betaFeedback

WorksEngineering

Excitable Behaviours of a First-Order High-Pass Neuron Model

Abstract

With the rise of neuromorphic technologies, simple and tunable neuron models are needed to analyse and design spiking systems. However, the simplest neuron models, e.g., Leaky Integrate-and-Fire (LIF) models, usually fail to capture the excitable behaviours of biological neurons. This could be a significant limitation, as excitability enables reliable spike timing under parametric uncertainties. Recently, a minimal excitable model, the so-called High-Pass Model (HPM), was introduced in neuroengineering. It simply consists of a static sigmoid with a high-pass filter in positive feedback. Given its novelty, it is still unclear how easily the HPM can be tuned, and how much of the excitable behaviours of a biophysical neuron it can reproduce. To address these questions, we investigate the tunability of a first-order HPM sub- and supra-rheobase, i.e., below and above the onset of the tonic spiking regime (where the neuron spikes periodically), and compare its behaviours with those of a simple Hodgkin-Huxley-type Conductance-Based Model (CBM). Below rheobase, both models fire spikes in response to transient inputs, including post-inhibitory rebound spikes. Above rheobase, an affine reparametrization of the input of the HPM suffices to fit its frequency-current curve to that of a CBM. The two models mostly differ near their respective rheobases, where the first-order HPM lacks bistability and spike latency. We suggest that modifying the frequency-selective element of the HPM may be sufficient to overcome this limitation. These results highlight the potential of the HPM as a fundamental computational primitive in neuromorphic, spike-based technologies.

Block diagram of the first-order high-pass neuron model

Figure 1. The first-order high-pass neuron model.

Introduction

Neurons are excitable cells: they emit spikes with reproducible timing in response to transient inputs [1, 2]. Since Hodgkin and Huxley [3], Conductance-Based Models (CBMs) have been the reference description of the ionic mechanisms behind excitability. However, dimensionality and nonlinearities of CBMs make their dynamics, and in particular their spiking threshold, challenging to analyse and exploit in the design of neuromorphic, spike-based technologies.

While this complexity has motivated reduced models of CBMs that retain the essence of their biophysics [4], the models mostly used in neuromorphic designs are from the integrate-and-fire family [5], since they are very easy to analyse. In particular, a common choice is the linear Leaky Integrate-and-Fire (LIF) neuron. However, a LIF neuron acts as a low-pass system, and thus is non-excitable: it spikes exactly when its input crosses a static threshold, so its response depends only on the input amplitude, not on how rapidly the input changes. Excitable neurons, in contrast, have a dynamic spiking threshold that depends on both the amplitude and the rate of change of the input [6, 7]. Although extensions such as the adaptive exponential model recover some of these excitable behaviours [8, 9], this is at the price of additional states, parameters and nonlinearities, reintroducing some of the complexity that LIF models seek to avoid.

Excitability is, however, at the heart of the spike timing robustness of neurons. Under the same fluctuating input, perturbed copies of excitable models such as CBMs spike more reliably than non-excitable models (Figure 2). While the LIF also shows some amount of reliability, its spike timings are also less consistent, and, crucially, its response is qualitatively different from that of a CBM. This suggests that neuromorphic technologies could benefit from easily tunable and theoretically tractable spiking primitives that retain neuronal excitability, particularly for bio-inspired spike-based coding in the presence of hardware mismatch.

Recently, a minimal excitable model, the so-called High-Pass Model (HPM), was introduced in neuroengineering [10]. It consists of a static high-gain sigmoid with a high-pass filter in positive feedback. According to the mixed-feedback principle, excitability in biophysical neurons results from the interplay of a fast positive and a slow negative feedback [11, 12]. The first-order HPM might be the simplest realization of this principle available: the feedthrough branch of the high-pass acts as a fast positive feedback, while the low-pass branch acts as a slow negative feedback [10] (Figure 1). Like other excitable models, the HPM can generate spikes in response to sub-rheobase transient inputs and produce post-inhibitory rebound spikes (where a spike is produced at the release of an inhibitory input). It also exhibits robust spike timing (Figure 2). In short, the HPM is a neuron model that appears to reproduce the excitable behaviours of a CBM while retaining minimal model complexity.

