Accepted Manuscript

Title: Linked Gauss-Diffusion processes for modeling a

finite-size neuronal network

Author: M.F. Carfora E. Pirozzi

PII: S0303-2647(17)30073-4
DOI: http://dx.doi.org/doi:10.1016/j.biosystems.2017.07.009
Reference: BIO 3760

To appear in: BioSystems

Received date: 3-3-2017

Revised date: 20-7-2017
Accepted date: 31-7-2017

Please cite this article as: M.F. Carfora, E. Pirozzi, Linked Gauss-Diffusion processes
for modeling a finite-size neuronal network, <![CDATA[BioSystems]]> (2017),
http://dx.doi.org/10.1016/j.biosystems.2017.07.009

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 Linked Gauss-Diffusion processes for modeling a
finite-size neuronal network

t
M.F. Carforaa , E. Pirozzib

ip
a Istituto
per le Applicazioni del Calcolo “Mauro Picone”, Consiglio Nazionale delle
Ricerche, via Pietro Castellino 111 - 80131 Napoli, Italia

cr
b Dipartimento di Matematica e Applicazioni, Università degli Studi di Napoli “Federico II”,

Complesso di Monte S. Angelo via Cintia - 80126 Napoli, Italia

us
Abstract
A Leaky Integrate-and-Fire (LIF) model with stochastic current-based linkages

an
is considered to describe the firing activity of neurons interacting in a (2 × 2)-
size feed-forward network. In the subthreshold regime and under the assumption
that no more than one spike is exchanged between coupled neurons, the stochas-
tic evolution of the neuronal membrane voltage is subject to random jumps due
M
to interactions in the network. Linked Gauss-Diffusion processes are proposed
to describe this dynamics and to provide estimates of the firing probability den-
sity of each neuron. To this end, an iterated integral equation-based approach is
applied to evaluate numerically the first passage time density of such processes
d

through the firing threshold. Asymptotic approximations of the firing densities

of surrounding neurons are used to obtain closed-form expressions for the mean
te

of the involved processes and to simplify the numerical procedure. An extension

of the model to an (N ×N )-size network is also given. Histograms of firing times
obtained by simulations of the LIF dynamics and numerical firings estimates are
compared.
p

Keywords: Stochastic differential equations, Synaptic current-based linkages,

ce

Simulation, First passage time

1. Introduction
Ac

In the last two decades, an increasing attention has been paid to the repre-
sentation of the dynamics of interacting neurons in small and large networks
(Arbib, 2003; Brunel, 2000; Politi and Luccioli, 2010; Sirovich et al., 2010)
through different stochastic models. Many of them are essentially based on
integrate-and-fire equations, population dynamics, mean field theory (Brunel
and Hansel, 2006; Buice and Chow, 2013; Forcaud and Brunel, 2002; Galves

Email addresses: f.carfora@iac.cnr.it (M.F. Carfora), enrica.pirozzi@unina.it (E.

Pirozzi)

Preprint submitted to Elsevier July 20, 2017

Page 1 of 19
 and Locherbach, 2016; Ostojic et al., 2009; Soula and Chow, 2007). In the com-
plex study of the proposed network models, several aspects are explored as, for
instance, size and heterogeneity of the network, proportions of excitatory and
inhibitory neurons, spiking synchronization, correlated noisy inputs and some

t
others features (see Lansky and Ditlevsen, 2008; Postnova et al., 2010; Brunel

ip
and Hansel, 2006, and references therein). Anyway, all these models start with
the description of the structure of the considered network and, consequently,
they specify how adjacent neural nodes interact.

cr
In this paper, we intend to investigate all the above network features by de-
scribing the firing activity of each neuronal component; specifically, we take into
account the geometry of the network, the kind of interaction between neurons

us
and the position of each neuron within the network by proposing a stochas-
tic Leaky Integrate-and-Fire (LIF) model for a finite size network. Such small
networks constitute a useful model for some specific biological systems: as de-
scribed for instance in Hansson (1995), the larval antenna of several Lepidoptera

an
comprises just a few olfactory receptor neurons (about ten). In the olfactory
system of adult moths the sensory input, after being detected and processed
by some thousands first-order neurons, is transmitted synaptically to a small
network made by a few tens of second-order neurons that code for this sensorial
M
information (Rospars et al., 2014). On the other hand, small network models
allow to move from the detailed representation of the dynamics of individual
neurons to the description of the activity of large populations of neurons: in-
deed, a single neuronal unit can be assumed as representative of the proportion
d
of excitatory or inhibitory neurons in large networks (Greenwood and Ward,
2016; Sirovich et al., 2010).
Here we describe the firing density of any observed neuron as the first pas-
te

sage time (FPT) density through a threshold of a Gauss-Diffusion (GD) process

with jumps that are random in both time and amplitude. The model is then rep-
resented by a system of LIF-type equations (see Buonocore et al., 2014, 2015b)
p

including both external stimuli and linkage synaptic currents, expressed through
indicator functions of the firing of others neurons. Two different strategies for
ce

