CoChat

Input-driven dynamics for robust memory retrieval in Hopfield networks

Open access

The DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index; 1 of the 9 sources its claims cite were checked for support.

Abstract

The Hopfield model provides a mathematical framework for understanding the mechanisms of memory storage and retrieval in the human brain. This model has inspired decades of research on learning and retrieval dynamics, capacity estimates, and sequential transitions among memories. Notably, the role of external inputs has been largely underexplored, from their effects on neural dynamics to how they facilitate effective memory retrieval. To bridge this gap, we propose a dynamical system framework in which the external input directly influences the neural synapses and shapes the energy landscape of the Hopfield model. This plasticity-based mechanism provides a clear energetic interpretation of the memory retrieval process and proves effective at correctly classifying mixed inputs. Furthermore, we integrate this model within the framework of modern Hopfield architectures to elucidate how current and past information are combined during the retrieval process. Last, we embed both the classic and the proposed model in an environment disrupted by noise and compare their robustness during memory retrieval.

Overview

CoChat Summary

## TL;DR

This paper proposes an input-driven plasticity (IDP) variant of the Hopfield model in which external input directly modulates synapses, reshaping the energy landscape to support more robust memory retrieval. The authors show analytically and in simulations that the model can suppress mixed-input ambiguity, create a “confusion state” when inputs are too weakly informative, and be more resilient to noise than the classic Hopfield model.

## Key findings

  • In the IDP model, a retrievable memory \(x_\mu=\gamma_\mu \xi^\mu\) exists only if its saliency weight satisfies \(\alpha_\mu > \alpha_{\text{existence}} = 1\); stability requires the stronger condition \(\alpha_\mu > \alpha_{\text{stability}}\) (with \(\alpha_{\text{stability}} > 1\)).
  • The paper states that the modified energy landscape ranks memories by saliency: if \(\alpha_\nu < \alpha_\mu\), then \(E(x_\mu;W(u)) < E(x_\nu;W(u))\), so stronger input components create deeper minima.
  • For inputs that are well mixed and satisfy \(\alpha_\mu < 1\) for all memories, the IDP dynamics converge to the origin, which the authors call a “confusion state.”
  • In noise simulations, the classic Hopfield model is reported to fail under moderate noise: at \(\sigma=1\), overlaps remain at chance level, and larger \(\sigma\) leads to total retrieval failure; by contrast, the IDP model shows correct sequential retrieval even at relatively high noise and performs best at an intermediate value, \(\sigma=2\), for retrieval timing.

## Methods

  • The paper develops a continuous-time Hopfield framework with input-modulated synapses: \( \dot{x}(t) = -x(t) + W(u(t))\Psi(x(t)) \), where \(W(u(t))\) depends on saliency weights \(\alpha_\mu(t) = [\xi_\mu^\top u(t)]^2\).
  • Simulations for Figs. 2 and 5 use networks of \(N=1024\) neurons, Euler-Maruyama integration, and activation \(\psi(x_i)=\tanh(5x_i)\); inputs are piecewise constant and constructed as mixtures \(u=\sum_\mu \omega_\mu \xi^\mu\) with dominant weights sampled from \(U[7,10]\) and non-dominant weights from \(U[2,5]\) over intervals of length 10.
  • The analysis combines energy-function arguments, existence/stability thresholds for equilibria, and stochastic dynamics with additive Brownian noise \( \sigma dB(t) \) to compare classic and IDP Hopfield behavior.

## Limitations

  • The paper explicitly notes that a full theoretical investigation of the biologically plausible firing-rate version of the IDP model is future work, and that empirical validation of input-dependent associative memory models is still needed.
  • It also states that a complete characterization of learning dynamics, including long-term memory formation and short-term modulation of saliency weights, remains open.
  • Evident from the text: the simulations use ad hoc piecewise-constant inputs constructed for experimental control, so the reported robustness results are tied to a stylized setup rather than naturalistic input streams.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1126/sciadv.adu6991

Cites

47 of 47 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1. Comparison between classic Hopfield and IDP Hopfield models. (A) A slowly morphing sequence of noisy images is presented as a input to the observer, who updates its belief state to retrieve the memory closest to the current image u. This adaptation process occurs continuously. (B) In the classic model, the network state is set to an initial condition x(0) equal to the current image u, and, then, the Hopfield dynamics performs the memory retrieval task. (C) In the proposed input-driven plasticity (IDP) model, the network initial condition is arbitrary, the image u modifies the synaptic weights W(u), and the Hopfield dynamics with modified synaptic weights performs the memory retrieval task. This dynamics is well posed and naturally tracks the morphing images also when the image is time-varying u=u(t). (D) In the classic model, the Hopfield dynamics is a gradient descent for the energy E(x;W): The blue ball, representing the neural state, rolls from an initial condition toward a stable minimum point (cat memory). Therefore, the retrieval process is successful when the noisy image x(0)=u (dotted cat) lies in the region of attraction of the correct memory (cat memory). (E) In the proposed model, the noisy image u directly modifies the synaptic matrix W(u) and, in turn, the energy landscape E[x;W(u]), thereby extending the region of attraction of the correct memory. The retrieval process is successful from generic initial conditions when the correct memory is the unique minimum of the landscape. Specifically, in the panel, the noisy image (dotted cat) renders the correct memory a minimum (cat memory) and the incorrect memory (dog memory) no longer an equilibrium of E[x;W(u]).
Fig. 1. Comparison between classic Hopfield and IDP Hopfield models. (A) A slowly morphing sequence of noisy images is presented as a input to the observer, who updates its belief state to retrieve the memory closest to the current image u. This adaptation process occurs continuously. (B) In the classic model, the network state is set to an initial condition x(0) equal to the current image u, and, then, the Hopfield dynamics performs the memory retrieval task. (C) In the proposed input-driven plasticity (IDP) model, the network initial condition is arbitrary, the image u modifies the synaptic weights W(u), and the Hopfield dynamics with modified synaptic weights performs the memory retrieval task. This dynamics is well posed and naturally tracks the morphing images also when the image is time-varying u=u(t). (D) In the classic model, the Hopfield dynamics is a gradient descent for the energy E(x;W): The blue ball, representing the neural state, rolls from an initial condition toward a stable minimum point (cat memory). Therefore, the retrieval process is successful when the noisy image x(0)=u (dotted cat) lies in the region of attraction of the correct memory (cat memory). (E) In the proposed model, the noisy image u directly modifies the synaptic matrix W(u) and, in turn, the energy landscape E[x;W(u]), thereby extending the region of attraction of the correct memory. The retrieval process is successful from generic initial conditions when the correct memory is the unique minimum of the landscape. Specifically, in the panel, the noisy image (dotted cat) renders the correct memory a minimum (cat memory) and the incorrect memory (dog memory) no longer an equilibrium of E[x;W(u]).CC BY-NC · PubMed Central
Fig. 2. Exploration of the response of different associative memory models to a time-varying input. Each retrieval plot displays a sample average of the memory overlap parameter over 50 retrieval tasks, with the associated SD shown through the shaded, color-matched surrounding area. (A to C) The response of three different models to a time-varying input. The simulation time is segmented, with each subinterval featuring a constant input. The horizontal bars on top of each retrieval plot display the duration of the external input and the color associated to its dominant saliency weight. At each external input switch, the previous dominant saliency weight shrinks below the stability threshold αstability. (A) Classic Hopfield model (Eq. 4). The network dynamics converge to a mixed state, precluding exact retrieval of any individual memory. (B) Input-modulated Hopfield model (Eq. 5). The modulator (Eq. 6) shuts off the input after intervals of t=2 simulation time. The network dynamics then freely recall the prototypical memory associated with the dominant component in the input. (C) IDP Hopfield model (Eq. 7). The network naturally recalls the prototypical memory associated with the dominant component in the input. The model exhibits memory retention even after input switching, showcasing an intriguing integrative feature for past and present information.
Fig. 2. Exploration of the response of different associative memory models to a time-varying input. Each retrieval plot displays a sample average of the memory overlap parameter over 50 retrieval tasks, with the associated SD shown through the shaded, color-matched surrounding area. (A to C) The response of three different models to a time-varying input. The simulation time is segmented, with each subinterval featuring a constant input. The horizontal bars on top of each retrieval plot display the duration of the external input and the color associated to its dominant saliency weight. At each external input switch, the previous dominant saliency weight shrinks below the stability threshold αstability. (A) Classic Hopfield model (Eq. 4). The network dynamics converge to a mixed state, precluding exact retrieval of any individual memory. (B) Input-modulated Hopfield model (Eq. 5). The modulator (Eq. 6) shuts off the input after intervals of t=2 simulation time. The network dynamics then freely recall the prototypical memory associated with the dominant component in the input. (C) IDP Hopfield model (Eq. 7). The network naturally recalls the prototypical memory associated with the dominant component in the input. The model exhibits memory retention even after input switching, showcasing an intriguing integrative feature for past and present information.CC BY-NC · PubMed Central
Fig. 3. Energy landscapes for IDP Hopfield model for varying saliency weights. Stable and unstable equilibria are depicted as green stars and orange dots, respectively. Recall the existence threshold is αexistence=1 and, when multiple memories exist in the input, the stability threshold satisfies αstability>1. (A) “No memories” α1<1, α2<1: When no memory is sufficiently strong in the input, the only global minimum is at the origin, and it is globally attractive for the dynamics. This situation corresponds to a confusion state for the network, in which the input is not strong enough to evoke any retrieval. (B) “One memory” α1<1, α2>1: When the saliency weight for precisely one memory is above αexistence, two symmetric equilibria appear corresponding to the memory, and they are attractive from almost all initial conditions. In this case, the origin is a saddle point. (C) “one stable and one unstable memory” α2>αstability>α1>1: Two symmetric equilibria appear corresponding to the stable memory, and they are attractive from almost all initial conditions. The other two symmetric equilibria are saddle points and the origin becomes an unstable maximum. (D) “Two stable memories” α2>α1>αstability: Four symmetric stable equilibria appear, corresponding to the two memories of the model. The memory associated to the dominant saliency weight carves deeper valley in the energy landscape than the other memory. When α1≈αstability, the shallowness of the valley associated to the first memory facilitates outward jumps due to stochastic fluctuations.
Fig. 3. Energy landscapes for IDP Hopfield model for varying saliency weights. Stable and unstable equilibria are depicted as green stars and orange dots, respectively. Recall the existence threshold is αexistence=1 and, when multiple memories exist in the input, the stability threshold satisfies αstability>1. (A) “No memories” α1<1, α2<1: When no memory is sufficiently strong in the input, the only global minimum is at the origin, and it is globally attractive for the dynamics. This situation corresponds to a confusion state for the network, in which the input is not strong enough to evoke any retrieval. (B) “One memory” α1<1, α2>1: When the saliency weight for precisely one memory is above αexistence, two symmetric equilibria appear corresponding to the memory, and they are attractive from almost all initial conditions. In this case, the origin is a saddle point. (C) “one stable and one unstable memory” α2>αstability>α1>1: Two symmetric equilibria appear corresponding to the stable memory, and they are attractive from almost all initial conditions. The other two symmetric equilibria are saddle points and the origin becomes an unstable maximum. (D) “Two stable memories” α2>α1>αstability: Four symmetric stable equilibria appear, corresponding to the two memories of the model. The memory associated to the dominant saliency weight carves deeper valley in the energy landscape than the other memory. When α1≈αstability, the shallowness of the valley associated to the first memory facilitates outward jumps due to stochastic fluctuations.CC BY-NC · PubMed Central
Fig. 4. Illustration of the modern Hopfield reformulation of the dynamics for the IDP Hopfield model with input filter. The symbol ⊙ denotes a Hadamard entrywise product. (A) Neural network representation with interconnected layers and synaptic weights Mz, Mx, and My. (B) Block diagram representation, where each block 1/(1+τs), with τ=τz,τx,τy, denotes a low-pass filter with cutoff frequency 1/τ, each block Mz, Mx, and My denotes a matrix-vector multiplication, and each block Ψz(⋅), Ψx(⋅), and Ψy(⋅) denotes an activation function.
Fig. 4. Illustration of the modern Hopfield reformulation of the dynamics for the IDP Hopfield model with input filter. The symbol ⊙ denotes a Hadamard entrywise product. (A) Neural network representation with interconnected layers and synaptic weights Mz, Mx, and My. (B) Block diagram representation, where each block 1/(1+τs), with τ=τz,τx,τy, denotes a low-pass filter with cutoff frequency 1/τ, each block Mz, Mx, and My denotes a matrix-vector multiplication, and each block Ψz(⋅), Ψx(⋅), and Ψy(⋅) denotes an activation function.CC BY-NC · PubMed Central
Fig. 5. Robustness to noise in the IDP and classic Hopfield model. (A and B) Differently from Fig. 2C, the prototypical memory associated to the red weight remains stable after each input switching. (A) IDP model insulated from noise. The model correctly retrieves the first prototypical memory but is unable to track the dominant component of next inputs due to persistent stability of the red component. (B) IDP model with high amplitude noise. The model correctly retrieves the first prototypical memory. At input switch, the memory remains stable, but the associated energy well considerably flattens (see Fig. 3D as reference case). The noise pushes the activity out of the flat wells and directs it toward the deepest one, which is associated to the dominant saliency weight. The process is then repeated at each input switch. (C) The classic Hopfield model preserves good memory retrieval performance only for small noise amplitudes. As the noise amplitude increases, the memory retrieval performance quickly deteriorates until total failure. (D and E) Robustness to brief input perturbations (glitches). (D) The IDP Hopfield model successfully corrects the glitch integrating past internal and present external information. (E) The classic Hopfield model fails at correcting the glitch and fails in the retrieval task once the input is switched off. (F) Noise amplitude and speed of the transient toward the next memory. Intermediate values of noise amplitude result in optimal speed for the transition between two memories. Insulated and high amplitude noise models take considerably longer to complete the switch.
Fig. 5. Robustness to noise in the IDP and classic Hopfield model. (A and B) Differently from Fig. 2C, the prototypical memory associated to the red weight remains stable after each input switching. (A) IDP model insulated from noise. The model correctly retrieves the first prototypical memory but is unable to track the dominant component of next inputs due to persistent stability of the red component. (B) IDP model with high amplitude noise. The model correctly retrieves the first prototypical memory. At input switch, the memory remains stable, but the associated energy well considerably flattens (see Fig. 3D as reference case). The noise pushes the activity out of the flat wells and directs it toward the deepest one, which is associated to the dominant saliency weight. The process is then repeated at each input switch. (C) The classic Hopfield model preserves good memory retrieval performance only for small noise amplitudes. As the noise amplitude increases, the memory retrieval performance quickly deteriorates until total failure. (D and E) Robustness to brief input perturbations (glitches). (D) The IDP Hopfield model successfully corrects the glitch integrating past internal and present external information. (E) The classic Hopfield model fails at correcting the glitch and fails in the retrieval task once the input is switched off. (F) Noise amplitude and speed of the transient toward the next memory. Intermediate values of noise amplitude result in optimal speed for the transition between two memories. Insulated and high amplitude noise models take considerably longer to complete the switch.CC BY-NC · PubMed Central

Verification

Claims the paper makes

1 Valid · 4 Not valid · 4 Not checked · 34 Own results and interpretations

23 more claims

Create a free account to see all 43 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 10
  • The quote only partly states this claim. 1
  • Probably not a substantial claim. 5

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 5

  1. H. Ramsauer B. Schäfl J. Lehner P. Seidl M. Widrich T. Adler L. Gruber M. Holzleitner M. Pavlović G. K. Sandve V. Greiff D. Kreil M. Kopp G. Klambauer J. Brandstetter S. Hochreiter “Hopfield networks is all you need” in International Conference on Learning Representations (ICLR 2021) pp. 1–95.
  2. P. Dayan L. F. Abbott Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems (MIT Press 2005).
  3. F. Moss P. McClintock Noise in Nonlinear Dynamical Systems (Cambridge Univ. Press 2012).
  4. A. Karuvally T. Sejnowski H. T. Siegelmann “General sequential episodic memory model” in International Conference on Machine Learning (PMLR 2023) pp. 15900–15910.
  5. H. Khalil Nonlinear Systems (Prentice Hall 2002).

Cited by

See all 12

Metadata

Authors
Betteti, S., Baggio, G., Bullo, F., & Zampieri, S.
Affiliations
  • University of Padua (IT)
  • University of California, Santa Barbara (US)
Corresponding author
  • Betteti, Simone
  • Baggio, Giacomo
  • Bullo, Francesco
  • Zampieri, Sandro
Venue
Science Advances
Published
2025-04-25
Volume
11
Issue
17
Pages
eadu6991-eadu6991
ISSN
2375-2548 (ISSN-L)
OpenAlex
W4409723675
Cited by
12
References
43
Citations by year
2026: 7 · 2025: 5
FWCI
4.81
Type
journal article
Language
en
Open access
Licence
cc-by-nc
Locations
  • Science Advances (publishedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • Padua Research Archive (University of Padova) (submittedVersion, cc-by-nc)
  • arXiv (Cornell University) (cc-by)
Funders
Air Force Office of Scientific Research (FA9550-, #FA9550-21-1-0203, -21-1-)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
Keywords
Hopfield neural network, memory retrieval, external inputs, synaptic plasticity, energy landscape, neural dynamics, memory storage, noise robustness, associative memory

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.