DiveMem: GMLG in Devices & Neural Systems

• DiveMem is a class of mechanisms using Globally-Modulated Local Gating to coordinate precise local updates with system-wide contextual signals.
• Its applications range from nanoscale charge gating in 2D materials and mixture-of-experts neural architectures to robust Bayesian filtering and lifelong learning in SNNs.
• By enabling context-aware modulation, DiveMem enhances performance, efficiency, and adaptability across condensed matter devices, neural networks, and signal processing systems.

DiveMem refers to a class of computational mechanisms—most precisely characterized as Globally-Modulated Local Gating (GMLG)—that coordinate localized, high-resolution updates or activations (“local” gates) with system-wide, context-dependent signals (“global” modulators). Originally motivated both by advances in condensed matter device engineering and by neural and artificial computing architectures, DiveMem mechanisms are used to sharply modulate memory, information flow, or material properties with nanoscale spatial precision or context-dependent selectivity. This entry provides a technical survey of the concept, underlying models, and principal applications across physical devices, neural network architectures, Bayesian filtering, and lifelong learning systems.

1. Physical Realization: Nanoscale Charge Gating in 2D Materials

DiveMem mechanisms first arose in the context of dielectrics-free, self-aligned gating of two-dimensional (2D) conductors, most notably in graphene-based optoelectronic devices. In this realization, a two-level structure is used: a global solid-polymer electrolyte gate (Poly(ethylene oxide)–LiClO₄) covers the active device, while a lithographically patterned, ion-impenetrable resist mask (cross-linked PMMA) defines lateral boundaries for local electrostatic doping. Under a top-gate voltage, mobile ions accumulate only on unmasked regions, forming a sharp electrical double layer (EDL) with a Debye length of approximately 1 nm.

The carrier density contrast is set by the effective specific capacitance of the EDL:

$$C_\mathrm{eff} \simeq \frac{\varepsilon_\mathrm{elec}}{\lambda_D}$$

yielding $C_\mathrm{eff} \approx 3.2$–$4.4\,\mu$F cm⁻². The local charge density step between masked and unmasked regions is:

$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$

Charge density modulations of up to $2\times10^{14}$ cm⁻² and in-plane fields exceeding $6\times10^{8}$ V m⁻¹ are achieved across lateral transition zones as narrow as 10 nm—resolution set by the convolution of Debye length and lithography (Peng et al., 2016).

This GMLG realization enables both abrupt and smoothly-resolved spatial doping profiles, supporting functionality such as locally tunable p–n junctions (demonstrated by intersecting Dirac ridges in four-quadrant resistance maps) and periodic thermoelectric arrays (e.g., radial thermopile photodetectors with responsivities up to 26.2 mV W⁻¹). Table 1 summarizes key device performance metrics:

Parameter	Typical Value	Setting
Δn (carrier contrast)	$>10^{14}$ cm⁻²	PEO–LiClO₄, $V_\mathrm{tg}\lesssim 2$ V
Electric field, Eₓ	$6\times10^8$ V m⁻¹	Δn across 10 nm boundary
Mask resolution	10 nm half-pitch	E-beam cross-linked PMMA
Leakage current	$<10$ pA	PMMA mask, cyclic voltammetry

2. Algorithmic GMLG: Mixture-of-Experts and Sparse Attention

The DiveMem (GMLG) principle is central to efficient conditional computation architectures, especially in multimodal mixture-of-experts models and sparse attention mechanisms.

In Mixture-of-Experts Multimodal LLMs (MLLMs), MoDES implements GMLG by combining layer-wise global importance scores with local per-token routing probabilities. Each MoE layer $C_\mathrm{eff} \approx 3.2$0 computes local gating probabilities via softmax over router logits:

$C_\mathrm{eff} \approx 3.2$1

Global layer importance $C_\mathrm{eff} \approx 3.2$2 is calibrated as the average Kullback-Leibler divergence between model outputs when layer $C_\mathrm{eff} \approx 3.2$3 is ablated:

$C_\mathrm{eff} \approx 3.2$4

Normalized importance $C_\mathrm{eff} \approx 3.2$5 multiplies the local score, yielding the final selection criterion:

$C_\mathrm{eff} \approx 3.2$6

Dual-Modality Thresholding (DMT) then sets separate expert-skip thresholds for text and visual tokens, with these thresholds learned via a monotonic “frontier search” to optimize accuracy vs. expert ratio (Huang et al., 19 Nov 2025). MoDES achieves up to 88% expert skipping at less than 0.1% accuracy loss on large MLLMs, with up to 2.16× prefilling and 1.26× decoding speedups.

Sparse attention via All-or-Here Attention (AHA) implements a binary-choice GMLG: each token–head pair chooses, by a router and hard thresholding, between executing local sliding-window or full global attention. The router is trained with an L1 regularizer to bias toward locality. For a window of 256 tokens, 93% of heads activate only local attention, reducing quadratic complexity to $C_\mathrm{eff} \approx 3.2$7 with virtually no quality drop (Luo et al., 27 Dec 2025).

