Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability

arXiv:2512.13859 · q-bio.NC, cond-mat.stat-mech · Submitted 2025-12-15 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "Neuromodulation-inspired gated associative memory networks".

Marcus: Here's the gist of the paper. Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability introduces a minimal, biophysically motivated network where neuropeptide-like signals are modeled by a self-adaptive, activity-dependent gating mechanism that fundamentally reorganizes the attractor structure to bypass classical capacity limits The gated network can bypass the sharp loss of retrieval (the “memory cliff”) at the classical critical capacity αc Gating further converts each discrete memory attractor into a continuous cluster of fixed points, so that the retrieved state varies with the initial overlap rather than being quantized to the stored pattern — a feature useful for input-driven computation with persistent activity Model and Biological Motivation The study introduces a continuous-state (graded-response) associative memory network augmented with a multiplicative gating mechanism to mimic the effect of neuropeptides or neuromodulators on individual neurons This mechanism is based on experimental evidence that neuromodulation often acts multiplicatively, effectively gating neuronal inputs or regulating integration timescales The model consists of two recurrently connected layers: a primary layer of N neurons with activity x(t) and an auxiliary layer of N auxiliary units z(t) that mimic neuropeptides or neuromodulators The dynamics are governed by equations where the neuronal activity is influenced by the gating function σ(z), which acts as a switch-like multiplicative neuromodulatory gate Extended Retrieval and Dynamical Stabilization The primary finding is that the gated network can bypass the sharp loss of retrieval (the “memory cliff”) at the classical critical capacity αc Above αc, where a cued pattern otherwise survives only as a high-overlap transient, multiplicative gating renders this transient retrieval permanent when the gates act as fully open or fully closed switches, while graded gating substantially prolongs its lifetime The dynamics show that introducing the self-adaptive gating fundamentally alters this behavior, extending the high-overlap plateau significantly as the steepness γ increases Emergent Multistability and Landscape Reshaping The gating mechanism leads to a novel memory retrieval phase of continuous multistability, where the system supports a continuous manifold of stable fixed-point attractors In the binary gating limit γ → ∞, the system reaches a continuum of distinct fixed points depending on the specific initial overlap, creating a “cloud” of stable fixed points This suggests that in the gated networks, the steady-state overlap can be continuously controlled by choosing a different initial neural condition rather than being quantized to a stored pattern Thermodynamic Verification and Spectral Analysis The findings are corroborated by Dynamical Mean-Field Theory (DMFT), which establishes that the observations made in finite-size simulations persist in the thermodynamic limit and in the long-time limit The analysis of the instantaneous Jacobian reveals a characteristic spectral signature of frozen-stabilization, which is reproduced during retrieval This stabilization mechanism involves a subpopulation ceasing to evolve and supplying a static background field to the rest, which can stabilize active units away from the high-gain region Robustness and Biological Relevance The study demonstrates that hard-gate retrieval remains accessible even when modulatory variables evolve substantially more slowly than the neural variables, depending strongly on the initial gate configuration Furthermore, the results show that gated retrieval is largely robust to rank and density of the modulatory coupling W in the hard-gate limit, requiring W to be neither dense nor full rank for finite gate steepness The emergent multistability suggests that the retrieved state is set continuously by the initial overlap, a feature that can be exploited for input-driven computation with persistent activity The utility of gating may extend from biological systems to machine learning more broadly, such as in gated attention in large language models The work demonstrates that a minimal model of self-adaptive gating fundamentally enhances the classical associative memory properties by circumventing catastrophic spin-glass transitions and inducing emergent multistability The key insights are the circumvention of the capacity ceiling, the stabilization of transients into true fixed points, and the emergence of a continuous manifold of stable fixed points Future work should focus on learning Wij couplings from measurements and adapting the model for spiking networks The gating-induced stabilization of ghost-like transients is identified as a novel physical perspective driving extended retrieval without assuming a Lyapunov function The utility of gating may extend to modeling complex biological systems where neuropeptides are essential for capturing key aspects of brain function The retrieved state being set continuously by the initial overlap is a feature that can be exploited for input-driven computation with persistent activity The study suggests that neuromodulatory gating may provide richer computational capabilities in biological and neuromorphic architectures Supplementary Information The supplementary material provides detailed methods, including direct many-body simulations, DMFT derivations, and analysis of the instantaneous Jacobian spectrum This includes illustrations of realistic neuromodulator networks and the finite-size behavior of the retrieval boundary The results confirm that the spread of steady-state overlaps vanishes as N → ∞, meaning mss is a deterministic function of α and m(0) in the thermodynamic limit The analysis confirms that the gating mechanism inherently induces a certain type of effective higher-order interactions, such as three-body interactions in the adiabatic limit The results confirm that the gated network bypasses the sharp order–to–spin-glass transition, extending the mss = O(1) retrieval phase beyond the standard critical capacity threshold αc The paper concludes that neuromodulatory gating provides richer computational capabilities in biological and neuromorphic architectures References The paper references various foundational works in associative memory, spin glasses, and neural dynamics Key references include Hopfield networks

