Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Modelling Membrane Potentials by Diffusion Leaky Integrate-and-Fire Models Patrick Jahn DEPARTMENT OF MATHEMATICAL SCIENCES UNIVERSITY OF COPENHAGEN jahn@math.ku.dk CIRM, Marseille - January 18, 2009 Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Contents 1 Diffusion Leaky Integrate-and-Fire Neuronal Models 2 Modelling Membrane Potentials in Motoneurons by time-inhomogeneous Diffusion Models Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Biological Background Figure: The Neuron Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models −30 −40 −50 potential [mV] −20 Membrane Potential recorded by W. Kilb, Mainz 0 1 2 3 4 5 6 time [s] Figure: This membrane potential was recorded in vitro from a pyramidal neuron belonging to cortical slice preparation of a juvenile C57bl/6 mouse. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Diffusion Leaky Integrate-and-Fire Neuronal Model S | {z } | {z } T1 ... T2 x0 | {z } Tn Assume the process X between spikes is a Diffusion process given by dXt = 1 (a − Xt )dt + σ(Xt )dBt τ Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Diffusion Leaky Integrate-and-Fire Neuronal Model S | {z } | {z } T1 ... T2 x0 | {z } Tn Assume the process X between spikes is a Diffusion process given by dXt = 1 (a − Xt )dt + σ(Xt )dBt τ Goal: Estimate β(·) := τ1 (a − ·), σ(·), x0 , S from discrete observations of X and from the observation of iid ISIs (level crossing times) Ti , i = 1, . . . , n.. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Diffusion Leaky Integrate-and-Fire Neuronal Model S | {z } | {z } T1 ... T2 x0 | {z } Tn Assume the process X between spikes is a Diffusion process given by dXt = 1 (a − Xt )dt + σ(Xt )dBt τ Goal: Estimate β(·) := τ1 (a − ·), σ(·), x0 , S from discrete observations of X and from the observation of iid ISIs (level crossing times) Ti , i = 1, . . . , n.. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models −48 −49 −51 −50 S ? x0 ? −52 potential [mV] −47 −46 Problem of Finding Excitation Threshold and Reset Value 1.0 1.5 2.0 2.5 3.0 time [s] Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models 3.5 Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Problem of Finding Excitation Threshold and Reset Value −50 −40 −30 [mV] −20 −10 0 17Sept08_023.asc alle spikes uebereinanderlegen / spikezeiten auf 0 transformieren −0.04 −0.02 0.00 0.02 17Sept08_023.asc Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models 0.04 Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Strategy θ2 | {z T1 } | {z } ... T2 θ1 | {z } Tn 1. Fixing drift and diffusion coefficient with nonparametric methods proposed by R. Höpfner (2006). So assume X is given by 1 dXt = (a − Xt )dt + σ(Xt )dBt τ Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Strategy θ2 | {z T1 } | {z } ... T2 θ1 | {z } Tn 1. Fixing drift and diffusion coefficient with nonparametric methods proposed by R. Höpfner (2006). So assume X is given by 1 dXt = (a − Xt )dt + σ(Xt )dBt τ 2. Estimate θ1 = x0 and θ2 = S from the observation of iid inter spike times n o (θ ) Ti := inf t ≥ 0 | Xt 1 = θ2 , i = 1, . . . , n. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Strategy θ2 | {z T1 } | {z } ... T2 θ1 | {z } Tn 1. Fixing drift and diffusion coefficient with nonparametric methods proposed by R. Höpfner (2006). So assume X is given by 1 dXt = (a − Xt )dt + σ(Xt )dBt τ 2. Estimate θ1 = x0 and θ2 = S from the observation of iid inter spike times n o (θ ) Ti := inf t ≥ 0 | Xt 1 = θ2 , i = 1, . . . , n. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Some Facts. . . In the most famous cases, where X is an Ornstein-Uhlenbeck (OU) or a Cox-Ingersoll-Ross (CIR) process, no explicit expression for the density of the level crossing time T is known. However, we know its Laplace transfrom (LT) as a ratio of special functions. (Jahn, PhD-Thesis 2009): For the OU and CIR cases the estimation problem of θ1 = x0 and θ2 = S was solved by using Millar’s framework (1984) for minimum distance estimators via the comparison of empirical and theoretical LT with respect to a suitable Hilbert space norm . . . . Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Some Facts. . . In the most famous cases, where X is an Ornstein-Uhlenbeck (OU) or a Cox-Ingersoll-Ross (CIR) process, no explicit expression for the density of the level crossing time T is known. However, we know its Laplace transfrom (LT) as a ratio of special functions. (Jahn, PhD-Thesis 2009): For the OU and CIR cases the estimation problem of θ1 = x0 and θ2 = S was solved by using Millar’s framework (1984) for minimum distance estimators via the comparison of empirical and theoretical LT with respect to a suitable Hilbert space norm . . . . Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models The Minimum Distance Estimator w.r.t. the LT Theorem (J. 2009) Let X be an OU or a CIR process. Define the empirical LT of the level crossing time T of X from θ1 to θ2 and the corresponding family of possibly true LTs by n L̂n (α) := 1 X −αTi e n and Lθ (α) := Eθ [e −αT ], i=1 then the MDE w.r.t. the LT θn∗ := arg inf kL̂n − Lθ kH θ1 <θ2 is strongly consistent and asymptotically normal. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models The Pearson Diffusion Case (work with Jesper L. Pedersen. . . ) Pearson Diffusion dXt = τ1 (a − Xt )dt + σ p (Xt − c)2 + ddBt , X0 = θ1 ≥ 0 Jesper computed the Laplace transform of T : Eθ [e −αT ] = g Re (θ1 , α) − K (α) · hRe (θ1 , α) g Re (θ2 , α) − K (α) · hRe (θ2 , α) where g , h and K are expressions of hypergemetric functions where the parameters of the diffusion X enter. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models The Pearson Diffusion Case (work with Jesper L. Pedersen. . . ) Pearson Diffusion dXt = τ1 (a − Xt )dt + σ p (Xt − c)2 + ddBt , X0 = θ1 ≥ 0 Jesper computed the Laplace transform of T : Eθ [e −αT ] = g Re (θ1 , α) − K (α) · hRe (θ1 , α) g Re (θ2 , α) − K (α) · hRe (θ2 , α) where g , h and K are expressions of hypergemetric functions where the parameters of the diffusion X enter. Goal: Solve the identifiability problem and apply the developed Framework to this case... Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models The Pearson Diffusion Case (work with Jesper L. Pedersen. . . ) Pearson Diffusion dXt = τ1 (a − Xt )dt + σ p (Xt − c)2 + ddBt , X0 = θ1 ≥ 0 Jesper computed the Laplace transform of T : Eθ [e −αT ] = g Re (θ1 , α) − K (α) · hRe (θ1 , α) g Re (θ2 , α) − K (α) · hRe (θ2 , α) where g , h and K are expressions of hypergemetric functions where the parameters of the diffusion X enter. Goal: Solve the identifiability problem and apply the developed Framework to this case... Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Contents 1 Diffusion Leaky Integrate-and-Fire Neuronal Models 2 Modelling Membrane Potentials in Motoneurons by time-inhomogeneous Diffusion Models Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Joint work with R. Berg, J. Hounsgaard and S. Ditlevsen −70 −80 −90 −110 V(t) (mV) −60 −50 trace = 32 0 5000 10000 15000 20000 25000 30000 time (ms) Figure: Data collected by Rune Berg form a motoneuron of an active network in the spinal cord of a turtle, during mechanical stimulation. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models A First Approach Assumption: X is a Diffusion process with time inhomogeneous drift, dXt = β(Xt , t)dt + σ(Xt )dBt . Analysis: Nonparametric estimation of drift and diffusion coefficient. For the drift analysis: Cutting the trajectory into ”homogeneous” parts and analyse them separately. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models A First Approach Assumption: X is a Diffusion process with time inhomogeneous drift, dXt = β(Xt , t)dt + σ(Xt )dBt . Analysis: Nonparametric estimation of drift and diffusion coefficient. For the drift analysis: Cutting the trajectory into ”homogeneous” parts and analyse them separately. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Nonparametric estimators Consider the SDE dXt = β(Xt )dt + σ(Xt )dBt then the nonparametric estimators [Höpfner (2006), Florens-Zmirou (1993)] are given by Pi1 −M Xi∆ −x X(i+M)∆ −Xi∆ i=i0 K h ∆M b β(x) := βb(∆,M,h) (x) = Pi1 −M Xi∆ −x i=i0 K h Pi1 −M Xi∆ −x X(i+M)∆ −Xi∆ 2 √ i=i0 K h ∆M c2 (x) := σ c2 σ (x) = (∆,M,h) Pi1 −M Xi∆ −x i=i0 K h We use for K a rectangular and a triangular kernel. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Nonparametric estimators Consider the SDE dXt = β(Xt )dt + σ(Xt )dBt then the nonparametric estimators [Höpfner (2006), Florens-Zmirou (1993)] are given by Pi1 −M Xi∆ −x X(i+M)∆ −Xi∆ i=i0 K h ∆M b β(x) := βb(∆,M,h) (x) = Pi1 −M Xi∆ −x i=i0 K h Pi1 −M Xi∆ −x X(i+M)∆ −Xi∆ 2 √ i=i0 K h ∆M c2 (x) := σ c2 σ (x) = (∆,M,h) Pi1 −M Xi∆ −x i=i0 K h We use for K a rectangular and a triangular kernel. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Estimation of the Diffusion coefficient sigma^2: 05414032.acm_V.txt , LT: 500 , h: 0.1 , M: 20 500 500 1000 1500 sigma^2 1500 1000 sigma^2 2000 2000 2500 2500 sigma^2: 05414013.acm_V.txt , LT: 500 , h: 0.1 , M: 20 −85 −80 −75 −70 −65 −60 −55 −50 potential −90 −80 −70 −60 potential Figure: Nonparametric estimation of the Diffusion coefficient σ(·) works, since the estimator only considers the squared increments. Hence the effect of the inhomogeneous drift (finite variation) is negligible. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Analysis of the Drift: Cutting the trajectory into ”homogeneous” parts. . . −90 −80 −70 −60 −50 −110 potential 05414032.acm_V.txt 10 15 20 time Figure: Separate analysis of the drift in the quiescent period, On- and Off-cycles. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Analysis of the Drift: Cutting the trajectory into ”homogeneous” parts. . . −70 −80 −90 potential −60 −50 05414032.acm_V.txt 12.5 13.0 13.5 14.0 14.5 time Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models 15.0 Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Results: data set τ ON-cycle τ OFF-cycle Q-τ σ 2 (x) 02.acmV 03.acmV 11.acmV 13.acmV 14.acmV 17.acmV 20.acmV 26.acmV 27.acmV 28.acmV 31.acmV 32.acmV 35.acmV 36.acmV 38.acmV 41.acmV 7.498 10.461 8.074 7.025 5.648 8.767 8.391 9.318 8.751 8.731 9.869 8.077 9.977 7.866 8.928 10.088 13.465 11.813 13.59 16.496 10.443 14.513 14.339 9.356 11.282 11.242 9.042 9.679 9.179 13.528 9.123 14.102 19.966 31.783 27.388 33.156 22.203 25.954 38.494 17.27 33.192 23.89 31.393 26.122 28.975 16.238 17.581 22.892 45.59 · ( x - -88.662 ) 39.542 · ( x - -93.113 ) 25.426 · ( x - -116.197 ) 54.204 · ( x - -87.967 ) 57.916 · ( x - -81.793 ) 34.146 · ( x - -114.114 ) 40.984 · ( x - -101.679 ) 48.239 · ( x - -89.122 ) 47.846 · ( x - -99.089 ) 65.637 · ( x - -83.247 ) 38.66 · ( x - -101.546 ) 44.244 · ( x - -103.347 ) 23.315 · ( x - -135.819 ) 73.516 · ( x - -75.379 ) 69.325 · ( x - -75.265 ) 43.987 · ( x - -93.613 ) Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models An Extended LIF-model dXt = p 1 (a + g (t) − Xt )dt + σ Xt − c dBt τ (Xt ) τ (Xt ) = τ ∗ e −γ(Xt −a) √ σ(·) = σ · − c was already estimated. a is the mean of the quiescent period. τ ∗ and γ are found by regressing log(τ ) on (Xt − a). Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models An Extended LIF-model dXt = p 1 (a + g (t) − Xt )dt + σ Xt − c dBt τ (Xt ) τ (Xt ) = τ ∗ e −γ(Xt −a) √ σ(·) = σ · − c was already estimated. a is the mean of the quiescent period. τ ∗ and γ are found by regressing log(τ ) on (Xt − a). Goal: Estimate g (t) at the observation time points t1 , . . . , tn . Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models An Extended LIF-model dXt = p 1 (a + g (t) − Xt )dt + σ Xt − c dBt τ (Xt ) τ (Xt ) = τ ∗ e −γ(Xt −a) √ σ(·) = σ · − c was already estimated. a is the mean of the quiescent period. τ ∗ and γ are found by regressing log(τ ) on (Xt − a). Goal: Estimate g (t) at the observation time points t1 , . . . , tn . Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Fitting the extended Model If we assume that between observations g (t) and τ (Xt ) are ”constant”, where ∆ = ti − ti−1 is small then we can write Xi ≈ E[Xi |Xi−1 ] ≈ Xi−1 e −∆/τ (Xi−1 ) +(g (ti−1 )+a)(1−e −∆/τ (Xi−1 ) ) Hence we derive an estimator ĝ (ti−1 ) = Xi − Xi−1 e −∆/τi−1 −a (1 − e −∆/τi−1 ) Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Fitting the extended Model If we assume that between observations g (t) and τ (Xt ) are ”constant”, where ∆ = ti − ti−1 is small then we can write Xi ≈ E[Xi |Xi−1 ] ≈ Xi−1 e −∆/τ (Xi−1 ) +(g (ti−1 )+a)(1−e −∆/τ (Xi−1 ) ) Hence we derive an estimator ĝ (ti−1 ) = Xi − Xi−1 e −∆/τi−1 −a (1 − e −∆/τi−1 ) Finally we smooth ĝ (·) by a smooth spline g s (·). Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Fitting the extended Model If we assume that between observations g (t) and τ (Xt ) are ”constant”, where ∆ = ti − ti−1 is small then we can write Xi ≈ E[Xi |Xi−1 ] ≈ Xi−1 e −∆/τ (Xi−1 ) +(g (ti−1 )+a)(1−e −∆/τ (Xi−1 ) ) Hence we derive an estimator ĝ (ti−1 ) = Xi − Xi−1 e −∆/τi−1 −a (1 − e −∆/τi−1 ) Finally we smooth ĝ (·) by a smooth spline g s (·). Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Comparing Network Activity and Estimated Input 1.0 Network Activity 0 5000 10000 15000 20000 time 10 20 The cyan line is the squared and smoothed network activity from above. It was rescaled to be compared to our estimate g s (·) given by the black line. 0 input (mV) 30 −1.0 40 −0.5 0.0 50 0.5 trace = 32 0 5000 10000 15000 20000 25000 30000 35000 time (ms) Patrick Jahn - jahn@math.ku.dk We can think of the network activity to be a threshold version of the Input! Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Comparing Network Activity and Estimated Input 1.0 Network Activity 0 5000 10000 15000 20000 time 10 20 The cyan line is the squared and smoothed network activity from above. It was rescaled to be compared to our estimate g s (·) given by the black line. 0 input (mV) 30 −1.0 40 −0.5 0.0 50 0.5 trace = 32 0 5000 10000 15000 20000 25000 30000 35000 time (ms) Patrick Jahn - jahn@math.ku.dk We can think of the network activity to be a threshold version of the Input! Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Simulated Data using the estimated quantities compared to real data: −80 V(t) (mV) −70 −60 a,γ,τ= −99.9 , 0.029 , 21.1 −110 −100 −90 −80 −90 −100 −110 V(t) (mV) −70 −60 −50 trace = 32 −50 trace = 32 0 5000 10000 15000 20000 25000 30000 time (ms) Patrick Jahn - jahn@math.ku.dk 0 5000 10000 15000 time (ms) Leaky Integrate-and-Fire Models 20000 25000 30000 Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models A Spiking Dataset V(t) (mV) −60 −50 −40 a,γ,τ= −89 , 0.04 , 31.6 −100 −90 −80 −70 −60 −70 −80 −90 −100 V(t) (mV) −50 −40 −30 trace = 13 −30 trace = 13 0 5000 10000 15000 20000 25000 30000 time (ms) Patrick Jahn - jahn@math.ku.dk 0 5000 10000 15000 time (ms) Leaky Integrate-and-Fire Models 20000 25000 30000 Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models R.Berg, S.Ditlevsen, J.Hounsgaard: “Intense Synaptic Activity Enhances Temporal Resolution in Spinal Motoneurons”. PLoS ONE (2008). R.Höpfner: “On a set of data for the membrane potential in a neuron”. Math. Biosci. 207 (2007). P.Millar: “A General Approach to the Optimality of Minimum Distance Estimators”. Trans. Amer. Math. Soc., Vol. 286, No. 1. (1984). P.Jahn, PhD. Thesis: “Statistical Problems Related to Excitation Threshold and Reset Value of Membrane Potentials”. http://archimed.uni-mainz.de (2009). P.Jahn: “Estimation of Excitation Threshold and Reset Value in Diffusion Leaky Integrate-and-Fire Neuronal Models”. submitted in J. Math. Biol. Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models Diffusion Leaky Integrate-and-Fire Neuronal Models Modelling by time-inhomogeneous Diffusion Models Thank you for your attention! Questions? Patrick Jahn - jahn@math.ku.dk Leaky Integrate-and-Fire Models