The spiking model ladder, priced in FLOPs, for people who close the loop

What the work claims

This is a survey, and a modestly scoped one: Parepko, Shulepin, and Nasybullin divide neuron models into single-entity (single-compartment) and composite (multi-compartment) classes, and review only the former. Their claim is organizational, not experimental: that the scattered literature on single-compartment models can be arranged into a clean ladder, from one-equation integrate-and-fire (IF) models up through the four-equation Hodgkin-Huxley (HH) model and its simplified descendants, with each rung's strengths, limitations, and use cases stated explicitly, plus a final section treating trained artificial neural networks as neuron surrogates.¹

The one piece of quantitative substance is their Table I, which compares models on discrete or continuous dynamics, number of variables, number of parameters, number of FLOPs per step, whether the model carries experimental evidence, and the dynamical regimes it can reproduce. The authors are careful to state their counting convention: every primary mathematical operation counts as one floating-point operation, and the external input current is excluded to keep realizations comparable. Weight your reading accordingly: this is a map, not a verdict, and the terrain it maps is textbook material.¹

How it works

The ladder, as the survey lays it out. At the bottom sits the integrate-and-fire family: the passive membrane is a single capacitor-resistor equation, C dV/dt = I, and a spike is a bookkeeping event, when V crosses a threshold the state resets. Add a leak conductance and you get the leaky IF; add a second equation that raises the threshold after each spike and you get an adaptive IF; make the equation quadratic and the model becomes bistable, able to switch its activity type with the input. The survey's own verdict is blunt: these models are, in their phrase, biophysically meaningless, but that is exactly why they are cheap and why they map effortlessly onto electrical circuits in neuromorphic processors.¹

At the top sits Hodgkin-Huxley: the membrane equation plus three gating variables (m, n, h) whose voltage-dependent kinetics were determined empirically from the squid axon. It is the reference standard for biophysical fidelity and, the authors note, computationally demanding, with activation and inactivation treated as kinetically independent, a known simplification. Between the poles they place the reductions: the FitzHugh-Nagumo model, a two-dimensional phenomenological rewrite that trades channel detail for phase-plane insight but whose parameters the authors flag as hard to calibrate; the Izhikevich model, two equations with a quadratic voltage term and a reset, whose five parameters were fitted to cortical dynamics and which the survey positions as the workhorse for large-scale simulation, noting it can be discretized at 1 millisecond resolution and run on modest hardware; the dendritic neuron model, a four-layer feedforward construction (synaptic, dendritic, membrane, cell) whose multiplicative dendritic branches capture interactions a weighted sum cannot; and finally trained ANNs, where a CNN-LSTM with a small custom spiking layer can imitate a modeled neuron's activity after training, at the price of behaving unpredictably outside the training distribution.¹

The table's headline numbers, under the authors' counting convention: IF, 1 variable, 3 parameters, 1 FLOP per step; LIF, 1 variable, 4 parameters, 5 FLOPs; Izhikevich, 2 variables, 5 parameters, 13 FLOPs; Hodgkin-Huxley, 4 variables, 11 parameters, about 1200 FLOPs. That is a factor of roughly ninety between the cheapest credible spiking model and the reference biophysical one, per neuron, per step.¹

Where a skeptic should push

The most load-bearing assumption in any such table is that FLOP counts mean something portable. They do not, very much: the survey itself warns the numbers vary with realization, a fair warning given that no two simulators integrate the HH equations identically, and a 1 millisecond discretization of a model with sodium kinetics faster than a millisecond is a claim that deserves scrutiny, not a footnote. Treat the table as an order-of-magnitude ladder (tens versus hundreds versus thousands of operations), which is robust, not as a benchmark.

Second, the survey's selection criteria are never made rigorous. Which models count as prevalent is asserted, not demonstrated; the meta-dynamic neuron material it cites is peripheral; and the ANN section reports simulation preferences without any protocol an instrumentation engineer could reproduce. Third, what the survey does not analyze is as telling as what it does: there is no identifiability discussion (which parameters can be recovered from which measurements), no fitting criteria, no absolute timing, and no treatment of numerical stiffness, which is where HH models actually cost you in practice. As a guide to choosing a model it is a starting bibliography, not a decision procedure.¹

What this means for the MEA closed-loop twin

The non-obvious point: a microelectrode array system already contains a neuron model, and it is at the bottom of this ladder. Every threshold-crossing spike detector with a refractory period is an integrate-and-fire neuron wearing a firmware costume. When the same system runs closed loop, predicting tissue state to time its stimuli, it needs a second model, a digital twin of the recorded tissue, and that twin must run faster than real time on hardware that sits next to the amplifier. The survey's table is the budget line for that requirement: at one millisecond steps, an Izhikevich-class twin costs about thirteen floating-point operations per neuron per step, which is trivial for hundreds of channels on any modern edge node, while an HH-class twin at roughly 1200 operations is where the fan-out starts to bite, before counting the stiffness-driven sub-stepping a stiff HH system imposes.

The opportunity, then, is to choose the ladder rung consciously rather than by habit. For state prediction and stimulus timing, an adaptive-IF or Izhikevich twin captures threshold fatigue and burst structure, the features that matter for loop stability, at a cost that leaves the edge node's budget for spike sorting. Reserve HH-class twins for offline characterization, where they earn their cost. The survey's evidence column quietly supports this division: the IF and LIF rows carry no experimental-evidence checkmark, while Izhikevich and HH do, so the cheapest defensible rung for a loop twin is not the bottom one.¹

The threat has two faces. One is identifiability, which the survey never raises: HH and Izhikevich parameters in the literature are fitted to intracellular voltage traces, and an extracellular array delivers spike times and waveforms, not membrane voltage. A twin whose parameters you cannot measure from your own instrument is a hypothesis dressed as a state estimate; the honest calibration target for MEA data is IF-class dynamics with adaptation, fitted from inter-spike statistics. The other threat is the survey's own section five: a trained ANN surrogate is the fastest twin of all, and the authors note it cannot handle out-of-domain data. Living tissue leaves the training distribution as a matter of course, drugs, maturation, drift. A twin that silently extrapolates is worse than a crude model that knows it is crude.

The bottom line

Established by this survey: the single-compartment model ladder and its relative costs, stated under an explicit counting convention, with honest limitations attached to each rung. Not established, because the survey attempts nothing of the kind: which model you should choose for a given workload. For MEA instrumentation the actionable reading is that the compute ladder is real (tens versus hundreds versus thousands of operations per neuron per step), that the loop twin should sit at the cheapest rung carrying an experimental-evidence checkmark, and that identifiability from extracellular data, not fidelity, is the binding constraint. What would confirm the framework is a comparison that fits each model class to the same MEA recordings and scores prediction quality per FLOP; what would break it is evidence that waveform-level features let you recover higher-rung parameters extracellularly, which would move the whole ladder up one rung for free.