interpretively valuable. The paper is also transparent in stating its LDS-based stability results and in including numerical support for the unproved Conjecture 1. If the WTA and column-depth results were valid, they would provide a falsifiable, analytically tractable model of hierarchical contrast enhancement. The current manuscript, however, does not establish those results in the parameter regime used for its central simulations: the stated WTA condition (19) and dissipation-excitation dominance (20) cannot hold simultaneously for wEE=0.8 and delta=0.1, the values used in Fig. 6. This is a load-bearing gap rather than a presentation issue.

major comments (3)
  1. [Eq. (19)-(20), 'Lateral inhibition and neural input sensitivity'] The WTA condition (19) is inconsistent with dissipation-excitation dominance (20) for the paper's own simulation parameters. With D=I, zero bias, and centered inputs (Assumption 4 in the SI), the equilibrium x*=(1,1,0) requires wEE - wEI + delta >= 1. Combining with (20), wEE < 1, gives delta >= 1 - wEE + wEI > 1 - wEE. For wEE=0.8 and delta=0.1 (Fig. 6), this requires delta > 0.2, which is impossible for any wEI > 0. If a nonzero bias b_E is intended, then (19) is not the correct condition: it becomes wEE - wEI + b_E + delta >= 1, and the centered-input Assumption 4 must be stated in the main text. As written, the existence and uniqueness claims for the WTA equilibrium have no valid parameter regime for the simulations shown.
  2. [Eq. (25) and SI Prop. 13] The depth formula (25) inherits the same infeasibility. The recursion in SI Eqs. (186)-(188) assumes that each layer operates in the unsaturated linear regime until the final threshold is crossed, but that final threshold is exactly the saturated WTA condition (19), which is inconsistent with (20) for the stated parameters. Moreover, Eq. (25) is stated in the main text without the conditions under which it is derived: the inhibitory compensation law (SI Eq. 148) and the admissible input-mean range in SI Theorem 6. Without these, the claim that the column depth can be computed analytically for arbitrary inputs with |u_E1-u_E2|=2epsilon<2delta is not supported.
  3. [SI Prop. 9 and Theorem 3; 'Different firing frequency' section] The global asymptotic stability assertion that follows Eq. (20) is not fully established for the non-scalar-timescale E2I network. SI Theorem 3 proves GAS only under the additional assumption WD=DW, which for D=diag(d_E,d_I,d_E) and the E2I synaptic matrix holds only when d_E=d_I. The paper's own Conjecture 1, which would cover the non-commuting case, is unproved and supported only by Monte Carlo simulation. The main text's unconditional wording ('the equilibrium point is unique and is globally asymptotically stable') should be qualified, and the role of Conjecture 1 should be stated where the result is used.
minor comments (5)
  1. [Notation, Eq. (9)] There are small formatting errors: in Eq. (9), F_i(x) is written with a stray comma as 'integral Phi^{-1}_i(s), ds'; several places use 'proxEi_act' or 'E^i_int' without defining the superscript convention. Please clean up the notation.
  2. [Eq. (19) and SI Lemma 3] The main text gives WTA condition (19) without the parameters d_E, d_I, and bias b used in SI Lemma 3. Since the main text Eq. (16) includes timescales tau_E, tau_I but no dissipation or bias, the relation between the two formulations should be made explicit.
  3. [Eq. (25) and SI Prop. 13] Eq. (25) writes ln((1-wEE)/wEE), while SI Prop. 13 has ln(dE/wEE-1). These coincide only for dE=1; the main text should state this normalization explicitly.
  4. [Fig. 6 caption] The caption reports wEE=0.8 and delta=0.1 but does not list wEI, wIE, or wII, so the reader cannot verify that the WTA conditions (19) are satisfied. Please report all parameters used in the simulation.
  5. [SI Conjecture 1 numerical validation] The sentence 'we performed N=27,000 independent simulations, which provides at least p-hat=0.99 confidence' is imprecise; it should state the confidence level and margin, e.g., as in Lemma 2. Also, sampling 'uniformly from matrices with Frobenius norm in [0,100]' is not uniform over the LDS subset; this should be clarified.