This raises two practical questions: how expressive and easily tunable is the HPM? In particular, can it be tuned to reproduce the behaviours of a simple CBM? The answers will clarify which neuronal behaviours the HPM can easily support for neuromorphic applications, and which properties of a CBM can be captured and analysed with a HPM. We address these questions using a simple Hodgkin-Huxley-type CBM as a reference, and we deliberately keep the HPM as simple as possible: the high-pass filter is fixed to first-order. Our contributions are the following.

  1. Below the rheobase, we show that the HPM reproduces the main excitable behaviours of the CBM, and that its parameters have a clear and intuitive influence on its input sensitivity.

  1. Above the rheobase, we show that the frequency-to-input-current (ff-II) curve of the CBM can be matched by the HPM using an analytical least-squares fit and an affine parametrization of its input.

  1. We show that the remaining differences between the HPM and the simple CBM concern specific mechanisms of the CBM near its rheobase: the CBM possesses a noticeable bistability region [13, 6], and it can significantly delay its spike initiation after a slow excitation. In contrast, the first-order HPM is monostable and generally fires immediately after excitation.

This work comes with three accompanying papers. In [14], the excitability mechanism of the HPM is introduced, together with an explicit dynamical threshold separating the resting and excited states. In [15], an analytical expression for the tonic spiking frequency of the HPM is obtained. Finally, in [16], both parabolic and elliptic bursting are reproduced with minimal modifications to the HPM. Collectively, these results establish a set of design principles for the use of the HPM as a neuromorphic computational primitive.

The remainder of the paper is organized as follows. Section II introduces the HPM and CBM models. Section III compares their responses to transient inputs and how the HPM can be tuned. Section IV compares their tonic spiking regimes under constant inputs and fits the HPM ff-II curve. Finally, Section V discusses the limitations of first-order HPMs and future directions. All code and data are available on GitHub: https://github.com/JeromeAdriaens/EB-FO-HPM-ECC27.

Hodgkin-Huxley and high-pass models

This section introduces the two models compared in this paper: a Hodgkin-Huxley CBM, used as a simple but representative biological neuron model, and a first-order HPM. While the CBM is rooted in the biophysical, circuit-level implementation of excitability, the HPM is derived from teleological, i.e., function-driven, principles, retaining only the minimal input-output elements underlying excitability. In both models, tonic spiking emerges when their input crosses the rheobase, a point at which their resting equilibrium bifurcates into a limit cycle. By contrast, a neuron is excitable if a transient input, possibly below the rheobase, can still trigger a spike. Both models follow a mixed-feedback structure and are excitable.

A Hodgkin-Huxley conductance-based model

We consider a Hodgkin-Huxley CBM with a usual minimal set of ionic currents for action potential generation [3]: a leak current IlI_l, a transient sodium current INaI_{\mathrm{Na}}, and a delayed-rectifier potassium current IK,dI_{\mathrm{K,d}}. With an applied current IextI_{\mathrm{ext}}, the membrane potential obeys Kirchhoff’s current balance:

CV˙=−∑Iion+Iext,C\dot{V}=-\sum I_{\mathrm{ion}}+I_{\mathrm{ext}},

i.e.

CV˙=−Gˉl(V−El)−GˉNamNa(V,t)hNa(V−ENa)−GˉK,dmK,d(V,t)(V−EK)+Iext(t)(1)C\dot{V}=-\bar{G}_{l}(V-E_l)-\bar{G}_{\mathrm{Na}}m_{\mathrm{Na}}(V,t)h_{\mathrm{Na}}(V-E_{\mathrm{Na}}) -\bar{G}_{\mathrm{K,d}}m_{\mathrm{K,d}}(V,t)(V-E_{\mathrm{K}})+I_{\mathrm{ext}}(t) \tag*{(1)}

with GNa=GˉNamNahNaG_{\mathrm{Na}}=\bar{G}_{\mathrm{Na}}m_{\mathrm{Na}}h_{\mathrm{Na}} and GK,d=GˉK,dmK,dG_{\mathrm{K,d}}=\bar{G}_{\mathrm{K,d}}m_{\mathrm{K,d}} the dynamic conductances. Each gating variable X∈{mNa,hNa,mK,d}X\in\{m_{\mathrm{Na}},h_{\mathrm{Na}},m_{\mathrm{K,d}}\} obeys