the investigation on the firing activity are proposed: by means of trajectories

simulations and by the more advantageous numerical evaluation of an auxiliary
system of Volterra integral equations (Buonocore et al., 2015a; D’Onofrio and
Pirozzi, 2015). Moreover, under suitable hypotheses related to the low firing
Ac

rate (we will specify them as typical of an asymptotic regime), a closed-form

expression for the firing density will also be derived.
From a modeling point of view, the obtained firing densities are an inter-
esting result of the application of the theory of linked Gauss Markov processes
to the firing activity of a small feed-forward network. They represent the first
step towards the modeling of larger and more complex networks. From a bio-
logical point of view, the knowledge of the firing densities of both the internal
and boundary neurons contributes to investigate how the sensory information
is detected, processed and encoded (Rospars et al., 2014). Finally, the obtained
densities can be assumed as statistical hypotheses for inferential studies on the
dynamics of more complex networks and for data analysis and parameters esti-

Page 2 of 19
 mation techniques.
We start from the model proposed in Buonocore et al. (2013, 2014) and
Carfora and Pirozzi (2015) to describe the interaction between two neurons,
also subject to a time-dependent synaptic current conveying external stimuli;

t
in Section 2 we show how to extend it to a (2 × 2) network of linked neurons

ip
that can be suitably generalized to an (N × N ) network; we also outline the
main modeling choices and state the terminology. A (2 × 2) scheme of a small
neuronal network is considered in Section 3, with direction and weight of the

cr
coupling links between neurons to be specified according to the position of each
of them within the network. We describe the evolution of the membrane volt-
age of these four neurons by LIF-type stochastic differential equations (SDEs)

us
and focus on the corresponding FPTs, assumed as models for their firing times.
These firing densities will be evaluated by adopting suitable approximating GD
processes whose FPT densities can be estimated by numerical iterative proce-
dures. In Section 4, the numerical strategy is given along with an asymptotic

an
approximation of the firing densities through a closed-form expression for the
associated FPT densities, that holds under special hypotheses (low firing or
subthreshold regime). The numerical evaluation of these firing densities will
be compared with histograms of first crossing times through a fixed membrane
M
voltage threshold of the simulated paths obtained by an Euler discretization of
the LIF equations. The case of three neurons will also be considered as a special
case of the (2 × 2) scheme in which a specific neuron has been disconnected from
the others. Finally, in Section 5, by using asymptotic approximations for the
d
firing activity of all the surrounding neurons, we provide the mean function of
the GD process that approximates the membrane voltage of a generic neuron
within a finite size (N × N ) network with specified geometry and connections;
te

in Section 6 some concluding remarks are presented.

p

2. The model

We consider a small network of interacting neurons, each of them subject to a

ce

stochastic LIF-type dynamics with a time dependent stimulus I(t). We call such
a model a time inhomogeneous LIF model. An additional stochastic synaptic
current realizes the connection between neuronal units. As in the classical LIF
model, when the membrane potential V (t) of a generic neuron reaches a firing
Ac

threshold VS , it is immediately reset on the deterministic value v0 < VS ; at the

same time, its firing activates an additional current in the dynamics of adjacent
neurons. We are then interested in characterizing the probability density of the
firing of any neuron in the network by investigating the FPT probability density
function (pdf) of the stochastic process V (t). Specifically, in such a model the
evolution of the membrane voltage of the neuron, embedded in the network, can
be described by the following SDE for t ≥ 0
 
1 ρ + µθ
dV (t) = − V (t) + + I(t) dt + σ(t)dW (t) (1)
θ θ

Page 3 of 19
 with V (0) = v0 ; here, θ stands for the decay time of the membrane potential V (t)
in absence of noise, ρ stands for the resting potential, µ represents a constant
signal, σ > 0 is the intensity of noise and W (t) is a standard Brownian motion.
The stimulus I(t) (that is dimensionally the ratio between the input current

t
and the membrane capacity) is specified as

ip
h i
I(t) = i0 e−t/α + k 1 − e−t/α H(t), (2)

cr
where i0 is the initial value of the stimulus and α > 0 is its decay time. The
stochastic jump induced by the spikes of adjacent neurons has maximum am-
plitude k and it is modeled by a linear combination of Heaviside step functions:

us
(
X 0, if t < τj ,
H(t) = βj Hj (t) and Hj (t) := (3)
1, if t ≥ τj ,
where τj is the emission time of the first spike by Neuron j. Note also that k

an
measures both the intensity and the sign (positive forP excitatory/negative for
inhibitory) of the interaction. Finally, we also assume βj = 1. In Buonocore
et al. (2013, 2014) and Carfora and Pirozzi (2015) we developed such a model to
describe the firing activity of two neurons having one-way connection, mutual
M
connection and a connection driven by a periodic signal, respectively. Here the
sum over j in the stochastic function H(t) defined in (3) is extended to all the
neurons whose firing activity affects the dynamics of the observed one, since
for any j Hj (t) is the indicator function of the first firing time τj of Neuron j.
d
Furthermore, the coefficient βj in the first part of (3) represents the coupling
weight between the observed neuron, whose dynamics is described by (1), and
the Neuron j linked to it. These parameters (that add up to 1) are to be chosen
te