Circularity Check

1 steps flagged

Per-neuron energy is constructed to reproduce the firing-rate fixed point, making the Nash-equilibrium equivalence definitional; WTA and depth results remain independent.

specific steps
  1. self definitional [Section 'To each neuron its own energy', Eqs. (7)-(8); SI Section 2.2, Eqs. (56)-(59) and Remark 2]
    "Owing to the equivalence with the single neuron dynamics in (1), the following relationship must hold xi − ∂xi Ei_int(x, ui) = Σ_j W_ij x_j + u_i ... from which it easy to see that the interaction cost of each neuron is Ei_int(x, ui) = −xi[Σ_j (1 − 1/2 δij)Wij xj − 1/2 xi + ui]. ... Notice that this is the minimal function that yields the desired partial derivative, but not the unique one."

    The energy is defined by Eq. (7) so that its partial derivative matches the firing-rate vector field; hence the Nash equilibrium/fixed-point coincidence in Eq. (10) is an identity by construction, not a prediction. Remark 2 confirms non-uniqueness (adding K(x−i) leaves dynamics unchanged), showing the energies are a parametrization of the dynamics. Later WTA/LDS/depth results derive from Eqs. (16)/(24) directly and are not affected.

full rationale

The only substantive circularity is the central variational reformulation: the per-neuron energy is chosen so that the firing-rate fixed point is the Nash equilibrium, making the 'game-energetic interpretation' a definitional recasting of the dynamics. The paper is transparent about the non-uniqueness, which lowers the severity. The remaining functional analysis is not circular: the E2I WTA conditions (19)-(20) and the global-stability conclusions are derived from the equilibria and Lyapunov diagonal stability of the firing-rate model; the cortical-column depth estimate (25)/Prop. 13 follows by repeated application of the unsaturated equilibrium equations, not from the fitted game. Self-citations ([Betteti et al. 2025a,b], [Nozari and Cortes 2019,2021]) are used as background or confirmation and are not load-bearing for the asymmetric construction. The alleged infeasibility of condition (19) with the Fig. 6 parameters (wEE=0.8, δ=0.1 requires wEE−wEI+δ ≥ 1 with wEI > 0) is a correctness/parameter-regime concern, not a circularity, so it does not further raise the score. Overall: partial circularity from one construction-by-definition step, with independent dynamical results, giving 6.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 1 invented entities

The central results rely on domain assumptions about network structure (reciprocity, homogeneity, slope-restricted activation) and on treating the per-neuron energies as meaningful despite their non-uniqueness; no parameters are fit to data, but δ and wEE are hand-chosen thresholds.

free parameters (2)
  • Per-layer precision δ = e.g., 0.1 in Fig. 6
    Central to the δ-precision WTA inequality (18) and the depth formula (25); it is a design threshold chosen by hand, not estimated from data.
  • Self-excitation wEE = 0.8 in Fig. 6; varied 0.5-3.5 in Fig. 3
    Principal regime switch; the depth formula's gain depends on wEE; no fitting procedure is given.
axioms (5)
  • domain assumption Assumption 1 (E-I reciprocity): every E-to-I edge has a reciprocal I-to-E edge
    Used to make the LDS construction vanish off-diagonal E-I blocks in Theorem 2; many real circuits are not perfectly reciprocal.
  • domain assumption Assumption 2 (synaptic homogeneity): all weights of a given class are identical
    Used in E_kI and cortical-column WTA/depth results; heterogeneity is deferred to future work in the conclusions.
  • domain assumption Assumption 3 (slope-restricted, diagonal, bounded activation function)
    Required for the homeomorphism and global-stability results; the model uses saturating ReLU.
  • domain assumption Unsaturated linear-regime recursion and exact inhibitory compensation for stacked layers
    Depth formula (25) and Proposition 13 assume each layer's soft-WTA equilibrium is linear and compensated by the b'_I law (148); saturation or compensation drift breaks the multiplicative gain.
  • standard math Standard facts from convex/game theory and matrix stability (Forti-Tesi, P-matrices, proximal operators, Poincaré-Bendixson)
    Unproved background tools invoked throughout the derivations.
invented entities (1)
  • Neuron-as-agent with private energy E^i no independent evidence
    purpose: To attach a game-theoretic meaning to asymmetric rate dynamics
    The per-neuron energies are constructed so that their stationary conditions reproduce the firing-rate fixed points; no independent measurement or prediction outside the model is provided for the individual energies.