τX(V)X˙=X∞(V)−X(2)\tau_X(V)\dot{X}=X_{\infty}(V)-X \tag*{(2)}

where X∞(V;a,b)=(1+exp⁡V+ab)−1X_{\infty}(V;a,b)=\left(1+\exp\frac{V+a}{b}\right)^{-1} is a Boltzmann function and τX(V;a,b,d,e)=a−bd⋅(1+exp⁡V+de)−1\tau_X(V;a,b,d,e)=a-\frac{b}{d}\cdot\left(1+\exp\frac{V+d}{e}\right)^{-1} a shifted Boltzmann function. Contrary to the original Hodgkin-Huxley model [3], where the gating variables are raised to higher powers to fit the kinetics of the ion channels precisely, here they enter with exponent one, since we only care about the qualitative structure of the model. In this model, sodium activation mNam_{\mathrm{Na}} provides a fast positive feedback on VV, whereas potassium activation mK,dm_{\mathrm{K,d}} and sodium inactivation hNah_{\mathrm{Na}} provide slow negative feedbacks, so the system is excitable and fits the mixed-feedback principle [11, 12].

The first-order high-pass neuron model

The HPM was introduced in [10]. It consists of a static high-gain sigmoid σ\sigma with a first-order high-pass filter H(s)H(s) in positive feedback. At first-order, H(s):=τs/(τs+1)H(s) := \tau s/(\tau s + 1) (Figure 1). In state-space form, the model reads

y=σ(u+y−x),(3)y = \sigma(u + y - x), \tag*{(3)}
τx˙=−x+y,(4)\tau\dot{x} = -x + y, \tag*{(4)}

with uu the input signal, yy the output signal, xx the internal state of high-pass. We moreover define v:=u+y−xv := u + y - x as the sigmoid input. Since yy appears on both sides of y=σ(u+y−x)y = \sigma(u + y - x), the motif is acausal. To simulate it, we use the causal regularization εy˙=−y+σ(u+y−x)\varepsilon\dot{y} = -y + \sigma(u + y - x). The state-space equations (3) are recovered at the singular limit ε→0+\varepsilon\to0^{+}, corresponding to the classical singularly-perturbed limit of neurodynamics.

The first-order high-pass combines an instantaneous positive feedback with a slow negative feedback, which fits the mixed-feedback principle. The model can also be understood as a lead-excitation motif: the high-pass filter computes the trend of the output (lead), which is then amplified by positive feedback (excitation). Collapsing the feedthrough branch of the high-pass filter with the sigmoid yields a bistable block, so the system can also be read as a hysteresis-and-adaptation motif, much like the FitzHugh-Nagumo neuron model [14].

The tunability of the first-order HPM, beyond the high-pass timescale τ\tau, lies entirely in the choice of the high-gain sigmoid σ\sigma, which we parametrize as follows

σ(v;k,p,α)=α⋅v+(k−α)⋅v(1+∣k⋅v∣p)1/p.(5)\sigma(v; k, p, \alpha) = \alpha\cdot v + (k - \alpha) \cdot\frac{v}{\left(1 + \lvert k \cdot v\rvert^{p}\right)^{1/p}}. \tag*{(5)}

The effects of the three parameters (k>1k > 1, p>0p > 0 and α≥0\alpha\ge0) on the sigmoid are illustrated in Figure 3. The slope at the origin is σ′(0)=k\sigma'(0) = k, so the corners of σ\sigma lie near (±1/k,±1)(\pm1/k, \pm1). The exponent pp sets the sharpness of these corners, so σ′\sigma' becomes discontinuous at ±1/k\pm1/k as p→∞p \to\infty. Finally, α\alpha is the slope of σ\sigma as v→±∞v \to\pm\infty. The rheobase of the model lies at the point where σ′=1\sigma' = 1 [15, 14, 16]. Consequently, α\alpha, pp

Parametrized sigmoid function

Figure 3. Parametrized sigmoid function. Arrows show the effect of increasing kk (red), pp (blue), and α\alpha (green).

and kk set the input sensitivity at rest, near the rheobase, and during tonic spiking, respectively. Setting (p,α)=(∞,0)(p, \alpha) = (\infty, 0) yields a non-excitable HPM neuron, since only inputs above the rheobase affect the neuron.