67: , dynamical mean field theory

77: , and works on neuropeptide connectomes in C.

Ines: elegans

14: The literature also includes studies on long short-term memory (LSTM) architectures

27, 28: and neuromodulation effects on neuronal conductivity The paper discusses conceptual parallels with previous studies on three-body interactions in coupled phase oscillators The supplementary references provide further context on the mathematical formalism, including path-integral methods and spectral decomposition of the Jacobian The paper also discusses the distinction between neuromodulation and synaptic plasticity mechanisms Contents The contents outline the structure of the paper, covering definitions, main results from simulations, DMFT formalism, visualization of continuous multistable attractors, and a summary of findings The supplementary information includes detailed methods for direct many-body simulations and numerical integration of DMFT equations The paper summarizes the key insights regarding the impact of gating on memory retrieval properties, including the circumvention of catastrophic spin-glass transition and the emergence of continuous multistability The paper provides a natural question about how applicable these findings are for biological systems across all species investigated The paper notes that dopamine has been found to lead to increased working memory in the hippocampus by affecting individual neurons The paper suggests that the neuromodulatory couplings Wij should contain structure and low rank based on the neuropeptidergic connectome of C.

Marcus: elegans The paper notes that Costacurta et al.

Title and authors: Ines: So we’re discussing "Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability". The title tells us immediately that they are focusing on how these modulation signals change the way a network stores and retrieves information, specifically looking for extended retrieval.

Marcus: And the authors are Goto, Lopez Rios, Scholz, and Vaikuntanathan. They’re bringing together physics with neuroscience here to model this biophysical mechanism.

Yuki: For me, it suggests that we shouldn't treat memory as purely a structural property of the connections; it has to include these dynamic modulatory components that change the system's behavior during retrieval.

Ines: Precisely. The main point they make is that this minimal, biophysically motivated model fundamentally reorganizes the attractor structure, allowing the network to go past traditional capacity limits <ref:2512.13859#pg1>.

Marcus: And they show this through simulations comparing an ungated benchmark against their gated version, where the standard critical capacity alpha c is around zero point one three in the ungated case <ref:2512.13859#pg1>.

Yuki: So, what does this mean for us on the ground? It means if we are modeling biological systems, we can't just use these standard Hopfield limits without considering these neuromodulatory effects that can stabilize retrieval where it otherwise would collapse.

The paper's summary: Ines: They summarize the central finding around how the gating mechanism alters the retrieval phase, specifically noting that instead of a sharp loss of retrieval at capacity alpha c, which is what happens in ungated networks <ref:2512.13859#pg1>, they get extended retrieval.

Marcus: The paper explains that by making the gates act as fully open or fully closed switches, they can make that transient retrieval permanent, and graded gating just substantially prolongs its lifetime <ref:2512.13859#pg1>.

Yuki: That's interesting because it suggests that even if the pattern is slightly above the classical capacity threshold alpha c of about zero point one three, we might still get a good retrieval quality if the gate configuration is right <ref:2512.13859#pg3>.

Ines: But they go further with emergent multistability, which means the system supports a continuous manifold of stable fixed point attractors instead of just one sharp solution <ref:2512.13859#pg3>.

Marcus: In the hard gate limit, when you look at the overlap as a function of memory load alpha and initial overlap m(zero), they get this "cloud" of stable fixed points, where the steady state overlap can be continuously controlled by choosing a different initial neural condition <ref:2512.13859#pg3>.

Yuki: So instead of the stored pattern being strictly quantized to a specific point, it varies continuously depending on how similar your starting input is to that pattern, which sounds like it could be really useful for input-driven computation <ref:2512.13859#pg3>.