to modulate the relative weight of the interaction according to both the number
of surrounding neurons and their position with respect to the observed one.
Moreover, note that the jumps in the time evolution of V (t) induced by the
p

synaptic linkage function H(t) have random amplitude and occur in random
times. The final goal of this approach is to provide a prediction of the firing
ce

density of each neuron in the network by numerical evaluation of the FPT of

a GD process Ve (t) approximating V (t). It is worth noting that we focus on
the effect of single presynaptic spikes on the firing probability of a postsynaptic
neuron: then, only the first spike time of each one of the presynaptic neurons
Ac

is recorded and the effects of these spikes on the dynamics of a postsynaptic

neuron are considered just until its first spike, and the consequent reset of its
membrane potential.
Here, we limit ourselves to illustrate the model based on the SDE of type
(1) with stochastic linkages (2)-(3) specified on the basis of the dimension N
of the network and of the location of the embedded neuron; furthermore, we
provide a prediction of the firing densities, numerically evaluated, by adopting
a GD process approximating the process subject to the dynamics of SDE (1),
suitably specified for the observed neuron in the network. Indeed, the FPT pdf
of this GD process is obtained through the numerical quadrature of a related
integral equation.

Page 4 of 19
 3. A (2 × 2)-size network
Let us consider a small (2 × 2) network, made of four interconnected neurons
and graphically represented as a square matrix with connecting edges along

t
which information is transmitted (see Fig. 1). The related stochastic model

ip
involves a system of four SDEs of type (1), each one describing the evolution of
the membrane voltage of a neuron embedded in the network. In the following,
we will denote these neurons by the labels N11 , N12 , N21 , N22 , as in Fig. 1, and

cr
describe the related dynamics in detail; the linkage is realized by means of
the stochastic functions Hij (t), all built as linear combinations of the indicator
functions of the firing activity of the connected neurons. The next Subsections

us
will be specifically devoted to show results for these four neurons.

an
M
d

Figure 1: Schematic of a 4-neurons network.

te

3.1. N11 and N12 dynamics

p

Let us start by introducing the basic model made of just two interacting
neurons, denoted by N11 and N12 . The evolution of their membrane potentials
V11 (t) and V12 (t), respectively, under a common fixed potential threshold VS
ce

can be described by two coupled SDEs:

 1 ρ1j + µ1j θ1j 

dV1j (t) = − V1j (t) + + I1j (t) dt + σ1j (t)dW1j (t) (4)
Ac

θ1j θ1j
with h i
I1j (t) = i0,1j e−t/α1j + k1j 1 − e−t/α1j H1j (t), j = 1, 2. (5)
In the case of one-way interaction, N11 does not receive any input from N12
so that H11 (t) ≡ 0 ∀t > 0 and consequently the stimulus I11 (t) only represents
a contribution due to external biological or injected currents. On the contrary,
for N12 we have H12 (t) = β11 H11 (t), with β11 =1 and
(
0, if t < τ11 ,
H11 (t) := (6)
1, if t ≥ τ11 .

Page 5 of 19
 Moreover, k12 is a signed coefficient, whose sign takes into account the type
of interaction (i.e., excitatory or inhibitory) and whose modulus represents the
intrinsic response of the neuron N12 , i.e. the amplitude of the jump, induced
by the interaction with N11 , on the stimulus driving its dynamics. This model

t
has been studied in Buonocore et al. (2013) and Buonocore et al. (2014). The

ip
mean functions of the two processes V11 (t) and V12 (t) are given by
h i
m11 (t) ≡ E[V11 (t)] = v0,11 e−t/θ11 + (ρ11 + µ11 θ11 ) 1 − e−t/θ11

cr
e−t/α11 − e−t/θ11 (7)
+ i0,11
(1/θ11 − 1/α11 )

us
and
h i
m12 (t|T11 ) ≡ E[V12 (t)|T11 ] = v0,12 e−t/θ12 + (ρ12 + µ12 θ12 ) 1 − e−t/θ12

an
e−t/α12 − e−t/θ12
Z th i (8)
+ i0,12 + k12 e−t/θ12 1 − e−s/α12 es/θ12 H11 (s)ds.
(1/θ12 − 1/α12 ) 0

Note that the mean function m12 (t|T11 ) of the process V12 (t) is conditioned
M
upon the random variable T11 , that describes the FPT of the process V11 (t), by
which we model the spike time τ11 of N11 . Here, H11 (t) is the indicator function
of the event {T11 ≤ t}.
Hence, we consider the GD process Ve12 (t) approximating V12 (t) whose mean
is derived as
d

m
e 12 (t) = E [m12 (t|T11 )] = ET11 [E(V12 (t)|T11 )]
te

so that, by observing that E[H11 (t)] = P11 (t) := P rob(T11 ≤ t), we obtain from
(8)
p

h i
me 12 (t) ≡ E[Ve12 (t)] = v0,12 e−t/θ12 + (ρ12 + µ12 θ12 ) 1 − e−t/θ12
ce

Z th
e−t/α12 − e−t/θ12 i
+ i0,12 + k12 e−t/θ12 1 − e−s/α12 es/θ12 P11 (s)ds.
(1/θ12 − 1/α12 ) 0
(9)
Ac