The rest of the paper studies the input-output behaviours of the first-order HPM (3), compares them to those of the CBM (1), and examines how the parameters τ\tau, kk, pp, and α\alpha can shape them.

Responses to transient inputs

Excitability is best defined as an all-or-none sensitivity to transient inputs. This transient sensitivity underlies spikes triggered below the rheobase, including rebound spikes at the release of an inhibitory input, as well as the reliability of spike timing in the presence of parametric mismatch (Figure 2). In this section, we show that the CBM and the HPM respond similarly to transient inputs, despite showing some differences near the rheobase. We then investigate how the HPM parameters can tune its sub-rheobase sensitivity.

Raster plots and frequency-to-input-current curves for CBM, HPM, LIF, and non-excitable HPM models

Figure 2. Parametric robustness in spike timing of excitable models. Raster of the spike times of 50 copies of each model (one row per copy), whose parameters are perturbed by a 5% relative standard deviation noise, driven by the same fluctuating current (top left). The initial models are chosen to share a similar tonic spiking regime (top right). The CBM and the HPM fire at nearly the same times across copies (left). The non-excitable HPM and the LIF do not (right). All code and data are available on GitHub.

Sensitivity of neuron models to transient inputs

Inspired by [2], we apply a transient wave input to four neuron models: the CBM (1), an excitable HPM (3), a non-excitable HPM, and a regular LIF. The transient wave input is defined for t≥0t \ge0 as

w(t)=w0+wpktτwexp⁡(1−tτw),(6)w(t) = w_0 + w_{\mathrm{pk}} \frac{t}{\tau_w} \exp\left(1 - \frac{t}{\tau_w}\right), \tag*{(6)}

where τw\tau_w is the timescale of the wave and wpkw_{\mathrm{pk}} is its peak amplitude (reached at t=τwt = \tau_w) with respect to the baseline w0w_0. The baselines are set individually for each model below its rheobase so that the neurons are initially at rest. Two protocols are performed, one with excitatory waves (wpk>0w_{\mathrm{pk}} > 0) and one with inhibitory (hyperpolarizing) waves (wpk<0w_{\mathrm{pk}} < 0). For each protocol, the number of spikes produced by each model in response to a single wave is reported in Figure 4 and Figure 5 as a function of the amplitude ∣wpk∣\lvert w_{\mathrm{pk}}\rvert and the timescale τw\tau_w, both on logarithmic axes. Transposing the xx and yy axes, the figures give sensitivity curves as a function of the input’s timescale range (the inverse of frequency), reminiscent of amplitude Bode plots.

Rebound spikes triggered by inhibitory waves

Figure 4. Rebound spikes triggered by inhibitory waves (5), log-log scale. Non-excitable HPM and LIF do not exhibit rebound excitability.

Transient sensitivity of the models to excitatory waves

Figure 5. Transient sensitivity of the models to excitatory waves (5), log-log scale. Grey levels indicate the number of spikes each model produces in response.

1) Response to inhibitory waves: Let us first examine the results of the inhibitory protocol in Figure 4, as they are the simplest to interpret. The non-excitable models (LIF and non-excitable HPM) never produce rebound spikes, whereas the excitable ones (HPM and CBM) do, so only the latter are shown. The regions of the (wpk,τw)(w_{\mathrm{pk}}, \tau_w)-plane in which rebound spikes occur are qualitatively similar for the HPM and the

CBM: rebounds are triggered after an inhibition of sufficient amplitude, and within a restricted range of timescales. This timescale range reflects the frequency selectivity to transient inputs of both models. On one hand, the mixed feedback of the CBM behaves like a band-pass filter with a large bandwidth, since its positive feedback (the sodium activation mNam_{\mathrm{Na}}) possesses a fast, but not negligible, timescale. On the other hand, the mixed feedback of the HPM is a high-pass filter, so its sensitivity to small timescales is larger.