The paper's improvements: Ines: Now let’s talk about the specific improvements the authors suggest for this model. They highlight that the gated network can bypass that sharp loss of retrieval at capacity alpha c, which is a big deal because it means we can maintain order-one retrieval even when the load exceeds zero point one three <ref:2512.13859#pg3>.

Marcus: The paper also points out that the dynamics show that introducing this self-adaptive gating fundamentally alters this behavior, extending the high-overlap plateau significantly as the steepness gamma increases <ref:2512.13859#pg1>.

Yuki: And they’re discussing how hard-gate retrieval remains accessible even when modulatory variables evolve much more slowly than the neural variables, provided you have a specific initial gate configuration <ref:2512.13859#pg4>.

Ines: They also mention that gated retrieval is quite robust to the rank and density of the modulatory coupling W in the hard-gate limit, as long as W isn't too dense or full rank for a finite gate steepness <ref:2512.13859#pg4>.

Marcus: And they’ve proposed a way to tune retrieval access by using an operational threshold theta, which can either broaden or narrow the range of cues from which retrieval happens, allowing for tailored memory access based on fidelity requirements Ines

Conclusion: Ines: So we’re wrapping up the paper "Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability". The authors show that this minimal model of self-adaptive gating enhances classical associative memory properties by circumventing catastrophic spin-glass transitions <ref:2512.13859#pg1>.

Marcus: The main points are the circumvention of the capacity ceiling, the stabilization of transients into true fixed points in the binary gate limit, and finally, the emergence of a continuous manifold of stable fixed point attractors <ref:2512.13859#pg3>.

Yuki: What this means for population genetics or biology is that neuromodulatory gating might provide richer computational capabilities in biological and neuromorphic architectures because it allows the retrieved state to be set continuously by the initial overlap Ines

Ines: Yeah, and they flag a limitation: they note that the spread of steady-state overlaps actually vanishes as N goes to infinity, meaning the steady state overlap becomes a deterministic function of alpha and m(zero) in that limit Supplementary Information <ref:2512.13859#pg3>.

Marcus: So, despite these limitations, the idea is that neuromodulatory gating provides richer computational capabilities in biological and neuromorphic architectures because it allows for input-driven computation with persistent activity <ref:2512.13859#pg4>.

Ines: We’ve discussed how this gated network overcomes the capacity ceiling and introduces multistability, which is a significant step from standard models. Thank you for tuning in to discuss these findings on the paper "Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability".

Daiki Goto, Hector Manuel Lopez Rios, Monika Scholz, Suriyanarayanan Vaikuntanathan

Department of Physics, The University of Chicago · The James Franck Institute, The University of Chicago Department of Chemistry, The University of Chicago Department of Chemistry · Max Planck Institute for Neurobiology of Behavior—caesar

q-bio.NC, cond-mat.stat-mech

Submitted: 2025-12-15

Updated: 2026-11-01

Journal ref: Proc. Natl. Acad. Sci. U.S.A. 123 (44), e2607559123 (2026)

DOI: 10.1073/pnas.2607559123

Code: https://github.com/FelixRoy/dmft_rLV

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability introduces a minimal…

Key concepts

Gating Mechanism
This is a self-adaptive multiplicative switch applied to neuronal inputs, mimicking how neuropeptides or neuromodulators regulate neural activity. It acts like a gate that can either fully open or fully close, fundamentally reorganizing the network's structure and allowing it to bypass classical capacity limits by controlling which signals are integrated into the memory state.
Memory Cliff Bypass
Classical associative memories fail sharply when overloaded, losing retrieval capability above a critical capacity ($\alpha_c$). The gated network avoids this 'memory cliff' because the gating mechanism converts discrete memory states into continuous clusters of fixed points. This means retrieval does not abruptly collapse; instead, it is prolonged or transformed into a stable state.
Emergent Multistability
The gating mechanism causes the network to support a continuous manifold of stable fixed-point attractors rather than just one stored pattern. In the limit where gates are fully open or closed, this creates a 'cloud' of stable states. This allows the retrieved memory to be continuously controlled by adjusting the initial conditions, rather than being strictly quantized to a single stored pattern.