pith-pipeline@v1.3.0-alltime-deepseek · 47209 in / 24438 out tokens · 246259 ms · 2026-08-03T18:26:38.147949+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Competition, stability, and functionality in excitatory-inhibitory neural circuits." pith.science (2026). https://pith.science/paper/GCR5L2HT

@misc{pith2026251205252,
  author       = {Pith},
  title        = {Pith review of: Competition, stability, and functionality in excitatory-inhibitory neural circuits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GCR5L2HT}},
  note         = {Machine review of arXiv:2512.05252}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Energy-based models have become a central paradigm for understanding computation and stability in both theoretical neuroscience and machine learning. However, the energetic framework typically relies on symmetry in synaptic or weight matrices - a constraint that excludes biologically realistic systems such as excitatory-inhibitory (E-I) networks. When symmetry is relaxed, the classical notion of a global energy landscape fails, leaving the dynamics of asymmetric neural systems conceptually unanchored. In this work, we extend the energetic framework to asymmetric firing rate networks, revealing an underlying game-theoretic structure for the neural dynamics in which each neuron is an agent that seeks to minimize its own energy. In addition, we exploit rigorous stability principles from network theory to study regulation and balancing of neural activity in E-I networks. We combine the novel game-energetic interpretation and the stability results to revisit standard frameworks in theoretical neuroscience, such as the Wilson-Cowan and lateral inhibition models. These insights allow us to study cortical columns of lateral inhibition microcircuits as contrast enhancer - with the ability to selectively sharpen subtle differences in the environment through hierarchical excitation-inhibition interplay. Our results bridge energetic and game-theoretic views of neural computation, offering a pathway toward the systematic engineering of biologically grounded, dynamically stable neural architectures.

Figures

Figures reproduced from arXiv: 2512.05252 by Alexander Davydov, Francesco Bullo, Jorge Cort\'es, Simone Betteti, William Retnaraj.

Figure 1
Figure 1. Figure 1: A qualitative comparison of neural interactions and collective actions in symmetric (left) and asymmetric (right) firing rate networks. (A) Dynamics of a recurrent neural network with firing rates x, activation function Φ, synaptic matrix W, and external input u. Firing-rate models capture the temporal evolution of mean neural activity across interconnected populations. (B) Graphic visualization of neural … view at source ↗
Figure 2
Figure 2. Figure 2: Graphic interpretation of the proxEact operator action for any fixed external input u. (A) Energy landscape associated to a quadratic interaction cost Eint and activation cost Eact for a saturated activation function. In particular, Eact is equal to zero over the entire domain of linear activation, and equal to +∞ over the entire saturated region. The dynamics are therefore steered inside (or at the bounda… view at source ↗
Figure 3
Figure 3. Figure 3: Self-excitatory weight wEE drives the emergence of limit cycles through Hopf-like bifurcation. Hopf￾like bifurcation observed as the E-I firing-rate network transitions from the antagonistic weak decision regime to the antagonistic indecision regime. Plot shows the limit set for each wEE as it is varied starting in the consensual regime (wEE = 0.5 < 1) and capturing the point of bifurcation at wEE = wII + … view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the Wilson-Cowan model and complete dynamical and game characterization of the exhibited regimes (A) Schematic of the Wilson-Cowan model, with one excitatory neuron (E) and one inhibitory neuron (I). Excitatory synapses (wEE, wIE, light blue) are positive, while inhibitory synapses (wII , wEI , light orange) are negative. Each neuron receives an external input (uE, uI ). (B) Summary table of t… view at source ↗
Figure 5
Figure 5. Figure 5: Schematic of a E 2 I excitatory-inhibitory circuit and visualization of the interacting Energies under LDS constraints. (A) Schematic of a minimal E 2 I circuit composed of two excitatory neurons (E1, E2) interacting through a shared inhibitory interneuron. Each excitatory neuron receives an external input (uE1 , uE2 ), which is forwarded to the inhibitory neuron. The interneuron in turn relays equal inhib… view at source ↗
Figure 6
Figure 6. Figure 6: Schematic of a cortical column of E 2 I networks and layer-wise forward contrast enhancement. (A) Vertical arrangement of E 2 I modules, each forwarding its excitatory activity to the layer above. In the first layer, the two excitatory neurons receive distinct external inputs, uE1 and uE2 . Within the δ-FLI subthreshold regime, the layer converges to a Nash equilibrium (x 1,∗ E1 , x 1,∗ E2 ) that lies betw… view at source ↗
Figure 7
Figure 7. Figure 7: Schematic representation of an excitatory-inhibitory network. The network is comprised of [PITH_FULL_IMAGE:figures/full_fig_p026_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Saturated activation function at variable steepness [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Schematic depiction of the Wilson-Cowan model, with one excitatory and one inhibitory neuron receiving [PITH_FULL_IMAGE:figures/full_fig_p034_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Zero-sum Game (ZSG) saddle-like structure connected to the two-dimensional Wilson-Cowan firing-rate [PITH_FULL_IMAGE:figures/full_fig_p037_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: xE1 xE2 xI wIE wEI wIE wEI wEE wEE wII uE1 uE2 [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Star topology of a biologically plausible Winner-takes-All network. [PITH_FULL_IMAGE:figures/full_fig_p041_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Schematic representation of a cortical column composed by [PITH_FULL_IMAGE:figures/full_fig_p044_13.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TIDE: Asymmetric Neural Circuits for Stabilized Temporal Inhibitory-Excitatory Dynamics