  1. Response to excitatory waves: Let us now examine the results of the excitatory protocol in Figure 5. A first striking observation is that the figures become progressively simpler, going from the CBM to the HPM, the non-excitable HPM, and finally the LIF. To structure the discussion, we divide the (wpk,τw)(w_{\mathrm{pk}},\tau_w)-plane of the CBM into several qualitatively distinct regions: R1 to R4, interpret each of them, and examine whether they appear in the other models.

R1: Sub-rheobase excitability: This region corresponds to excitatory waves that trigger a spike, despite their peak amplitude remaining below the rheobase (vertical lines in the figure). Only the excitable models (the CBM and the HPM) exhibit this behaviour. Similarly to rebound spikes, the region appears on a restricted range of timescales, reflecting the frequency selectivity of both models to transient inputs. In models where this region exists, spike timing is reliable (Figure 2). This region is probably well-suited for robust spike-based coding.

R2: Tonic spiking: This region corresponds to inputs for which the models enter an oscillatory regime. Its interpretation is straightforward: the longer the input remains above the rheobase, the more spikes are produced. All models exhibit a version of this region. Interestingly, the purely low-pass filtering mechanism of the LIF delays significantly the first spike trigger, which explains the absence of spiking over a large region of R2 for small timescales. A similar effect can be found in the other models, but it is confined to much smaller timescales. This region is well-suited for rate-based coding.

R3: Rest-spike bistability: Because of accommodation, depending on the protocol, two rheobases can be found for the CBM: the lower one is identified using step inputs, and the upper one using a very slowly rising ramp [13], [6], [7]. We bound R1 by the lower rheobase, since spikes beyond it can lead to tonic spiking, and R2 by the upper rheobase, since this is where the resting state truly loses stability. R3 corresponds to a hysteresis effect which is absent from the other models [9], [17]. Neurons exhibiting this region can better retain memory of past inputs. We conjecture that a HPM with a second-order filter would recover this region.

R4: Spike latency: For slowly varying inputs, the CBM driven slightly above its upper rheobase may still take a substantial time before producing its first spike. This explains a gap in R2, which we call R4, where the neuron is on its way to tonic spiking, but the input wave returns it to rest before the first spike is produced. We attribute this delay to the timescale of the positive feedback (sodium activation mNam_{\mathrm{Na}}). Neglecting it, as in the HPM, removes the delay between excitation and spike production. Neurons exhibiting this region display a dynamic phase response to transient inputs. We conjecture that a HPM with a band-pass filter of wide bandwidth would recover this region.

Tuning the sub-rheobase sensitivity of a HPM

Figure 6 shows how the boundary of the R1 spiking region in the (wpk,τw)(w_{\mathrm{pk}},\tau_w)-plane changes when each parameter of the HPM is varied while the others are held fixed. The four parameters play distinct roles:

Parameter influence on the sub-rheobase response of the HPM

Figure 6. Parameter influence on the sub-rheobase response of the HPM. Boundary of the spiking region in the (wpk,τw)(w_{\mathrm{pk}},\tau_w)-plane. Spikes occur to the right of each curve. Each panel increases one parameter; the others are fixed.

  • τ\tau shifts the spiking boundary along the timescale axis, leaving both the amplitude range of R1 and the rheobase unchanged. This is expected, since τ\tau sets the frequency selectivity of the HPM.

  • α\alpha changes the amplitude range of R1, leaving the timescale sensitivity unchanged. The rheobase stays nearly constant at large pp, but can shift at smaller pp. This is consistent with α\alpha tuning the loop gain of the HPM in its excitable sub-rheobase regime.

  • pp changes both the rheobase and the amplitude range of R1, since the rheobase lies at the corner of the sigmoid while the sub-rheobase sensitivity is set by the gain below it, thus only leaving the timescale sensitivity

unchanged. For α=0\alpha= 0, the excitability of the HPM relies entirely on the corner of the sigmoid not being too sharp, i.e., on pp remaining small.

