Metaneurons: Adaptive Meta-Dynamic Neurons

• Metaneurons are neuron models with adaptive internal dynamics and meta-parameters that regulate computation beyond fixed weights.
• They incorporate features like state-dependent routing and meta-homeostasis to optimize performance in spiking networks and neuromorphic hardware.
• Their applications span from improved accuracy in SNN experiments to scalable, energy-efficient implementations in physical devices.

Metaneurons is a heterogeneous term rather than a standardized canonical object in contemporary computational neuroscience and neuromorphic engineering. In explicit usages, it can denote higher-level or meta-dynamic neuron models in spiking neural networks; in broader, interpretive usages, it refers to neural units whose effective computation is controlled by adaptive internal dynamics, state-dependent routing, developmental or homeostatic set-points, bioenergetic state, or physically embodied material dynamics rather than by a fixed weighted-sum-and-threshold rule alone (Cheng et al., 2020, Rudnicka et al., 24 Aug 2025). Across these strands, the recurring theme is a shift from the classical neuron as a static node toward a neuron, neuronal fragment, microcircuit, or device whose computational role is itself modulated, reconfigured, or meta-specified.

1. Terminological scope and principal usages

The literature does not provide a single agreed definition of metaneurons. Some papers use the term directly, some introduce closely related constructs under different names, and some are best read as conceptual precursors. This makes the topic encyclopedic only in an umbrella sense: “metaneuron” names a family resemblance among models that place a control layer, developmental layer, or hidden physical layer above ordinary spike transmission.

Usage strand	Core unit	Representative papers
Explicit SNN meta-types	Finite library of neuron dynamics	(Cheng et al., 2020)
Higher-level computational unit	Generalized weighted-sum unit with flexible activation and optional dynamics	(Rudnicka et al., 24 Aug 2025)
Meta-homeostatic neuron	Neuron with adaptive target firing rate controlling other plasticity rules	(Tosi et al., 2017)
Biochemical routing substrate	MR/GPGIC-controlled branch and pathway selection	(Nikolić, 2022)
Physical artificial neuron	Material device whose intrinsic dynamics realize neuron-like behavior	(Mendolia et al., 30 Nov 2025, Zhang et al., 2023, Azam et al., 2018, Palin et al., 21 Apr 2026, Boer et al., 2024)
Collective metastable unit	Neural population acting as an emergent higher-level stateful unit	(Löcherbach et al., 2020, Paliwal et al., 2024)

Two terminological caveats are central. First, the word is explicit in only part of this literature. Nikolić’s proposal that cognition is implemented by metabotropic receptors and G protein-gated ion channels never defines a “metaneuron,” but it is strongly interpretable as a meta-neuronal framework because the decisive computational state is shifted from voltages and synaptic weights to membrane-level gating states that open and close branches and pathways (Nikolić, 2022). Second, some papers that explicitly compare “metaneurons” with other neuron models do not fully formalize the model they evaluate. In particular, the SNN study of neuron-model choice describes metaneurons as a “higher-level computational unit” with weighted input aggregation, flexible activation functions, and optional time-varying internal dynamics, but it does not provide a unique governing dynamical equation, threshold law, reset equation, refractory term, or parameter set for the metaneuron used in experiments (Rudnicka et al., 24 Aug 2025).

A useful synthesis is therefore this: metaneurons are best understood as units whose effective computation depends on meta-parameters, hidden state variables, structural adaptation, or higher-order control loops that are not reducible to a fixed neuron-plus-synapse primitive. That interpretation is explicit in some works and only suggestive in others.

2. Explicit computational formulations in spiking and deep-learning models

The most direct formalization of metaneurons appears in the literature on Meta-Dynamic Neurons (MDNs). In “Finite Meta-Dynamic Neurons in Spiking Neural Networks for Spatio-temporal Learning,” MDNs are a finite library of representative neuron dynamic types for SNNs, learned from source tasks and then transferred to new tasks with fixed dynamic parameters (Cheng et al., 2020). The core second-order MDN is defined by