    cs.LG 2026-05 unverdicted novelty 6.0

    TIDE is a neuro-inspired architecture using stabilized asymmetric E-I networks with lateral inhibition and 80:20 balance that trains in under half the time of CTM while gaining +1.65% top-1 accuracy on perturbed ImageNet.

  2. Timescale Limits of Linear-Threshold Networks

    eess.SY 2026-04 unverdicted novelty 5.0

    Under the structural LDS condition, a parameterized family of LTNs converges to a globally exponentially stable PDS in the fast limit and a globally asymptotically stable HSS in the slow limit.

  3. Energy-Based Dynamical Models for Neurocomputation, Learning, and Optimization

    cs.LG 2026-04 unverdicted novelty 3.0

    The paper reviews and extends energy-based dynamical models that use gradient flows and energy landscapes for neurocomputation, learning, and optimization tasks.

Reference graph

Works this paper leans on

57 extracted references · 20 canonical work pages · cited by 3 Pith papers

  1. [1]

    Krotov and J

    D. Krotov and J. J. Hopfield. Dense associative memory for pattern recognition. In Advances in neural information processing systems, volume 29, 2016. doi:doi.org/10.48550/arXiv.1606.01164

    Pith/arXiv arXiv doi:10.48550/arxiv.1606.01164 2016

  2. [2]

    Ramsauer, B

    H. Ramsauer, B. Sch \"a fl, J. Lehner, P. Seidl, M. Widrich, L. Gruber, M. Holzleitner, T. Adler, D. Kreil, M. K. Kopp, G. Klambauer, J. Brandstetter, and S. Hochreiter. Hopfield networks is all you need. In International Conference on Learning Representations, 2021. doi:10.48550/arXiv.2008.02217

    Pith/arXiv arXiv doi:10.48550/arxiv.2008.02217 2021

  3. [3]

    Kozachkov, J

    L. Kozachkov, J. J. Slotine, and D. Krotov. Neuron–astrocyte associative memory. Proceedings of the National Academy of Sciences, 122 0 (21), May 2025. ISSN 1091-6490. doi:10.1073/pnas.2417788122

    doi:10.1073/pnas.2417788122 2025

  4. [4]

    J. J. Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79 0 (8): 0 2554–2558, April 1982. ISSN 1091-6490. doi:10.1073/pnas.79.8.2554

    doi:10.1073/pnas.79.8.2554 1982

  5. [5]

    M. A. Cohen and S. Grossberg. Absolute stability of global pattern formation and parallel memory storage by competitive neural networks. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13 0 (5): 0 815–826, September 1983. ISSN 2168-2909. doi:10.1109/tsmc.1983.6313075

    arXiv 1983

  6. [6]