Terminology

Summary

Neuromodulation-inspired gated associative memory networks: extended memory retrieval and emergent multistability introduces a minimal, biophysically motivated network where neuropeptide-like signals are modeled by a self-adaptive, activity-dependent gating mechanism that fundamentally reorganizes the attractor structure to bypass classical capacity limits The gated network can bypass the sharp loss of retrieval (the “memory cliff”) at the classical critical capacity αc Gating further converts each discrete memory attractor into a continuous cluster of fixed points, so that the retrieved state varies with the initial overlap rather than being quantized to the stored pattern — a feature useful for input-driven computation with persistent activity

Model and Biological Motivation

The study introduces a continuous-state (graded-response) associative memory network augmented with a multiplicative gating mechanism to mimic the effect of neuropeptides or neuromodulators on individual neurons This mechanism is based on experimental evidence that neuromodulation often acts multiplicatively, effectively gating neuronal inputs or regulating integration timescales The model consists of two recurrently connected layers: a primary layer of N neurons with activity x(t) and an auxiliary layer of N auxiliary units z(t) that mimic neuropeptides or neuromodulators The dynamics are governed by equations where the neuronal activity is influenced by the gating function σ(z), which acts as a switch-like multiplicative neuromodulatory gate

Extended Retrieval and Dynamical Stabilization

The primary finding is that the gated network can bypass the sharp loss of retrieval (the “memory cliff”) at the classical critical capacity αc Above αc, where a cued pattern otherwise survives only as a high-overlap transient, multiplicative gating renders this transient retrieval permanent when the gates act as fully open or fully closed switches, while graded gating substantially prolongs its lifetime The dynamics show that introducing the self-adaptive gating fundamentally alters this behavior, extending the high-overlap plateau significantly as the steepness γ increases

Emergent Multistability and Landscape Reshaping

The gating mechanism leads to a novel memory retrieval phase of continuous multistability, where the system supports a continuous manifold of stable fixed-point attractors In the binary gating limit γ → ∞, the system reaches a continuum of distinct fixed points depending on the specific initial overlap, creating a “cloud” of stable fixed points This suggests that in the gated networks, the steady-state overlap can be continuously controlled by choosing a different initial neural condition rather than being quantized to a stored pattern

Thermodynamic Verification and Spectral Analysis

The findings are corroborated by Dynamical Mean-Field Theory (DMFT), which establishes that the observations made in finite-size simulations persist in the thermodynamic limit and in the long-time limit The analysis of the instantaneous Jacobian reveals a characteristic spectral signature of frozen-stabilization, which is reproduced during retrieval This stabilization mechanism involves a subpopulation ceasing to evolve and supplying a static background field to the rest, which can stabilize active units away from the high-gain region

Robustness and Biological Relevance

The study demonstrates that hard-gate retrieval remains accessible even when modulatory variables evolve substantially more slowly than the neural variables, depending strongly on the initial gate configuration Furthermore, the results show that gated retrieval is largely robust to rank and density of the modulatory coupling W in the hard-gate limit, requiring W to be neither dense nor full rank for finite gate steepness The emergent multistability suggests that the retrieved state is set continuously by the initial overlap, a feature that can be exploited for input-driven computation with persistent activity The utility of gating may extend from biological systems to machine learning more broadly, such as in gated attention in large language models

Conclusion and Outlook

The work demonstrates that a minimal model of self-adaptive gating fundamentally enhances the classical associative memory properties by circumventing catastrophic spin-glass transitions and inducing emergent multistability The key insights are the circumvention of the capacity ceiling, the stabilization of transients into true fixed points, and the emergence of a continuous manifold of stable fixed points Future work should focus on learning Wij couplings from measurements and adapting the model for spiking networks The gating-induced stabilization of ghost-like transients is identified as a novel physical perspective driving extended retrieval without assuming a Lyapunov function The utility of gating may extend to modeling complex biological systems where neuropeptides are essential for capturing key aspects of brain function The retrieved state being set continuously by the initial overlap is a feature that can be exploited for input-driven computation with persistent activity The study suggests that neuromodulatory gating may provide richer computational capabilities in biological and neuromorphic architectures

Supplementary Information

The supplementary material provides detailed methods, including direct many-body simulations, DMFT derivations, and analysis of the instantaneous Jacobian spectrum This includes illustrations of realistic neuromodulator networks and the finite-size behavior of the retrieval boundary The results confirm that the spread of steady-state overlaps vanishes as N → ∞, meaning mss is a deterministic function of α and m(0) in the thermodynamic limit The analysis confirms that the gating mechanism inherently induces a certain type of effective higher-order interactions, such as three-body interactions in the adiabatic limit The results confirm that the gated network bypasses the sharp order–to–spin-glass transition, extending the mss = O(1) retrieval phase beyond the standard critical capacity threshold αc The paper concludes that neuromodulatory gating provides richer computational capabilities in biological and neuromorphic architectures