  • kk, unlike the other parameters, has a non-monotonic effect on the spiking boundary: as kk increases, the rheobase first increases and then decreases, while R1 flattens. This effect is harder to interpret, as this gain influences the sharpness of the sigmoid corner in a way not captured by pp. Since kk indirectly affects the entire spiking boundary, we recommend keeping it fixed.

In short, τ\tau sets the timescale at which the HPM responds, α\alpha and pp mainly set how far below the rheobase it responds, and pp and kk set the rheobase itself. More generally, we expect the R1 spiking boundary of HPMs to be shifted along the amplitude axis by the sigmoid parameters, which set the amplitude sensitivity, and along the timescale axis by the frequency-selector parameters, which set the timescale sensitivity. This is an important advantage, as it allows HPMs to be tuned intuitively by acting on each sensitivity separately.

Responses to constant inputs

As discussed previously, for constant inputs, both the CBM and the HPM fire tonically above their rheobase. In this section, we aim to fit the ff-II curve of the HPM to that of a given CBM, such as the one illustrated in Figure 7. Fitting reduced neuron models to detailed ones has a long history [18], [8], [4], but our goal here is narrower: we only aim to evaluate how easily the tonic spiking regime of a HPM can be made to coincide with that of a CBM. A first difficulty is that the ff-II curves of the two models differ qualitatively in shape. First, as mentioned in Section III-A.2, the CBM has two rheobases, which we now denote I‾rheo\underline{I}_{\mathrm{rheo}} and

f-I curve for the CBM

Figure 7. ff-II curve for the CBM. The inset zooms on the neighbourhood of the rheobases, indicated by dashed vertical lines: the lower rheobase I‾rheo\underline{I}_{\mathrm{rheo}} (red) and the upper rheobase I‾rheo\overline{I}_{\mathrm{rheo}} (teal).

I‾rheo\overline{I}_{\mathrm{rheo}}, between which it is bistable. However, the first-order HPM is monostable and has a single rheobase. Moreover, the ff-II curve of the CBM increases first concavely and then almost linearly, whereas that of the HPM is typically more bell-shaped [15].

To overcome this difficulty, we assume that the bistable window of the CBM is sufficiently narrow for a single rheobase in the HPM to be a reasonable approximation, and we only use the first half of the HPM ff-II curve, where the frequency ff increases with uu. This latter point is achieved through an affine parametrization of the input, such that

u=φ(Iext):=1ρ(Iext−I0),ρ>0,(7)u = \varphi(I_{\mathrm{ext})} := \frac{1}{\rho}(I_{\mathrm{ext}} - I_0), \qquad\rho> 0, \tag*{(7)}

where I0I_0 is the current mapped to u=0u = 0, the maximum spiking frequency of the HPM. The fit is performed on [Ir,I0][I_r,I_0], where IrI_r is the rheobase to match. The rheobase urheo(k,p,α)u_{\mathrm{rheo}}(k,p,\alpha) of the HPM [15] fixes the scaling factor ρ\rho: requiring φ(Ir)=−urheo\varphi(I_r) = -u_{\mathrm{rheo}} imposes

ρ=I0−Irurheo(k,p,α),(8)\rho= \frac{I_0-I_r}{u_{\mathrm{rheo}}(k,p,\alpha)}, \tag*{(8)}

so ρ\rho is not a free parameter. Given that the CBM has two rheobases and the HPM one, we take the midpoint Ir=12(I‾rheo+I‾rheo)I_r = \frac{1}{2}\left(\underline{I}_{\mathrm{rheo}}+\overline{I}_{\mathrm{rheo}}\right). The remaining parameters θ=(k,p,τ,α,I0)\theta= (k,p,\tau,\alpha,I_0) are obtained by solving the nonlinear least-squares problem

min⁡θ∑i=1N[fCBM(Ii)−fHPM(φ(Ii);θ)]2,Ii∈[Ir,I0],(9)\min_{\theta} \sum_{i=1}^{N}\left[f_{\mathrm{CBM}}(I_i)-f_{\mathrm{HPM}}(\varphi(I_i);\theta)\right]^2,\qquad I_i \in[I_r,I_0], \tag*{(9)}