    J. J. Hopfield. Neurons with graded response have collective computational properties like those of two-state neurons. Proceedings of the National Academy of Sciences, 81 0 (10): 0 3088–3092, May 1984. ISSN 1091-6490. doi:10.1073/pnas.81.10.3088

    doi:10.1073/pnas.81.10.3088 1984

  7. [7]

    Sherrington and S

    D. Sherrington and S. Kirkpatrick. Solvable model of a spin-glass. Physical Review Letters, 35 0 (26): 0 1792–1796, December 1975. ISSN 0031-9007. doi:10.1103/physrevlett.35.1792

    doi:10.1103/physrevlett.35.1792 1975

  8. [8]

    Mezard, G

    M. Mezard, G. Parisi, and M. Virasoro. Spin Glass Theory and Beyond: An Introduction to the Replica Method and Its Applications. WORLD SCIENTIFIC, November 1986. ISBN 9789812799371. doi:10.1142/0271

    doi:10.1142/0271 1986

  9. [9]

    Sequential state generation by model neural networks

    D Kleinfeld. Sequential state generation by model neural networks. Proceedings of the National Academy of Sciences, 83 0 (24): 0 9469–9473, December 1986. ISSN 1091-6490. doi:10.1073/pnas.83.24.9469

    work page doi:10.1073/pnas.83.24.9469 1986

  10. [10]

    Russo and A

    E. Russo and A. Treves. Cortical free-association dynamics: Distinct phases of a latching network. Physical Review E, 85 0 (5), May 2012. ISSN 1550-2376. doi:10.1103/physreve.85.051920

    work page doi:10.1103/physreve.85.051920 2012

  11. [12]

    Hoover, Y

    B. Hoover, Y. Liang, B. Pham, R. Panda, H. Strobelt, D. H. Chau, M. Zaki, and D. Krotov. Energy transformer. Advances in neural information processing systems, 36: 0 27532--27559, 2023. doi:10.48550/arXiv.2302.07253

    Pith/arXiv arXiv doi:10.48550/arxiv.2302.07253 2023

  12. [13]

    Ambrogioni

    L. Ambrogioni. In search of dispersed memories: Generative diffusion models are associative memory networks. Entropy, 26 0 (5): 0 381, April 2024. ISSN 1099-4300. doi:10.3390/e26050381

    doi:10.3390/e26050381 2024

  13. [14]

    Horn and J

    D. Horn and J. Weyers. Hypercubic structures in orthogonal hopfield models. Physical Review A, 36 0 (10): 0 4968–4974, November 1987. ISSN 0556-2791. doi:10.1103/physreva.36.4968

    work page doi:10.1103/physreva.36.4968 1987

  14. [15]

    G. Parisi. Asymmetric neural networks and the process of learning. Journal of Physics A: Mathematical and General, 19 0 (11): 0 L675–L680, August 1986. ISSN 0305-4470. doi:10.1088/0305-4470/19/11/005

    work page doi:10.1088/0305-4470/19/11/005 1986

  15. [16]

    H. Yan, L. Zhao, L. Hu, X. Wang, E. Wang, and J. Wang. Nonequilibrium landscape theory of neural networks. Proceedings of the National Academy of Sciences, 110 0 (45), October 2013. ISSN 1091-6490. doi:10.1073/pnas.1310692110

    work page doi:10.1073/pnas.1310692110 2013

  16. [17]

    Haider, A

    B. Haider, A. Duque, A. R. Hasenstaub, and D. A. McCormick. Neocortical network activity in vivo is generated through a dynamic balance of excitation and inhibition. The Journal of Neuroscience, 26 0 (17): 0 4535–4545, April 2006. ISSN 1529-2401. doi:10.1523/jneurosci.5297-05.2006

    work page doi:10.1523/jneurosci.5297-05.2006 2006

  17. [18]

    J. C. Eccles, P. Fatt, and K. Koketsu. Cholinergic and inhibitory synapses in a pathway from motor‐axon collaterals to motoneurones. The Journal of Physiology, 126 0 (3): 0 524–562, December 1954. ISSN 1469-7793. doi:10.1113/jphysiol.1954.sp005226

    work page doi:10.1113/jphysiol.1954.sp005226 1954

  18. [19]