The covariances of the processes V11 (t) and Ve12 (t) have the same form, given by

σ 2 h s/θ i
c(s, t) = θ e − e−s/θ e−t/θ , 0 ≤ s ≤ t; (10)
2
in particular, the corresponding variance is

σ2 h i
V ar(t) = c(t, t) = θ 1 − e−2t/θ , t ≥ 0. (11)
2

As well known, the FPT of such a GD process Ve12 (t) through the threshold VS
is an absolutely continuous random variable. Its pdf gVe12 [VS , t|v0,12 , 0] can be

Page 6 of 19
 obtained as a solution of the following integral equation (IE) with non-singular
kernel:
gVe12 [VS , t|v0,12 , 0] = − ψVe12 [VS , t|v0,12 , 0]

t
Z t (12)

ip
+ ψVe12 [VS , t|VS , ξ]gVe12 [VS , ξ|v0,12 , 0] dξ,
0

where, for τ < t,

cr
e 12 (t) [1 + e−2(t−τ )/θ12 ]

VS − m
ψVe12 [VS , t|y, τ ] = −me 012 (t) −
θ12 [1 − e−2(t−τ )/θ12 ]

us
e 12 (t) 2e−(t−τ )/θ12

y−m
+ fVe12 [Vs , t|y, τ ]
θ12 1 − e−2(t−τ )/θ12
and ( )

an
2
1 [Vs − M (t|y, τ )]
fVe12 [Vs , t|y, τ ] = p exp −
2
2πD (t|τ ) 2D2 (t|τ )

is the Gaussian transition pdf of the GD process Ve12 , whose conditional mean
and variance are given by
M
c(τ, t) c2 (τ, t)
M (t|y, τ ) = m
e 12 (t) + [y − m
e 12 (t)], D2 (t|τ ) = V ar(t) −
V ar(τ ) V ar(τ )
d
with c(·, ·) and V ar(·) are as in (10) and (11) for θ = θ12 , respectively. Then,
quadrature of IE (12) allows us to evaluate the FPT pdf gVe12 by a numerical
te

procedure (see Buonocore et al., 2014, for further details). However, it has to
be observed that, as clearly seen in the expressions for ψVe12 andfVe12 , to apply
the quadrature algorithm to the IE (12), the mean and covariance functions
of the involved GD process are required, along with the threshold VS . Now,
p

while the firing threshold VS and the covariance (10) of Ve12 (t) are known, to
specify the mean function m e 12 (t) of Ve12 (t) given by (9) we need an estimate of
ce

the distribution function P11 (s). To this end, we first apply the procedure to
the corresponding integral equation (12) written for the GD process V11 (t) (in
place of Ve12 (t)) to obtain the numerical evaluation of both gV11 [VS , t|v0,11 , 0] and
Ac

P11 (s) by quadrature of gV11 [VS , t|v0,11 , 0]; then, we can solve (12) for Ve12 (t).
Alternatively, under suitable hypotheses of an asymptotic regime (Buonocore
et al., 2014), namely p
VS − sup m11 (t) ≥ σ11 θ11 , (13)
t≥0

for t ≥ max{α11 , θ11 }, we found an asymptotic approximation for P11 (t):

P11 (t) ≈ 1 − e−h11 t (14)

with
[VS − (ρ11 + µ11 θ11 )]2
 
VS − (ρ11 + µ11 θ11 )
h11 = √ exp − 2 ; (15)
θ11 σ11 πθ11 θ11 σ11

Page 7 of 19
 h11 can also be used as an estimate of the firing rate of N11 in the asymptotic
regime. The asymptotic approximation (14) of the FPT distribution of T11
allows us to obtain a closed-form approximation for the mean of the process
Ve12 (t). We report it here, for the particular case θ12 = α12 ,

t
ip
h i
e ∗12 (t) = v0,12 e−t/θ12 + (ρ12 + µ12 θ12 ) 1 − e−t/θ12 + i0,12 te−t/θ12
m
" #
−t/θ12 e−h11 t − e−t/θ12 e−t/θ12 (e−h11 t − 1)
+ k12 θ12 − (t + θ12 )e − − .

cr
1
θ12 − h11
h11
(16)
This expression discloses how the rate parameter h11 of the firings of N11 affects

us
the mean of the process Ve12 (t). It is worth noting that for t → ∞ the term with
k12 θ12 survives in the expression (16) of the mean m e ∗12 (t). It arises from the
last term in (9), that reads for θ12 = α12

an
Z th i
k12 e−t/θ12 1 − e−s/θ12 es/θ12 P11 (s)ds.
0

Now, as t → ∞ the distribution function P11 (t) = P rob(T11 ≤ t) tends to 1,

as does the exponential distribution function (14) that approximates it when
M
neuron N11 is in the asymptotic regime. Hence, for t → ∞, m e ∗12 (t) tends to
ρ12 +µ12 θ12 +k12 θ12 , embodying the additional term k12 θ12 due to the interaction
with the nearest neuron. This result, that could appear as counterintuitive, is
indeed easily explained: although the neuron N11 fires after a long time with
d

high probability, there is a non-zero probability that it fires in short times and
this affects the mean of the membrane potential of neuron N12 , resulting in (16)
te

indicated above and in its asymptotic form.