$$\left\{ \begin{array}{l} \frac{dV_j(t)}{dt}=V_j(t)^2 - V_j(t) - U_j(t)+ \sigma(\sum_{i=1}^NW_{i,j}I_i(t))\ \frac{dU_j(t)}{dt}=\theta_a (\theta_b V_j(t)-U_j(t)) \ \begin{matrix} V_j(t)=\theta_c, & U_j(t)=U_j(t)+\theta_d & \text{if}(V_j(t)>V_{th}) \end{matrix} \ S_j(t) = V_j(t)>V_{th} \end{array} \right.$$

where the dynamic hyperparameters $(\theta_a,\theta_b,\theta_c,\theta_d)$ define the meta-type. The paper extracted spatial meta neurons from MNIST and temporal meta neurons from TIDigits, then retained four representative types: 2nd-FS, 2nd-RS, 2nd-SDS, and 2nd-WDS. Their transfer behavior was task-selective: spatial MDNs performed better on MNIST, Fashion-MNIST, NETtalk, and CIFAR-10, while temporal MDNs performed better on TIDigits and TIMIT. For example, on TIDigits the temporal MDNs reached $78.00 \pm 0.44$ and $76.32 \pm 1.08$, compared with $61.64 \pm 1.03$ for the first-order baseline; on Fashion-MNIST, the spatial MDNs reached $90.50 \pm 0.10$ and $89.58 \pm 0.07$, compared with $85.37 \pm 0.06$ and $85.79 \pm 0.07$ for the temporal MDNs (Cheng et al., 2020).

A looser but explicitly labeled usage appears in the SNN classification study comparing LIF, metaneurons, and probabilistic Levy–Baxter neurons. There, metaneurons are described as abstractions of collective neuron activity that compute a weighted sum like a perceptron, support binary step, sigmoid, or ReLU activations, and may process time-varying inputs with evolving internal states “akin to LIF,” yet the exact state equation is omitted (Rudnicka et al., 24 Aug 2025). Despite this under-specification, the reported empirical pattern is clear: metaneurons often achieved high accuracy with relatively small network sizes, especially $32$–$(\theta_a,\theta_b,\theta_c,\theta_d)$0 neurons per layer, and the paper states that they showed an accuracy range of $(\theta_a,\theta_b,\theta_c,\theta_d)$1 to $(\theta_a,\theta_b,\theta_c,\theta_d)$2 in SNNs with $(\theta_a,\theta_b,\theta_c,\theta_d)$3 to $(\theta_a,\theta_b,\theta_c,\theta_d)$4 neurons while other neuron models typically required $(\theta_a,\theta_b,\theta_c,\theta_d)$5 or more to achieve comparable performance (Rudnicka et al., 24 Aug 2025). This suggests that, within that paper’s vocabulary, a metaneuron is not a biophysical single-cell model but a scalable computational unit positioned between a perceptron and a fully specified spiking neuron.

An important deep-learning analogue is the Deep Artificial Neuron (DAN). In “Continual Learning with Deep Artificial Neurons,” each neuron in the outer network is itself a small neural network, and all such inner neuron-models share a single meta-learned parameter vector $(\theta_a,\theta_b,\theta_c,\theta_d)$6, called a neuronal phenotype (Camp et al., 2020). Synapses between neurons are vectorized rather than scalar, and the DAN computes a learned map from an $(\theta_a,\theta_b,\theta_c,\theta_d)$7-dimensional synaptic slice to a scalar activation. During meta-training, $(\theta_a,\theta_b,\theta_c,\theta_d)$8 is optimized across sequential tasks so that, at deployment, standard backpropagation on the plastic synaptic vectors $(\theta_a,\theta_b,\theta_c,\theta_d)$9 produces minimal forgetting while $78.00 \pm 0.44$0 remains fixed. In this sense DANs are metaneurons in a strong architectural sense: the neuron’s internal computation, rather than only the network weights, is the object of meta-learning (Camp et al., 2020).

A more weakly related architecture-level usage appears in the Metabolize Neural Network, or MetaNet. There, metaneurons are not new dynamical cells but ordinary hidden units embedded in a metabolism-inspired framework of neuron proliferation and autophagy during training (Dai et al., 2018). The central trigger is structural rather than electrophysiological:

$78.00 \pm 0.44$1

after which neurons are added, initialized by one of several schemes, and potentially removed if under-expressed (Dai et al., 2018). This usage broadens the term toward adaptive neuron populations rather than adaptive neuron dynamics.