    H. R. Wilson and J. D. Cowan. A mathematical theory of the functional dynamics of cortical and thalamic nervous tissue. Kybernetik, 13 0 (2): 0 55–80, September 1973. ISSN 0023-5946. doi:10.1007/bf00288786

    work page doi:10.1007/bf00288786 1973

  19. [20]

    Y. Fang, M. A. Cohen, and T. G. Kincaid. Dynamics of a winner-take-all neural network. Neural Networks, 9 0 (7): 0 1141–1154, October 1996. ISSN 0893-6080. doi:10.1016/0893-6080(96)00019-6

    work page doi:10.1016/0893-6080(96)00019-6 1996

  20. [21]

    Jaeger and H

    H. Jaeger and H. Haas. Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science, 304 0 (5667): 0 78–80, April 2004. ISSN 1095-9203. doi:10.1126/science.1091277

    doi:10.1126/science.1091277 2004

  21. [22]

    Dayan and L

    P. Dayan and L. F. Abbott. Theoretical neuroscience: computational and mathematical modeling of neural systems. MIT press, 2005. ISBN 9780262041997

    2005

  22. [23]

    Gerstner and W

    W. Gerstner and W. M. Kistler. Mathematical formulations of hebbian learning. Biological Cybernetics, 87 0 (5–6): 0 404–415, December 2002. ISSN 0340-1200. doi:10.1007/s00422-002-0353-y

    work page doi:10.1007/s00422-002-0353-y 2002

  23. [24]

    Centorrino, A

    V. Centorrino, A. Gokhale, A. Davydov, G. Russo, and F. Bullo. Positive competitive networks for sparse reconstruction. Neural Computation, 36 0 (6): 0 1163–1197, 2024. doi:10.1162/neco_a_01657

    doi:10.1162/neco_a_01657 2024

  24. [25]

    Gokhale, A

    A. Gokhale, A. Davydov, and F. Bullo. Proximal gradient dynamics: Monotonicity , exponential convergence, and applications. IEEE Control Systems Letters, 8: 0 2853--2858, 2024. doi:10.1109/LCSYS.2024.3516632

    arXiv 2024

  25. [26]

    Treves and D

    A. Treves and D. J. Amit. Metastable states in asymmetrically diluted hopfield networks. Journal of Physics A: Mathematical and General, 21 0 (14): 0 3155–3169, July 1988. ISSN 1361-6447. doi:10.1088/0305-4470/21/14/016

    work page doi:10.1088/0305-4470/21/14/016 1988

  26. [27]

    Tatarenko, W

    T. Tatarenko, W. Shi, and A. Nedic. Geometric convergence of gradient play algorithms for distributed nash equilibrium seeking. IEEE Transactions on Automatic Control, 66 0 (11): 0 5342–5353, November 2021. ISSN 2334-3303. doi:10.1109/tac.2020.3046232

    arXiv 2021

  27. [28]

    Nozari and J

    E. Nozari and J. Cortes. Hierarchical selective recruitment in linear-threshold brain networks—part i: Single-layer dynamics and selective inhibition. IEEE Transactions on Automatic Control, 66 0 (3): 0 949–964, March 2021. ISSN 2334-3303. doi:10.1109/tac.2020.3004801

    arXiv 2021

  28. [29]

    Nozari and J

    E. Nozari and J. Cortes. Oscillations and coupling in interconnections of two-dimensional brain networks. In 2019 American Control Conference (ACC), page 193–198. IEEE, July 2019. doi:10.23919/acc.2019.8815118

    arXiv 2019

  29. [30]

    S. H. Hwang and L. Rey-Bellet. Strategic decompositions of normal form games: Zero-sum games and potential games. Games and Economic Behavior, 122: 0 370–390, July 2020. ISSN 0899-8256. doi:10.1016/j.geb.2020.05.003

    doi:10.1016/j.geb.2020.05.003 2020

  30. [31]

    Q. X. Meng and J. W. Liu. Proximal Point Method for Online Saddle Point Problem, page 399–414. Springer Nature Singapore, 2025. ISBN 9789819665792. doi:10.1007/978-981-96-6579-2_27

    work page doi:10.1007/978-981-96-6579-2_27 2025

  31. [32]