Finally, an asymptotic approximation can also be obtained for the FPT
distribution function P12 (t) of N12 . If conditions similar to (13) hold, i.e.
p

p
VS − sup m12 (t) ≥ σ12 θ12 , (17)
t≥0
ce

for t ≥ max{α12 , θ12 }, the following approximation can be considered

P12 (t) ≈ 1 − e−h12 t , (18)
with
Ac

[VS − (ρ12 + µ12 θ12 )]2

 
VS − (ρ12 + µ12 θ12 + k12 θ12 )
h12 = √ exp − 2 . (19)
θ12 σ12 πθ12 θ12 σ12

3.2. N21 dynamics

It suffices to notice that the subthreshold evolution of the membrane poten-
tial of N21 is similarly influenced only by N11 , so that it can be described by an
SDE formally identical to (4), with subscripts 21 replacing 1j. All the related
approximations for N21 are substantially analogous to the ones reported above
for N12 : they can be simply rewritten by switching the subscripts 12 and 21.
The asymptotic approximations for P12 (t) and P21 (t) will then be exploited to
obtain the mean of the GD process describing the N22 dynamics.

Page 8 of 19
 3.3. N22 dynamics
N22 receives inputs from different surrounding neurons, so that its dynamics
represents a key step towards the generalization of the model for a neuron Nij

t
embedded in an (N × N )-size network. The associated GD process can provide

ip
predictions on its firing density by the IE approach. Its characterization is the
main theoretical result in this network model.
The evolution of the membrane potential of N22 can be described by

cr
 
1 ρ22 + µ22 θ22
dV22 (t) = − V22 (t) + + I22 (t) dt + σ22 (t)dW22 (t) (20)
θ22 θ22

us
with V22 (0) = v0,22 and
 
I22 (t) = i0,22 e−t/α22 + k22 1 − e−t/α22 H22 (t).

an
It is influenced by all of the other three neurons in the considered network.
Then, the general stochastic linkage function H(t) defined in (3) in this case is
given by
H22 (t) = β11 H11 (t) + β12 H12 (t) + β21 H21 (t),
M
with (
0, if t < τij ,
Hij (t) := (21)
1, if t ≥ τij
d

where τij is the first spike time of neuron Nij for i, j = 1, 2 and i · j 6= 4.
The choice of the weights βij can model different impacts of the related presy-
te

naptic neurons on the basis of the network architecture. Here, we choose

β11 = 1/5, β12 = β21 = 2/5. Similarly, the coefficient k22 is related to the
internal response of the considered neuron to the stimuli and then it can mod-
ulate the balance in this response between incoming excitatory and inhibitory
p

inputs.
Along the lines of Buonocore et al. (2014), we are able to write in closed
ce

form also the mean of this last stochastic process V22 (t) conditioned upon the
spikes of the other three neurons,
h i
m22 (t) ≡ E[V22 (t)|T11 , T12 , T21 ] = v0,22 e−t/θ22 +(ρ22 + µ22 θ22 ) 1 − e−t/θ22
Ac

 −t/α22 Z t
− e−t/θ22

e h i
+i0,22 + e−t/θ22 k22 1 − e−s/α22 H22 (s)es/θ22 ds.
1/θ22 − 1/α22 0
(22)

Now, in order to exploit the integral equation approach, we consider a GD

process Ve22 (t) with covariance in the form (10) and mean function obtained
from (22) by replacing H22 (s) with its mean value
1 2 2
E[H22 (s)] = P11 (s) + P12 (s) + P21 (s)
5 5 5

Page 9 of 19
 where, as before, Pij (t) := P rob(Tij ≤ t) is the FPT distribution for Vij (t), for
i, j = 1, 2 and i · j 6= 4. This process has mean function
h i e−t/α22 − e−t/θ22

t
e 22 (t) = v0,22 e−t/θ22 + (ρ22 + µ22 θ22 ) 1 − e−t/θ22 + i0,22
m
(1/θ22 − 1/α22 )

ip
Z th
k22 −t/θ22 i
+ e 1 − e−s/α22 es/θ22 [P11 (s) + 2 P12 (s) + 2 P21 (s)] ds.
5 0

cr
(23)

Note that the distribution functions Pij (t) of the FPTs of V11 (t), V12 (t) and
V21 (t) are required, at least by means of numerical evaluation. However, if the

us
asymptotic conditions (13) and (17) are satisfied for all the other neurons, for
time t large enough, approximations of the required Pij (t) are available in closed
form. In particular, we can adopt the asymptotic exponential approximation
(14) for P11 and the corresponding expressions (18) for P12 and P21 . Hence,

an
under asymptotic regime, an hyperexponential-type distribution function can
be adopted for the firing activity of the surrounding neurons. In this case, the
mean of the process Ve22 (t) is approximated without any previous numerical
evaluation of the FPT distribution functions of other neurons:
M
h i e−t/α22 − e−t/θ22
e ∗22 (t) = v0,22 e−t/θ22 + (ρ22 + µ22 θ22 ) 1 − e−t/θ22 + i0,22
m
(1/θ22 − 1/α22 )
Z th
k22 −t/θ22 i
1 − e−s/α22 es/θ22 1 − e−h11 s + 2 1 − e−h12 s + 2 1 − e−h21 s ds