3. Meta-adaptive biological abstractions: set-points, routing, and energetics

One major line of work relocates meta-neuronal adaptation to the biological neuron itself. In MANA, the defining innovation is meta-homeostatic plasticity (MHP), which updates each neuron’s target firing rate $78.00 \pm 0.44$2, the set-point used by ordinary homeostatic plasticity and synaptic normalization (Tosi et al., 2017). The key variable is therefore not membrane voltage but the homeostatic goal toward which lower-level plasticity rules drive the neuron. MHP is given by

$78.00 \pm 0.44$3

with $78.00 \pm 0.44$4. Lower-rate presynaptic neighbors push a postsynaptic target upward, higher-rate neighbors push it downward, and the resulting target firing rates self-organize into a broad distribution while shaping threshold homeostasis, current budgets, and topology (Tosi et al., 2017). MANA therefore provides a precise example of a neuron-level meta-state that governs the action of other plasticity rules.

A far more radical proposal is Nikolić’s theory of cognition as transient pathway selection by metabotropic receptors and G protein-gated ion channels. The paper states that “Our knowledge and skills are stored in metabotropic receptors (MRs) and the G protein-gated ion channels (GPGICs),” and that “the mental state is defined by the states of GPGICs, not by the state of neuron voltages” (Nikolić, 2022). MRs detect ligands, often extrasynaptic “runaway” neurotransmitters, and GPGICs then open or close dendritic branches and axon terminals for hundreds of milliseconds to minutes, transiently re-routing neural traffic. The paper’s core claim is that “The process of selecting this new subnetwork is what constitutes a mental operation - be it in a form of directed attention, perception or making a decision” (Nikolić, 2022). The term metaneuron is not used, and this must be treated as an interpretation, but the framework is suggestive because the effective computational state is displaced upward from spikes to a membrane-level routing layer.

A related but distinct biological reformulation appears in “Metaboplasticity: The Reciprocal Regulation of Neuronal Activity and Cellular Energetics.” Here the neuron remains a conductance-based LIF unit, but intrinsic excitability, synaptic kinetics, and STDP are modulated by a proxy for metabolic state implemented through temperature-dependent $78.00 \pm 0.44$5 scaling in a Brian2 microcircuit with $78.00 \pm 0.44$6 neurons (Öner et al., 25 Dec 2025). The core scaling is

$78.00 \pm 0.44$7

with $78.00 \pm 0.44$8 for membrane and synaptic kinetics and $78.00 \pm 0.44$9 for STDP amplitudes (Öner et al., 25 Dec 2025). The paper reports five emergent properties—dynamics bifurcation, STDP window deformation, signal degradation, topological shift, and parametric robustness—and argues for an inverted-U relationship between bioenergetics and learning. This suggests a metabolically regulated metaneuron concept in which the unit’s integration timescale and learning kernel depend on energetic state rather than remaining fixed.

Taken together, these papers define a biophysical meaning of metaneurons: a neuron whose function is governed by higher-order variables such as target rates, branch-gating molecules, or energy-state parameters. In each case, the decisive computational quantity is not merely the instantaneous voltage.

4. Material and neuromorphic embodiments

A large hardware literature implements metaneuron-like behavior directly in material dynamics. The clearest silicon example is “A Neuromodulable Current-Mode Silicon Neuron for Robust and Adaptive Neuromorphic Systems,” which presents a mixed-feedback current-mode neuron with one fast, one slow, and one ultraslow state, together with positive and negative feedback loops on multiple timescales (Mendolia et al., 30 Nov 2025). Its full current-mode model is

$76.32 \pm 1.08$0

with slow and ultraslow variables inactivating positive feedback via

$76.32 \pm 1.08$1

A single modulatory bias, the slow positive feedback gain $76.32 \pm 1.08$2, experimentally shifts the same physical neuron from tonic spiking to tonic bursting. The circuit was fabricated in $76.32 \pm 1.08$3 nm CMOS, implemented in a $76.32 \pm 1.08$4-neuron array, and is reported to operate with rest power about $76.32 \pm 1.08$5 nW and instantaneous spike power about $76.32 \pm 1.08$6 nW (Mendolia et al., 30 Nov 2025). Here the metaneuron is explicit: a neuron whose excitability class is selected by higher-level biases.