    Langenberg

    N. Langenberg. Interior proximal methods for equilibrium programming: part i. Optimization, 62 0 (9): 0 1247–1266, September 2013. ISSN 1029-4945. doi:10.1080/02331934.2011.625028

    arXiv 2013

  32. [33]

    J. S. Isaacson and M. Scanziani. How inhibition shapes cortical activity. Neuron, 72 0 (2): 0 231–243, October 2011. ISSN 0896-6273. doi:10.1016/j.neuron.2011.09.027

    work page doi:10.1016/j.neuron.2011.09.027 2011

  33. [34]

    Gerstner, W

    W. Gerstner, W. M. Kistler, R. Naud, and L. Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. doi:doi.org/10.1017/CBO9781107447615

    doi:10.1017/cbo9781107447615 2014

  34. [35]

    Forti and A

    M. Forti and A. Tesi. New conditions for global stability of neural networks with application to linear and quadratic programming problems. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 42 0 (7): 0 354–366, July 1995. ISSN 1057-7122. doi:10.1109/81.401145

    work page doi:10.1109/81.401145 1995

  35. [36]

    Grossberg

    S. Grossberg. Competition, decision, and consensus. Journal of Mathematical Analysis and Applications, 66 0 (2): 0 470–493, November 1978. ISSN 0022-247X. doi:10.1016/0022-247x(78)90249-4

    work page doi:10.1016/0022-247x(78)90249-4 1978

  36. [37]

    B. W. Connors and M. J. Gutnick. Intrinsic firing patterns of diverse neocortical neurons. Trends in Neurosciences, 13 0 (3): 0 99–104, March 1990. ISSN 0166-2236. doi:10.1016/0166-2236(90)90185-d

    work page doi:10.1016/0166-2236(90)90185-d 1990

  37. [38]

    Marder and J

    E. Marder and J. M. Goaillard. Variability, compensation and homeostasis in neuron and network function. Nature Reviews Neuroscience, 7 0 (7): 0 563–574, July 2006. ISSN 1471-0048. doi:10.1038/nrn1949

    doi:10.1038/nrn1949 2006

  38. [39]

    A. A. Prinz, D. Bucher, and E. Marder. Similar network activity from disparate circuit parameters. Nature Neuroscience, 7 0 (12): 0 1345–1352, November 2004. ISSN 1546-1726. doi:10.1038/nn1352

    doi:10.1038/nn1352 2004

  39. [40]

    B. V. Atallah, W. Bruns, M. Carandini, and M. Scanziani. Parvalbumin-expressing interneurons linearly transform cortical responses to visual stimuli. Neuron, 73 0 (1): 0 159–170, January 2012. ISSN 0896-6273. doi:10.1016/j.neuron.2011.12.013

    work page doi:10.1016/j.neuron.2011.12.013 2012

  40. [41]

    R. J. Douglas and K. A. C. Martin. Neuronal circuits of the neocortex. Annual Review of Neuroscience, 27 0 (1): 0 419–451, July 2004. ISSN 1545-4126. doi:10.1146/annurev.neuro.27.070203.144152

    arXiv 2004

  41. [42]

    Riesenhuber and T

    M. Riesenhuber and T. Poggio. Hierarchical models of object recognition in cortex. Nature Neuroscience, 2 0 (11): 0 1019–1025, November 1999. ISSN 1546-1726. doi:10.1038/14819

    doi:10.1038/14819 1999

  42. [43]

    W. E. Vinje and J. L. Gallant. Sparse coding and decorrelation in primary visual cortex during natural vision. Science, 287 0 (5456): 0 1273–1276, February 2000. ISSN 1095-9203. doi:10.1126/science.287.5456.1273

    arXiv 2000

  43. [44]

    M. F. Burg, S. A. Cadena, G. H. Denfield, E. Y. Walker, A. S. Tolias, M. Bethge, and A. S. Ecker. Learning divisive normalization in primary visual cortex. PLOS Computational Biology, 17 0 (6): 0 e1009028, June 2021. ISSN 1553-7358. doi:10.1371/journal.pcbi.1009028

    doi:10.1371/journal.pcbi.1009028 2021

  44. [45]

