15 Fast Transients and Rate Models

15.4 Summary

The population activity of spiking neuron models responds to a big and rapid change in the input current much faster than the input potential. The response of the input potential is characterized by the membrane time constant τm\tau_{m} and therefore exhibits the properties of a low-pass filter. In an asynchronously firing population of neurons, however, there are always a few neurons with membrane potential just below the threshold. These neurons respond quasi-instantaneously to a step in the input current, despite the fact that the input potential, i.e., the contribution to the membrane potential that is caused by the input, responds slowly.

The details of the response depend on the neuron model as well as on the amplitude of the signal and the type of noise. With slow noise as the dominant noise source, model neurons respond quickly and reliably to a step input. For white noise, the picture is more complicated.

For Spike Response Model neurons with escape noise, the speed of the response depends on the noise level. While the response is fast for low noise, it is as slow as the membrane potential in the limit of high noise.

For a large amount of diffusive white noise and a small amplitude of the input signal, the choice of neuron model plays an important role. Leaky integrate-and-fire models respond fairly slow, but faster than the input potential. The response of exponential integrate-and-fire models follows that of the membrane potential, but the effective membrane time constant depends on the population activity.

The fact that spiking neuron models in the high-noise limit respond slowly can be used to derive rate models for the population activity. Such rate models are the basis for the analysis of cognitive dynamics in Part IV of the book. Nevertheless, it should be kept in mind that standard rate models miss the rapid transients that a population of spiking models exhibits in response to signals that are strong compared to the level of noise.


The rapid transients in spiking models without noise or with slow noise have been reported by several researchers, probably first by Knight (264) and have later been rediscovered several times (183; 77; 351).

The analysis of transients in the escape-rate has been performed in Gerstner (183) where also the limits of high noise and low noise are discussed. For the linearization of the membrane potential density equations and analysis of transient behavior in neuron models with diffusive noise see Brunel and Hakim (78); Lindner and Schimansky-Geier (298); Fourcaud and Brunel (155); Richardson (434). Experimental data on transients in the linearized regime can be found in Silberberg et al. (477); Tchumatchenko et al. (509).

Simoncelli et al. (478) gives an authoritative summary of LNP models. How generalized integrate-and-fire models can be mapped to LNP models has been discussed in Aviel and Gerstner (33) and Ostojic and Brunel (373). An excellent overview of the central concepts of rate models with adaptation can be found in Benda and Herz (48).


  1. 1.

    Population of noise-free neurons

    (i) Show that for noise-free neurons the population activity equation yields

    A(t)=11+T(t^)A(t^),A(t)={1\over 1+T^{\prime}(\hat{t})}\,A(\hat{t})\,, (15.28)

    where T(t^)T(\hat{t}) is the inter-spike interval of a neuron that has fired its last spike at time t^\hat{t} , and the prime denotes the derivative.

    Hints: In the limit of no noise, the input-dependent interval distribution PI(t|t^)P_{I}(t\,|\,\hat{t}) reduces to a Dirac δ\delta -function, i.e.,

    PI(t|t^)=δ[t-t^-T(t^)].P_{I}(t\,|\,\hat{t})=\delta[t-\hat{t}-T(\hat{t})]\,. (15.29)

    where T(t^)T(\hat{t}) is given implicitly by the threshold condition

    T(t^)=min{(t-t^)| u(t)=ϑ;u˙>0, t>t^}.T(\hat{t})={\rm min}\{(t-\hat{t})\,|\,u(t)=\vartheta;\dot{u}>0,\,t>\hat{t}\}\,. (15.30)

    Recall from the rules for δ\delta -functions that

    abδ[f(x)]g(x)   𝑑   x=g(x0)|f(x0)|,\int_{a}^{b}\delta[f(x)]\,g(x)\,{\text{d}}x={g(x_{0})\over|f^{\prime}(x_{0})|}\,, (15.31)

    if ff has a single zero-crossing f(x0)=0f(x_{0})=0 in the interval a<x0<ba<x_{0}<b with f(x0)0f^{\prime}(x_{0})\neq 0 .

    (ii) Assume SRM 0{}_{0} neurons with u(t)=η(t-t^)+h(t)u(t)=\eta(t-\hat{t})+h(t) . Show that

    A(t)=h(t)ηT)A(t^),A(t)={h^{\prime}(t)\over\eta^{\prime}T)}\,A(\hat{t})\,, (15.32)

    Hint: use the results from (i).

    (iii) An input current of amplitue I1I_{1} is switched on at time t=0t=0 . Assume an input potential h(t)=h0h(t)=h_{0} for t<0t<0 and h(t)=(R/τ)0texp(-s/τ)I1h(t)=(R/\tau)\int_{0}^{t}\exp(-s/\tau)I_{1} for t>0t>0 . Show that the transient of the population activity after the step at t=0t=0 is instantaneous, despite the fact the input potential responds slowly.

  2. 2.