References

The paper references various foundational works in associative memory, spin glasses, and neural dynamics Key references include Hopfield networks [67], dynamical mean field theory [77], and works on neuropeptide connectomes in C. elegans [14] The literature also includes studies on long short-term memory (LSTM) architectures [27, 28] and neuromodulation effects on neuronal conductivity The paper discusses conceptual parallels with previous studies on three-body interactions in coupled phase oscillators The supplementary references provide further context on the mathematical formalism, including path-integral methods and spectral decomposition of the Jacobian The paper also discusses the distinction between neuromodulation and synaptic plasticity mechanisms

Contents

The contents outline the structure of the paper, covering definitions, main results from simulations, DMFT formalism, visualization of continuous multistable attractors, and a summary of findings The supplementary information includes detailed methods for direct many-body simulations and numerical integration of DMFT equations The paper summarizes the key insights regarding the impact of gating on memory retrieval properties, including the circumvention of catastrophic spin-glass transition and the emergence of continuous multistability The paper provides a natural question about how applicable these findings are for biological systems across all species investigated The paper notes that dopamine has been found to lead to increased working memory in the hippocampus by affecting individual neurons The paper suggests that the neuromodulatory couplings Wij should contain structure and low rank based on the neuropeptidergic connectome of C. elegans The paper notes that Costacurta et al. p. 9

Where the paper says it

Furthermore, the emergence of continuous multistable attractors through gating means that the retrieved state is set continuously by the initial overlap, rather than being quantized to a stored pattern: a feature that can be exploited for input-driven computation with persistent activity [88–91]. In fact, Costacurta et al.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved memory retrieval capacity by bypassing catastrophic failure at high loads, allowing for robust computation far beyond standard Hopfield limits. The system can maintain order-one retrieval overlap even when the load exceeds a critical capacity of 0.13 without collapsing into a spin-glass phase.

  2. Emergent continuous multistability in the attractor landscape, where the retrieved state varies continuously with initial overlap rather than being quantized to a stored pattern. This allows for input-driven computation with persistent activity by setting the retrieved state by the initial overlap rather than quantized to the stored pattern.

  3. Stabilization of transient dynamics into true fixed points during retrieval, converting transient states into stable attractors in the binary gating limit. This mechanism is analogous to frozen-stabilization and can be exploited for input-driven computation with persistent activity.

  4. System behavior can be controlled by the initial gate configuration rather than just a single initial overlap value, as shown by how the dependence on z(0) provides an additional means of selecting persistent hard-gated states.

  5. The system's retrieval boundary can be tuned using an operational threshold theta to either broaden or narrow the range of cues from which retrieval occurs, allowing for tailored memory access based on fidelity requirements.

Abstract

Classical autoassociative memory models have been central to understanding emergent computations in recurrent neural circuits across diverse biological contexts. However, they typically neglect neuromodulatory agents that are known to strongly shape memory capacity and stability. Here we introduce a minimal, biophysically motivated associative memory network in which neuromodulator-like signals are modeled by a self-adaptive, activity-dependent gating mechanism. Using many-body simulations and dynamical mean-field theory, we show that such gating fundamentally reorganizes the attractor structure: the network bypasses the classical spin-glass transition, maintaining robust, high-overlap retrieval far beyond the standard critical capacity. Mechanistically, when the gates act as fully open or fully closed switches, the gate stabilizes transient "ghost" remnants of stored patterns even far above the Hopfield limit, converting a single memory fixed point attractor into a cluster of fixed points. The retrieved state is set by the initial overlap rather than quantized to the stored pattern. This graded dependence may be exploited for input-driven computation with persistent activity. These results establish that neuromodulation-like gating alone, without modifying synaptic weights or introducing explicit higher-order couplings, can dramatically expand memory capacity, eliminate catastrophic breakdown, and reshape a discrete attractor landscape into a rich continuum, suggesting a general and biologically accessible route to enhanced computational capabilities in neural circuits and neuromorphic architectures.

Sources

Related papers