    A. Gibbard. Manipulation of voting schemes: A general result. Econometrica, 41 0 (4): 0 587, July 1973. ISSN 0012-9682. doi:10.2307/1914083

    doi:10.2307/1914083 1973

  45. [46]

    Hurwicz and S

    L. Hurwicz and S. Reiter. Designing Economic Mechanisms. Cambridge University Press, May 2006. ISBN 9780511754258. doi:10.1017/cbo9780511754258

    work page doi:10.1017/cbo9780511754258 2006

  46. [47]

    Decarolis, M

    F. Decarolis, M. Goldmanis, and A. Penta. Marketing agencies and collusive bidding in online ad auctions. Management Science, 66 0 (10): 0 4433–4454, October 2020. ISSN 1526-5501. doi:10.1287/mnsc.2019.3457

    arXiv 2020

  47. [48]

    Nisan, T

    N. Nisan, T. Roughgarden, and V. V. Tardos, E. Vazirani. Algorithmic Game Theory. Cambridge University Press, September 2007. ISBN 9780511800481. doi:10.1017/cbo9780511800481

    doi:10.1017/cbo9780511800481 2007

  48. [49]

    Kadmon and H

    J. Kadmon and H. Sompolinsky. Transition to chaos in random neuronal networks. Physical Review X, 5 0 (4), November 2015. ISSN 2160-3308. doi:10.1103/physrevx.5.041030

    doi:10.1103/physrevx.5.041030 2015

  49. [50]

    Mastrogiuseppe and S

    F. Mastrogiuseppe and S. Ostojic. Intrinsically-generated fluctuating activity in excitatory-inhibitory networks. PLOS Computational Biology, 13 0 (4): 0 e1005498, April 2017. ISSN 1553-7358. doi:10.1371/journal.pcbi.1005498

    doi:10.1371/journal.pcbi.1005498 2017

  50. [51]

    Hassan-Moghaddam and M

    S. Hassan-Moghaddam and M. R. Jovanović. Proximal gradient flow and Douglas – Rachford splitting dynamics: Global exponential stability via integral quadratic constraints. Automatica, 123: 0 109311, January 2021. ISSN 00051098. doi:10.1016/j.automatica.2020.109311

    arXiv 2021

  51. [52]

    Parikh and S

    N. Parikh and S. Boyd. Proximal Algorithms . Foundations and Trends in Optimization, Vol. 1 0 (3): 0 123--231, 2013. doi:doi.org/10.1561/2400000003

    doi:10.1561/2400000003 2013

  52. [53]

    R. T. Rockafellar. Monotone operators and the proximal point algorithm. SIAM J. Control Optim., 14 0 (5): 0 877--898, August 1976. doi:doi.org/10.1137/0314056

    doi:10.1137/0314056 1976

  53. [54]

    Hershkowitz

    D. Hershkowitz. Recent directions in matrix stability. Linear Algebra and its Applications, 171: 0 161–186, July 1992. ISSN 0024-3795. doi:10.1016/0024-3795(92)90257-b

    work page doi:10.1016/0024-3795(92)90257-b 1992

  54. [55]

    Betteti, G

    S. Betteti, G. Baggio, F. Bullo, and S. Zampieri. Firing rate models as associative memory: Excitatory-inhibitory balance for robust retrieval. Neural Computation, pages 1--32, 08 2025 b . doi:10.1162/neco.a.28

    doi:10.1162/neco.a.28 2025

  55. [56]

    Del Vecchio and J

    D. Del Vecchio and J. J. E. Slotine. A contraction theory approach to singularly perturbed systems. IEEE Transactions on Automatic Control, 58 0 (3): 0 752–757, March 2013. ISSN 1558-2523. doi:10.1109/tac.2012.2211444

    arXiv 2013

  56. [57]

    F. Bullo. Lectures on Neural Dynamics. self, 2025. URL https://fbullo.github.io/lnd

    2025

  57. [58]

    Grossberg

    S. Grossberg. Contour Enhancement , Short Term Memory , and Constancies in Reverberating Neural Networks . Studies in Applied Mathematics, 52 0 (3): 0 213--257, September 1973. ISSN 0022-2526, 1467-9590. doi:10.1002/sapm1973523213

    work page doi:10.1002/sapm1973523213 1973