The electromechanical leaky memcapacitor provides a more physically compact embodiment. In “Electromechanical memcapacitive neurons for energy-efficient spiking neural networks,” the core device is a movable-plate capacitor whose state variable $76.32 \pm 1.08$7 simultaneously controls capacitance and leakage (Zhang et al., 2023). The device equations are

$76.32 \pm 1.08$8

$76.32 \pm 1.08$9

and

$61.64 \pm 1.03$0

with

$61.64 \pm 1.03$1

The paper shows that a single device can implement leaky integration, threshold-triggered discharge, self-sustained spiking, and synchronization, while richer regimes such as bursting require an added adaptive memristor (Zhang et al., 2023). This is metaneuronal in the sense that integration, leak, threshold, and reset are co-embodied in one material state variable rather than assembled from separate circuit blocks.

Other hardware papers instantiate metaneurons through specific physical modes. In the skyrmion resonate-and-fire neuron, a fixed magnetic skyrmion in a magnetic tunnel junction acts as the internal state, and its breathing mode under voltage-controlled anisotropy provides subthreshold oscillation and resonance; firing is defined by a threshold on the average out-of-plane magnetization, $61.64 \pm 1.03$2 (Azam et al., 2018). In the inhibitory neuristor based on metal-to-insulator transition, volatile low-to-high resistance switching in LSMO embedded in a simple RL circuit produces robust $61.64 \pm 1.03$3–$61.64 \pm 1.03$4 MHz oscillatory current suppression, positioned as an inhibitory complement to IMT excitatory neuristors (Palin et al., 21 Apr 2026). In the halide perovskite neuron, a volatile MAPbI$61.64 \pm 1.03$5-assisted memristive device in series with a $61.64 \pm 1.03$6 pF on-chip capacitor produces stochastic leaky integrate-and-fire behavior with $61.64 \pm 1.03$7 to $61.64 \pm 1.03$8 pJ per spike, and stochastic firing improves detection of sub-threshold inputs in populations (Boer et al., 2024). These are materially embodied metaneurons because the decisive state variables are magnetic texture, phase-transition state, or ionic/filamentary device physics rather than abstract digital logic (Azam et al., 2018, Palin et al., 21 Apr 2026, Boer et al., 2024).

A related hardware boundary case is the BLAST synaptic transistor. That work is not a metaneuron in the strict sense because it implements a metaplastic synapse, not a neuron with intrinsic excitability dynamics. Still, its conductance-dependent update law demonstrates that state-dependent device physics can improve generalization beyond ideal linear synapses, making it a partial building block for future meta-adaptive neuron–synapse systems (Kireev et al., 2022).

5. Collective and systems-level metaneurons

A different tradition interprets the effective computational unit not as a single neuron or device but as a collective metastable population. In “Metastability for systems of interacting neurons,” a finite all-to-all excitatory network has only one true invariant measure, $61.64 \pm 1.03$9, yet can remain in an active state for exponentially long times in $90.50 \pm 0.10$0 before collapsing (Löcherbach et al., 2020). The key macroscopic observable is the total firing rate

$90.50 \pm 0.10$1

and under appropriate assumptions the last spiking time obeys

$90.50 \pm 0.10$2

The paper further shows that normalized exit times from an active neighborhood converge toward an exponential law. This makes the active population function like an emergent higher-level unit with its own metastable state, lifetime, and collapse statistics (Löcherbach et al., 2020).

A related field-theoretic study shows how such metastable macrostates can arise from nonlinear single-neuron properties. In “Metastability in networks of nonlinear stochastic integrate-and-fire neurons,” recurrent stochastic LIF neurons with voltage-dependent hazard $90.50 \pm 0.10$3 and hard reset exhibit mean-field multistability and finite-size switching between stable states (Paliwal et al., 2024). The homogeneous mean-field equation is

$90.50 \pm 0.10$4

For threshold power-law hazards, superlinear exponents $90.50 \pm 0.10$5 can generate bistability between two active firing-rate states; for exponential hazards, two active states can also coexist though no truly silent state exists (Paliwal et al., 2024). This supports a metaneuron interpretation in which a local population, not a single membrane compartment, is the relevant stateful computational object.