    LNP and frequency-current curve . Around Eq. ( 15.22 ), it was argued that a model

    A(t)=F^(0GI(s)I(t-s)   𝑑   s)A(t)=\hat{F}(\int_{0}^{\infty}G_{I}(s)I(t-s){\text{d}}s)\, (15.33)

    with a choice F^(x)=gσ(x/[0GI(s)   𝑑   s])\hat{F}(x)=g_{\sigma}(x/[\int_{0}^{\infty}G_{I}(s){\text{d}}s]) is optimal. The aim is to make the notion of optimality more precise.

    (i) Show that for constant, but arbitrary, input I0I_{0} , Eq. ( 15.33 ) leads to A0=gσ(I0)A_{0}=g_{\sigma}(I_{0}) , consistent with the general results of Chapter 12 .

    (ii) Suppose that GI(s)G_{I}(s) is the linearization of the population activity equations around A0A_{0} which is achieved for a constant input I0I_{0} . Show that linearization of Eq. ( 15.33 ) leads to ΔA(t)=0G(s)ΔI(t-s)   𝑑   s\Delta A(t)=\int_{0}^{\infty}G(s)\,\Delta I(t-s)\,{\text{d}}s .

    Hint: Recall that the response at zero frequency G^(0)=0G(s)   𝑑   s\hat{G}(0)=\int_{0}^{\infty}G(s){\text{d}}s is related to the slope of the gain function.

    (iii) Interpret the results from (i) and (ii) and explain the range of validity of the model defined in Eq. ( 15.33 ). What can happen if the input varies about a mean I1I0I_{1}\neq I_{0} ? What happens if the variations around I0I_{0} are big?

  3. 3.

    Leaky integrate-and-fire with white diffusive noise According to the results given in Tab.  15.2 , the linear filter GI(s)G_{I}(s) of leaky integrate-and-fire neurons has a high-frequency behavior G~I(ω)=A0Rσ1ωτm\tilde{G}_{I}(\omega)=A_{0}{R\over\sigma}{1\over\sqrt{\omega\tau_{m}}} .

    (i) Calculate the response to a step current input.

    Hint: Use ΔA(t)=0G(s)ΔI(t-s)   𝑑   s\Delta A(t)=\int_{0}^{\infty}G(s)\,\Delta I(t-s)\,{\text{d}}s . Insert the step current, take the Fourier transform, perform the multiplication in frequency space, and finish with the inverse Fourier transform.

    (ii) Compare with the simulation results in Fig. 15.9 B.

  4. 4.

    Rate model for a population of exponential integrate-and-fire with white diffusive noise

    The aim is to derive the effective time constant given in Eq. ( 15.24 ) which characterizes a population of exponential integrate-and-fire neurons.

    (i) Write A(t)=F[h(t)]A(t)=F[h(t)] . Linearize about a reference value A0=F(h0)A_{0}=F(h_{0}) and prove that    𝑑   A/   𝑑   t=F   𝑑   h/   𝑑   t{\text{d}}A/{\text{d}}t=F^{\prime}\,{\text{d}}h/{\text{d}}t .

    (ii) Assume that Eq. ( 15.23 ) holds with the unknown time constant τeff\tau_{\rm eff} . Assume periodic stimulation I(t)=I0+ΔIexp(iωt)I(t)=I_{0}+\Delta I\exp(i\omega t) with a high frequency ω\omega . This will lead to a periodic perturbation ΔAexp[i(ωt+ϕ)]\Delta A\,\exp[i(\omega t+\phi)] . Find the ratio c(ω)=ΔA/ΔIc(\omega)=\Delta A/\Delta I .

    (iii) Match the high frequency behavior of c(ω)c(\omega) to G~I(ω)\tilde{G}_{I}(\omega) so as to find the time constant τeff\tau_{\rm eff} .

    Hint: Recall from Tab.  15.2 that the linear filter GI(s)G_{I}(s) of the exponential integrate-and-fire neurons has a high-frequency behavior G~I(ω)=A0RΔT1ωτm\tilde{G}_{I}(\omega)=A_{0}{R\over\Delta_{T}}{1\over\omega\tau_{m}} .

  5. 5.

    Equivalence of rate models . We use the rate model defined in Eqs. ( 15.1 ) and ( 15.3 ) with R=1R=1 in order to describe coupled populations

    τm   𝑑   hi   𝑑   t=-hi+Ii+kwikF(hk).\tau_{m}{{\text{d}}h_{i}\over{\text{d}}t}=-h_{i}+I_{i}+\sum_{k}w_{ik}F(h_{k})\,. (15.34)

    Compare this model to another rate model

    τm   𝑑   Ai   𝑑   t=-Ai+F(kwikAk+I^i).\tau_{m}{{\text{d}}A_{i}\over{\text{d}}t}=-A_{i}+F(\sum_{k}w_{ik}A_{k}+\hat{I}% _{i})\,. (15.35)

    Show that Eq. ( 15.35 ) implies Eq. ( 15.34 ) under the assumption that I=I^+τmdI^/dtI=\hat{I}+\tau_{m}d\hat{I}/dt .

    Hint: Set hi=kwikAk+Ih_{i}=\sum_{k}w_{ik}A_{k}+I and take the derivative ( 345 ) .