+ e
d

5 0
h i e−t/α22 − e−t/θ22
= v0,22 e−t/θ22 + (ρ22 + µ22 θ22 ) 1 − e−t/θ22 + i0,22
te

(1/θ22 − 1/α22 )
−t/α22 −t/θ22
h i e −e
+ k22 θ22 1 − e−t/θ22 − k22
(1/θ22 − 1/α22 )
p

 −hij t
− e−t/θ22 e−(hij +1/α22 )t − e−t/θ22

k22 X e
+ i·j −
ce

5 hij − 1/θ22 hij + 1/α22 − 1/θ22

i·j6=4
(24)

with h11 and h12 given by (15) and (19) and with a corresponding expression
Ac

for h21 ; also in this case note the role of the rate parameters of the adjacent
neurons. In the particular case θ = α, we find (omitting for simplicity all the

10

Page 10 of 19
 subscripts 22 in the parameters θ, ρ, µ)
h i
me ∗22 (t) = v0 e−t/θ + (ρ + µθ) 1 − e−t/θ + i0 te−t/θ

t
e−h11 t − e−t/θ e−t/θ (e−h11 t − 1)
 
k22 −t/θ

ip
+ θ − (t + θ)e − 1 −
θ − h11
5 h11
−h12 t −t/θ −t/θ −h12 t
 
2k22 −t/θ e −e e (e − 1)
+ θ − (t + θ)e − −

cr
1
θ − h12
5 h12
e−h21 t − e−t/θ e−t/θ (e−h21 t − 1)
 
2k22
+ θ − (t + θ)e−t/θ − 1 −
θ − h21
5 h21

us
h i (25)
= v0 e−t/θ + (ρ + µθ) 1 − e−t/θ + k22 θ + (i0 t − k22 (t + θ)) e−t/θ
k22 e−h11 t − e−t/θ e−t/θ (e−h11 t − 1)
 
− 1 +
θ − h11
5 h11

an
 −h12 t −t/θ −t/θ −h12 t

2k22 e −e e (e − 1)
− 1 +
θ − h12
5 h12
 −h21 t
− e−t/θ e−t/θ (e−h21 t − 1)

2k22 e
M
− 1 + .
θ − h21
5 h21

The obtained mean function m e ∗22 (t) allows us to adopt the investigation strategy
based on the IE (12), rewritten for the GD process Ve22 (t). We remark that, by
d

using the above approximation for the mean Ve22 (t), we avoid the preliminary
application of the iterative numerical procedure to Ve11 (t), Ve12 (t) and Ve21 (t); we
te

simply apply the procedure to the process Ve22 (t) to evaluate the related FPT
density and use it as a prediction of the firing density of N22 .
p

4. Simulations and numerical results

ce

To assess the effectiveness of our approximations, we perform some experi-

ments to compare the FPT densities of the involved processes Veij (t), as obtained
via the IE approach and by numerical quadrature of equations of type (12), with
the histograms of the FPT obtained through simulation of the trajectories of
Ac

the processes Vij (t) via discretization of the corresponding SDEs. In all exper-
iments we assume as a reference level the resting potential ρ = 0 = v0 : after
these settings, we rescale all terms in Eqs.(4),(20) accordingly.
Our first experiment simulates an inhibitory network: parts a), b) and c) of
Figure 2 show the histograms of the firing times of N12 , N21 , N22 as obtained
by numerical simulation of 20000 trajectories of the involved neurons until their
first spike, i.e. until the first crossing time for the threshold VS . An Euler dis-
cretization method is applied to the corresponding SDEs (4) for the simulation
of random paths of V11 (t), V12 (t), V21 (t) and to (20) for random paths of V22 (t).
An ad hoc simulation code has been implemented to simultaneously record the
spike times of the four neurons. The histogram is constructed on the sample of

11

Page 11 of 19
 t
ip
cr
us
an
M
d

Figure 2: Histograms of first spike times and numerical approximations of FPTpdf for N12 (a),
te

N21 (b) and N22 (c) and plot of the numerical approximations for all the 4 neurons FPT
densities (d) in the case of inhibition. Parameters values for all neurons: ρ = 0, µ = 0,
θ = α = 1, σ = 1, v0 = 0, VS = 2, i0 = 0.25, k12 = −0.2, k21 = −0.1, k22 = −0.25
p

simulated crossing times. The firing densities of the same neurons, obtained by
ce

numerical quadrature of the integral equation (12), are superimposed for com-
parison. The agreement between these curves and the corresponding histograms
is very satisfactory. Finally, part d) of the same Figure shows the differences
among the firing densities of all the four considered neurons. In this setting, the
Ac

asymptotic approximations lead to a very satisfactory agreement with simula-

tion results. The comparison of all these curves is reported in Figure 5 below.
The second experiment simulates an excitatory network: in Figure 3, parts
a), b) and c), the histograms of the firing times of neurons N12 , N21 , N22 , as ob-
tained by simulating 20000 trajectories, are compared with the firing densities
of the same neurons, as obtained by numerical quadrature of the corresponding
Volterra integral equations of type (12). The agreement between these curves
and the corresponding histograms is still very satisfactory. Again, part d) of
the same Figure shows the differences among the firing densities of all the four
considered neurons in the excitatory network. The parameter values are spec-
ified in the caption of the Figures; when they are reported without subscripts,