where fCBMf_{\mathrm{CBM}} is the frequency measured on the tonic-spiking branch and fHPMf_{\mathrm{HPM}} the analytical frequency of the HPM [15]. We enforce k>1k > 1 and p,α>0p,\alpha> 0 so the HPM remains excitable. The expression of fHPMf_{\mathrm{HPM}} contains a root-finding step, which is differentiable through the implicit function theorem [19]. This gives the Jacobian of the residuals, and we solve (8) with the Levenberg-Marquardt algorithm [20], [21]. We recall that the implementation details are available in the GitHub repository mentioned earlier.

Figure 8 shows the result for N=11N = 11 equally spaced currents. The HPM reproduces the onset frequency and the concave rise with a root-mean-square error of 0.290.29 Hz, and below 0.50.5 Hz at every point. This suggests that a simple affine remapping of the HPM input is sufficient to match its tonic spiking regime to that of a CBM.

Fitted $f$-$I$ curve

Figure 8. Fitted ff-II curve. HPM (orange) fitted to the frequencies of the CBM (dots) on [Ir,I0][I_r, I_0] with an affine input map. Vertical lines: IrheoI_{\mathrm{rheo}}, I‾rheo\overline{I}_{\mathrm{rheo}} and their mean IrI_r, imposed on the HPM. Fitted parameters: k=1.13k = 1.13, p=5.61p = 5.61, α=0.55\alpha= 0.55, τ=31.5\tau= 31.5ms, I0=221.4μA/cm2I_0 = 221.4\mu\mathrm{A}/\mathrm{cm}^{2}, ρ=303.5\rho= 303.5.

Conclusions and perspectives

In this work, we investigated how expressive and easily tunable the first-order HPM is, using a simple Hodgkin-Huxley-type CBM as a reference. Below the rheobase, we showed that the HPM reproduces the main excitable behaviours of the CBM. Subjected to transient wave inputs, both models exhibit sub-rheobase excitability within a restricted range of timescales, thus making their spike timing reliable. These behaviours are absent from non-excitable models such as LIF neurons. Moreover, the parameters of the HPM have a relatively clear and decoupled influence on its input sensitivity, which makes the model particularly intuitive to tune. Above the rheobase, we showed that, despite some qualitative differences, the ff-II curve of the CBM can be matched by that of an HPM using a simple affine parametrization of the HPM input. The remaining differences between the two models concern mechanisms of the CBM located near its rheobase. The CBM exhibits a bistability region, and it can delay its first spike after a slow excitation. In contrast, the first-order HPM is monostable and generally fires immediately after excitation. We conjecture that these behaviours can be recovered within the HPM framework, with a second-order high-pass filter for bistability and a band-pass filter of wide bandwidth for spike latency.

Overall, these results indicate that even in its simplest form, the HPM captures the essential excitable and tonic spiking behaviours of a conductance-based neuron, while remaining analytically tractable and intuitive to tune, making it a promising candidate for neuromorphic applications.

Significant work remains to facilitate the theoretical design, numerical simulation, and hardware implementation of the HPM, particularly for frequency selectors more complex than the first-order high-pass filter investigated here.