3. Lifelong Learning and Neuromorphic SNNs

In neuromorphic and lifelong learning settings, DiveMem mechanisms govern the coupling of fast, context-driven local plasticity and slow, task-general global learning. In spiking neural networks (SNNs), GMLG is instantiated by two parallel synaptic weight traces:

• $C_\mathrm{eff} \approx 3.2$8, trained by (global) surrogate gradient backpropagation.
• $C_\mathrm{eff} \approx 3.2$9, trained locally by Hebbian or spike-timing-dependent plasticity (STDP).

The effective synaptic strength is given by:

$4.4\,\mu$0

where the context-dependent gating mask $4.4\,\mu$1 is a sigmoidal function of the local weights. Training alternates between local phase (updating $4.4\,\mu$2 via STDP in the current context/task) and global phase (updating $4.4\,\mu$3, gated by the mask, to optimize reward or classification objective) (Shen et al., 2024).

Empirically, GMLG SNNs achieve $4.4\,\mu$4 retention on previously learned tasks in blocked curricula (vs. $4.4\,\mu$5 for standard SNNs), match human “congruency effects,” and yield task-selective neurons. The mechanism is fully compatible with parallel on-chip neuromorphic implementations and is biologically congruent with prefrontal cortex context gating.

4. Robust Bayesian Filtering: Potential-Energy Gating

DiveMem also encompasses mechanisms for physics-informed robust filtering in stochastic dynamical systems. In potential-energy gating, the local value of a known or assumed potential $4.4\,\mu$6 modulates the observation noise covariance in Bayesian filters:

$4.4\,\mu$7

where $4.4\,\mu$8 is the gating sensitivity parameter. When the state $4.4\,\mu$9 is near a well of $$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$0, observations are trusted ($$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$1); near the barrier, measurement trust is down-weighted. This continuous, energy-driven gating achieves $$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$2–$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$3 RMSE improvements over standard EKF and particle filters, especially under observation outliers or during Kramers-type barrier-crossing events (Simeone, 12 Feb 2026).

Table 2 lists RMSE improvements in EKF and PF variants:

Filter Type	RMSE Gain	Outlier Fraction
PG-EKF	$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$4	$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$5
PG-PF	$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$6	$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$7
NT-EKF	$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$8	$$\Delta n \simeq \frac{C_\mathrm{eff}(V_\mathrm{tg} - V_\mathrm{th})}{e}$$9

Parameter robustness is demonstrated for $2\times10^{14}$0 misspecification in $2\times10^{14}$1 (potential shape).

5. Extensions in Neural Architectures

Context-Gated Convolution (CGC) is a further extension of DiveMem to convolutional neural networks. Global context, extracted by pooling and shallow projections, is decoded into a gating tensor $2\times10^{14}$2 that multiplicatively modulates the spatial kernel weights $2\times10^{14}$3:

$2\times10^{14}$4

All convolutions with $2\times10^{14}$5 are replaced by this context-adaptive kernel. Across vision (ImageNet, Kinetics) and sequence tasks (IWSLT machine translation), CGC consistently yields $2\times10^{14}$6–$2\times10^{14}$7 absolute accuracy gains over strong baselines with $2\times10^{14}$8 run-time overhead (Lin et al., 2019).

CGC confers neuron-level adaptivity (“adaptive processor” functionality) analogous to PFC context gating, endowing deterministic CNNs with sample-level receptive field plasticity—a direct realization of the globally-modulated, locally-applied gating concept.

6. Theoretical and Implementation Considerations

The DiveMem (GMLG) paradigm can be summarized as the composition of:

• Global context extraction, quantifying system- or layer-wide importance or physical context (e.g., layer sensitivity, spatial pool, potential energy).
• Local gating application, where the global signal modulates localized computation, connectivity, or inference (e.g., expert selection, synapse activation, observation trust).

Resolution or selectivity is jointly controlled by the fidelity of local gating (e.g., lithographic sharpness, router accuracy, STDP learning) and the strength/specificity of the global modulator (e.g., Debye length, importance weighting, physical energy). Implementation constraints are application-specific: mask thickness and exposure in device fabrication; router sparsity and thresholding for model inference; or local plasticity and context selectivity in SNNs.

The paradigm supports both hard (binary or blockwise) and soft (continuous or probabilistic) gating, and is compatible with multi-modality, dynamic thresholding, or hierarchical context routing.

7. Empirical Performance and Impact

Across all identified domains, DiveMem mechanisms yield substantial gains in efficiency, robustness, and selectivity without significant loss in task performance. In multimodal mixture-of-experts settings, aggressive resource reduction (up to 88% expert skipping) maintains over 99.9% accuracy (Huang et al., 19 Nov 2025). In SNN lifelong learning, GMLG resolves catastrophic forgetting and matches human task selectivity (Shen et al., 2024). For robust Bayesian filtering, energy-aware gating improves estimation under heavy-tailed noise by up to 80% (Simeone, 12 Feb 2026). In CNNs, context-gated convolution yields systematic accuracy improvements for minimal overhead (Lin et al., 2019).

DiveMem, instantiated as GMLG, thus provides a general, principled template for coordinating localized precision with context-aware global modulation; empirical results from condensed matter, signal processing, and large-scale neural computation attest to both the versatility and practical efficacy of this mechanism family.