12

Page 12 of 19
 t
ip
cr
us
an
M
d

Figure 3: Histograms of first spike times and numerical approximations of FPTpdf for N12 (a),
te

N21 (b) and N22 (c) and plot of the numerical approximations for all the 4 neurons FPT
densities (d) in the case of excitation. Parameters values for all neurons: ρ = 0, µ = 0,
θ = α = 1, σ = 1, v0 = 0, VS = 2, i0 = 0.25, k12 = 0.2, k21 = 0.5, k22 = 0.35
p

they have the same value for all the processes.

ce

Finally, as a particular case, we consider a three-neuron network, in which

the neuron N21 has been excluded from the interaction. This biological phe-
nomenon can happen for natural reasons (death and renovation of neuronal
cells) or for diseases or drugs effects and it is here considered to show how the
Ac

model can describe networks with a different structure by suitable modifica-

tions of the parameters. The SDEs describing the sub-threshold evolution of
the membrane potentials for the involved neurons are formally identical to (4),
(20) apart from the linking weights βij in the function H22 that are now chosen
(3) (3)
as β11 = 1/3, β12 = 2/3 to account for different distances between neurons
(The superscript (3) is here adopted to specify the different setting of this net-
(3)
work). In Figure 4, the histogram of the simulated firing times of neuron N22
as obtained by discretization of the corresponding SDE, and the estimated FPT
pdf obtained by solving (numerically) the corresponding integral equation are
(3)
shown. It is interesting to compare the firing densities of neurons N22 and N22

13

Page 13 of 19
 t
ip
cr
us
Figure 4: Histogram of first spike times and numerical approximations of FPT pdf for
(3)
N22 (left) and plot of the numerical approximations for all the 3 neurons (right). Param-

an
eters values for all neurons: ρ = 0, µ = 0, θ = α = 1, σ = 1, v0 = 0, VS = 2, i0 = 0.25,
k12 = 0.2, k22 = 0.35

as shown in Figures 3 and 4: the model detects the presence of an additional ac-
M
tive neuron in the (2 × 2)-size network, that it missing in the 3-neurons scheme,
and its influence on the firing density of the neuron N22 . Asymptotic regime
can also be considered and the related approximated mean provided. The fol-
lowing Figure 5 shows histograms for neurons N11 and N12 , respectively, with
the corresponding firing densities and asymptotic approximations as given by
d

(14) and the first of (18); it also compares the corresponding firing densities of
(3)
N22 in the simple 3-neurons network as obtained by numerical (red line) and
te

asymptotic (blue line) approximations of the other two neurons firing distribu-
tions. The two curves are practically identical, confirming the very good results
of the asymptotic approximation.
p

5. A (N × N )-size network
ce

It is worth underlining that the full numerical procedure becomes very hard
to implement for a network of dimension higher than 2 and quite unfeasible
for large networks. The system of integral equations written for the linked GD
Ac

processes requires a cascade-solving numerical procedure, whose computational

time rapidly grows with the dimension of the network. The use of exponential
approximations for the firing activity of surrounding neurons can make this pro-
cedure more attractive. Here, we explain how the proposed model can describe
the firing density of a neuron embedded in a (N × N )-size network. In this case,
the evolution of the membrane voltage of the Neuron Nij , with i, j = 1, . . . , N,
can be described by the stochastic process Vij (t) solution of the following SDE,
for t ≥ 0,
 
1 ρij + µij θij
dVij (t) = − Vij (t) + + Iij (t) dt + σij (t)dWij (t) (26)
θij θij

14

Page 14 of 19
 t
ip
cr
us
an
M
d

Figure 5: Histograms, numeric (red) and asymptotic (blue) approximations for N11 , N12 and
te

(3)
corresponding numerical approximations for N22 . Parameter values: ρ = 0, µ = 0, θ = α = 1,
v0 = 0, VS = 2, i0 = 0.25, k12 = k22 = −0.1.
p

with Vij (0) = v0,ij and

ce

 
Iij (t) = i0,ij e−t/αij + kij 1 − e−t/αij Hij (t),

for i, j = 1, . . . , N. In this case, the linking function Hij (t) is given by:
Ac

i j
1 X X
Hij (t) = n · m · Hnm (t), (27)
M n=i−1 m=j−1
Pi Pj
with Hij (t) ≡ 0 and M = n=i−1 m=j−1 n · m − i · j.
We can also rewrite this linkage function as