The most expansive systems-level formulation appears in “Neurons as hierarchies of quantum reference frames,” which treats neurons as nested hierarchies of information-processing subsystems rather than point-like summation units (Fields et al., 2022). In that framework, synapses, dendritic branches, somata, and local circuits are modeled as quantum reference frames acting as hierarchical active-inference systems. The authors define classifiers, infomorphisms, and cone–cocone diagrams to represent compositional inference, and they argue that dendritic remodeling, trophic reward, and local signal interpretation should be understood as parts of a hierarchical measurement/action architecture (Fields et al., 2022). This is a speculative extension, but it clearly aligns with a metaneuron conception in which the neuron is itself a multi-level computational society.

A broader conceptual precursor is “Matter & Mind Matter,” which argues that future computing architectures should move away from point-neuron abstractions toward artificial spatio-temporal networks shaped by local dynamics, plasticity, growth, homeostasis, and topology (Birkoben et al., 2022). The paper does not define metaneurons, but it repeatedly suggests that useful computational units may be small structured dynamical subsystems rather than isolated neurons. A plausible implication is that metaneurons, in this systems-level sense, are local recurrent motifs whose state is distributed across morphology and time.

6. Open problems, misconceptions, and unresolved questions

The first misconception is that metaneurons designate a single well-defined neuron model. The surveyed literature does not support that reading. In one paper the term means a finite transferable library of neuron dynamics; in another it means a higher-level SNN unit without a fully specified dynamical equation; in others it is an interpretive label for biochemical routing, meta-homeostatic control, or material dynamics (Cheng et al., 2020, Rudnicka et al., 24 Aug 2025). Any rigorous use of the term therefore has to specify the level—algorithmic, biophysical, hardware, or collective—at which the “meta” property resides.

The second misconception is that every metaneuron proposal is equally formalized. Several are not. The SNN metaneuron comparison paper explicitly omits a standalone governing dynamical system for the metaneuron used in experiments (Rudnicka et al., 24 Aug 2025). Nikolić’s MR/GPGIC theory is bold architecturally but does not provide a dynamical systems formalism, a symbolic state calculus, or learning rules for how knowledge is encoded in receptor/channel distributions (Nikolić, 2022). MANA’s meta-homeostatic rule is fully specified, but it is explicitly phenomenological rather than a direct mechanistic model of a known biological process (Tosi et al., 2017). Metaboplasticity ties energy state to excitability and plasticity through temperature as a proxy rather than explicit ATP production, ion-pump energetics, or mitochondrial dynamics (Öner et al., 25 Dec 2025). The concept is therefore ahead of its canonical mathematics in much of the field.

Hardware work introduces a different set of open problems. The neuromodulable CMOS neuron demonstrates principled mode switching, but the prototype shares biases across neurons and some power figures are simulation-based rather than direct chip measurements (Mendolia et al., 30 Nov 2025). The MIT inhibitory neuristor currently operates at cryogenic temperatures and requires external inductance, with no compact array-level integration yet demonstrated (Palin et al., 21 Apr 2026). The halide perovskite neuron shows on-chip stochastic LIF behavior, but endurance, environmental stability, and scaling of the capacitor footprint remain practical questions (Boer et al., 2024). The leaky memcapacitor and skyrmion neurons are conceptually strong but remain primarily theoretical or simulation-based, with fabrication variability and reliability still unresolved (Zhang et al., 2023, Azam et al., 2018).

A final unresolved issue concerns the proper granularity of the computational unit. Some papers place the meta-layer inside the neuron, as an adaptive threshold, target rate, or hidden device state. Others place it above the neuron, as a metastable population or hierarchical active-inference structure. This suggests that metaneurons may be less a single ontology than a research program: a systematic attempt to identify which variables above ordinary spike transmission are necessary to explain adaptive computation, robust routing, continual learning, and efficient physical realization. The literature so far supports that program strongly, but it does not yet converge on a single canonical metaneuron.