References

  1. [1]Z. F. Mainen and T. J. Sejnowski, “Reliability of spike timing in neocortical neurons,” Science, vol. 268, no. 5216, pp. 1503–1506, 1995.pubmed.ncbi.nlm.nih.gov/7770778
  2. [2]R. Sepulchre, G. Drion, and A. Franci, Excitable Behaviors, pp. 269–280. Cham: Springer International Publishing, 2018.
  3. [3]A. L. Hodgkin and A. F. Huxley, “A quantitative description of membrane current and its application to conduction and excitation in nerve,” The Journal of Physiology, vol. 117, no. 4, pp. 500–544, 1952.
  4. [4]T. Van Pottelbergh, G. Drion, and R. Sepulchre, “From biophysical to integrate-and-fire modeling,” vol. 33, no. 3, pp. 563–589.
  5. [5]A. N. Burkitt, “A review of the integrate-and-fire neuron model: I. homogeneous synaptic input,” Biological Cybernetics, vol. 95, no. 1, pp. 1–19, 2006.pubmed.ncbi.nlm.nih.gov/16622699
  6. [6]R. Azouz and C. M. Gray, “Dynamic spike threshold reveals a mechanism for synaptic coincidence detection in cortical neurons in vivo,” Proceedings of the National Academy of Sciences, vol. 97, no. 14, pp. 8110–8115, 2000.pubmed.ncbi.nlm.nih.gov/10859358
  7. [7]J. Platkiewicz and R. Brette, “A threshold equation for action potential initiation,” PLoS Computational Biology, vol. 6, no. 7, p. e1000850, 2010.DOI
  8. [8]R. Brette and W. Gerstner, “Adaptive exponential integrate-and-fire model as an effective description of neuronal activity,” Journal of Neurophysiology, vol. 94, no. 5, pp. 3637–3642, 2005.
  9. [9]E. M. Izhikevich, “Neuronal excitability, spiking and bursting,” International Journal of Bifurcation and Chaos, vol. 10, no. 6, pp. 1171–1266, 2000.
  10. [10]G. Bainier, C. Fernandez Lorden, A. Fyon, J. Adriaens, L. Mendolia, G. Drion, P. Sacré, and A. Franci, “A two-block model of neuronal excitability.” Poster presented at Neuromorphic Computing Netherlands 2026, Maastricht, The Netherlands, Oct. 2026.DOI
  11. [11]G. Drion, A. Franci, J. Dethier, and R. Sepulchre, “Dynamic input conductances shape neuronal spiking,” eNeuro, vol. 2, no. 1, 2015.DOI
  12. [12]R. Sepulchre, G. Drion, and A. Franci, “Control across scales by positive and negative feedback,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 2, pp. 89–113, 2019.DOI
  13. [13]A. V. Hill, “Excitation and accommodation in nerve,” Proceedings of the Royal Society of London. B. Biological Sciences, vol. 119, no. 814, pp. 305–355, 1936.DOI
  14. [14]G. Bainier, A. Fyon, C. Fernandez Lorden, and A. Franci, “Excitability of high-frequency positive feedback,” 2026. arXiv:2611.06822.arxiv.org/abs/2611.06822
  15. [15]C. Fernandez Lorden, J. Adriaens, A. Franci, and G. Bainier, “Characterization of the frequency-current curve of first-order high-pass neurons,” 2026. arXiv:2611.09508.arxiv.org/abs/2611.09508
  16. [16]A. Fyon, J. Brandolt, L. Mendolia, P. Sacré, G. Drion, A. Franci, C. Fernandez Lorden, and G. Bainier, “Types of bursting with a two-block spiking primitive,” 2026. arXiv:2611.08045.arxiv.org/abs/2611.08045
  17. [17]Y. Lu, X. Xin, and J. Rinzel, “Bistability at the onset of neuronal oscillations,” Biological Cybernetics, vol. 117, pp. 61–79, 2023.DOI
  18. [18]R. Jolivet, T. J. Lewis, and W. Gerstner, “Generalized integrate-and-fire models of neuronal activity approximate spike trains of a detailed model to a high degree of accuracy,” Journal of Neurophysiology, vol. 92, no. 2, pp. 959–976, 2004.pubmed.ncbi.nlm.nih.gov/15277599
  19. [19]M. Blondel, O. Mensch, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-López, F. Pedregosa, and J.-P. Vert, “Efficient and modular implicit differentiation,” Advances in Neural Information Processing Systems (NeurIPS), 2022. arXiv:2105.15183.arxiv.org/abs/2105.15183
  20. [20]K. Levenberg, “A method for the solution of certain non-linear problems in least squares,” Quarterly of Applied Mathematics, vol. 2, no. 2, pp. 164–168, 1944.DOI
  21. [21]D. W. Marquardt, “An algorithm for least-squares estimation of non-linear parameters,” Journal of the Society for Industrial and Applied Mathematics, vol. 11, no. 2, pp. 431–441, 1963.

Paper details

vlnc:W_anddyj4c
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents