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

Technical Summary: Neuromodulation–Inspired Gated Associative Memory Networks

Problem Statement
Classical autoassociative memory models, such as Amari–Hopfield recurrent neural networks (RNNs), have been foundational for understanding emergent computations in biological circuits. However, these models typically rely on static pairwise Hebbian couplings and neglect the influence of neuromodulatory agents (e.g., neuropeptides and bioamines). While it is known that neuromodulators shape memory capacity and stability, their collective computational consequences remain poorly understood theoretically. Specifically, it is unknown how activity-dependent neuromodulation affects the nervous system's capacity for memory retrieval and computability, particularly regarding the "memory cliff" or catastrophic breakdown that occurs when the memory load exceeds the critical capacity ($\alpha_c$).

Methodology
The authors introduce a minimal, biophysically motivated model of a self-adaptively gated associative memory network. The architecture consists of two recurrently connected layers:

1. Primary Layer: $N$ neurons with activity $x(t)$, governed by standard Hebbian couplings $J_{ij}$ derived from $P$ stored patterns.
2. Auxiliary Layer: $N$ units $z(t)$ mimicking neuromodulators, which multiplicatively gate the primary dynamics.

The dynamics are governed by coupled differential equations where the neuronal update rate is modulated by a sigmoidal function $\sigma(z_i)$ of the neuromodulator state. The neuromodulator state $z_i$ evolves based on the integrated activity of the neuronal layer, creating an activity-dependent feedback loop. The gating function $\sigma(z)$ acts as a switch-like control on the effective integration time-constant of the neurons.

The study employs two primary analytical and numerical approaches:

• Many-Body Simulations: Direct numerical integration of the coupled equations for finite system sizes ($N=1000$) to observe transient and steady-state behaviors.
• Dynamical Mean-Field Theory (DMFT): A theoretical framework using a bipartite cavity approach to derive self-consistent stochastic equations for the order parameters (overlaps) in the thermodynamic limit ($N, P \to \infty$). This allows for the analysis of long-time asymptotic behavior and the verification of finite-size simulation results.

Key Contributions and Results

1. Extension of Memory Retrieval Beyond Critical Capacity:
   In standard ungated networks ($\gamma=0$), memory retrieval fails catastrophically when the memory load $\alpha$ exceeds the critical capacity $\alpha_c \approx 0.13$, leading to a spin-glass phase. The authors demonstrate that the gated network ($\gamma \to \infty$) bypasses this sharp transition. Even in the overload regime ($\alpha > \alpha_c$), the network maintains robust, high-overlap retrieval. While the ungated network exhibits only transient retrieval (lingering near the pattern before decaying), the gated network stabilizes these transient states into permanent fixed points.

2. Mechanism of Stabilization (Frozen-Stabilization):
   The paper identifies the mechanism behind this enhancement as "frozen-stabilization." In the binary gating limit, the gate acts as a fully open or fully closed switch. A subset of neurons becomes "frozen" (closed gate, $\sigma=0$), effectively halting their dynamics. These frozen neurons stabilize the remaining "active" neurons, preventing the system from escaping the vicinity of the stored pattern. This converts what would be a transient "ghost" attractor (a region of slow dynamics near a destabilized fixed point) into a stable fixed point.

3. Emergent Continuous Multistability:
   A significant finding is the transformation of the attractor landscape. In classical models, the system relaxes to discrete fixed points corresponding to stored patterns. In the gated network, the steady-state overlap $m_{ss}$ depends continuously on the initial overlap $m(0)$. Instead of quantizing to a stored pattern, the system settles into a continuous manifold of stable fixed points. This "cloud" of attractors allows for graded retrieval states, enabling input-driven computation with persistent activity where the output is determined by the strength of the initial cue rather than a binary recall.

4. Theoretical Validation:
   The DMFT analysis confirms that these phenomena persist in the thermodynamic limit. The time-asymptotic DMFT solutions reveal a continuous band of steady-state overlap values for the gated network, bounded by the overlap of the closed-gate subpopulation. This validates that the extended retrieval and continuous multistability are intrinsic features of the gating mechanism, not artifacts of finite system size.

Significance
The paper claims that neuromodulatory gating alone—without modifying synaptic weights or introducing explicit higher-order couplings—is sufficient to overcome fundamental limitations of classical associative memory. Specifically, it:

• Eliminates the catastrophic capacity ceiling, allowing retrieval far beyond the Hopfield limit.
• Reshapes the discrete attractor landscape into a rich continuum, enabling graded dependence of the retrieved state on the initial input.
• Suggests a biologically accessible route to enhanced computational capabilities in neural circuits and neuromorphic architectures.

The authors position this work as a bridge between biological observations of neuropeptidergic networks (such as in C. elegans) and computational models, proposing that activity-dependent multiplicative gating is a general mechanism for expanding memory capacity and enabling complex, persistent computational states. They note that while the binary limit ($\gamma \to \infty$) provides a theoretical idealization, the finite-$\gamma$ regime (where gates are not perfectly closed) is more biologically apt, offering long-lived transient states that persist over behavioral timescales.