Hij (t) = β(i−1)(j−1) H(i−1)(j−1) + β(i−1)j H(i−1)j + βi(j−1) Hi(j−1) , (28)

where βnm = nm/M for n = i−1, i; m = j −1, j and n·m 6= i·j. As an example,
Figure 6 shows the case of a (3 × 3)-size network with β weights specified along
arrows. In order to apply our investigation strategy to this network, we need to

15

Page 15 of 19
 t
ip
cr
us
Figure 6: A (3 × 3)-size network with specified values for the linkage weights βij along the

an
connection arrows.

specify the mean of the GD process Veij (t) describing the dynamics of a generic
Neuron Nij with i, j = 1, . . . , N. This mean function includes the term
M
i j
1 X X
E [Hij (s)] = n · m · P(Tnm ≤ s)
M n=i−1 m=j−1
d

representing the connections among neurons in the model. The evaluation of

the involved probability distributions defines the computational complexity of
te

the numerical procedure: even if only the firing distributions of the three clos-
est neurons are required, each of them needs to be previously evaluated by a
similar procedure. Alternatively, we can assume the asymptotic regime for the
p

three neurons which are closer to the neuron Nij ; hence the dimension N of the
network does not affect the numerical evaluation, and only the linkage parame-
ce

ters have to be specified according to the position of the neuron Nij within the
network. Along the lines of the approach developed in the previous Sections,
we can approximate
i j
Ac

1 X X
n · m 1 − e−hnm s
 
E [Hij (s)] ≈ (29)
M n=i−1 m=j−1

assuming that the asymptotic regime holds for the neurons having a direct
connection with the observed one Nij and with

[VS − (ρnm + µnm θnm + knm θnm )]2

 
VS − (ρnm + µnm θnm + knm θnm )
hnm = √ exp − 2
.
θnm σnm πθnm θnm σnm

It is then possible to obtain a closed form expression for the mean of this last
process and so a numerical approximation of the FPT pdf via the IE approach.

16

Page 16 of 19
 Indeed, the asymptotic approximation of the mean of Veij (t) is given by
h i e−t/αij − e−t/θij
e ∗ij (t) = v0,ij e−t/θij + (ρij + µij θij ) 1 − e−t/θij + i0,ij
m
(1/θij − 1/αij )

t
ip
i j Z th
kij X X i
n · me−t/θij 1 − e−s/αij es/θij 1 − e−hnm s ds
 
+
M n=i−1 m=j−1 0

(30)

cr
with hij = 0. In the case θ = α (also omitting the explicit dependence on i, j in
the parameters v0 , ρ, µ, θ) this mean function reduces to

us
i j
h i kij X X
e ∗ij (t) = v0 e−t/θ +(ρ + µθ) 1 − e−t/θ + i0 te−t/θ +
m n·m
M n=i−1 m=j−1

an
(e−hnm t + hnm t − 1)
 
θ
× θ(1 − e−t/θ ) − e−t/θ + (e−t/θ − e−hnm t ) .
hnm 1 − θhnm
(31)
M
We can adopt the GD process Veij (t) with the above mean function m e ∗ij (t)
and covariance (10) to approximate the evolution of the membrane voltage of
the neuron Nij embedded in an (N × N ) network. By numerical evaluation
of the FPT density gVeij [VS , t|v0,ij , 0] as a solution of the IE (12) rewritten for
d

the process Veij (t), we can finally estimate the firing density of Nij , without any
previous simulation or numerical evaluation of the FPT of other neurons.
te

6. Concluding Remarks
p

The proposed synaptic current-based model for the description of the firing
activity of neurons in a finite-size network is essentially constructed by using
ce

a class of linked GD processes that solve a system of SDEs of type (26). The
FPTs of such processes through the firing threshold VS , numerically evaluated
by quadrature of a system of IEs of type (12), provide estimates of the cor-
responding firing densities. The specified SDEs can also be used to simulate
Ac

the stochastic dynamics, but in large networks the computational complexity

and the reliability of the approximations are hard to quantify. Alternatively,
the numerical procedure applied to solve the IE provides FPT estimates with
high degree of accuracy for the GD processes approximating the solution of the
considered SDEs. Comparisons are here provided with the simulation results
to validate the assumption of GD processes as suitable models for the above
neuronal dynamics. Furthermore, in Section 5 we specify how this model can
be used for the firing activity of a neuron embedded in an (N × N ) network
with N ≥ 3. We note that the expression obtained for the mean (31) of the
approximating GD process, under the asymptotic regime assumption, allows to
describe the behavior of the membrane voltage of the generic neuron Nij and

17

Page 17 of 19
 to provide a prediction of the related firing activity without any previous nu-
merical evaluations. The fundamental role in building this network model is
played by the network of size (2 × 2): indeed, any considered larger network is
a collection of a finite number of (2 × 2) networks. Finally, we remark that it

t
is possible to extend our model to investigate other interesting cases, such as

ip
mutual interaction between neurons, colored noise inputs, periodic driving sig-
nals, joint effect of inhibitory and excitatory neurons and also synchronization
phenomena. Some of these aspects will be object of future work.

cr
Acknowledgments

us
The authors would like to thank the anonymous reviewers whose thoughtful
and accurate comments and suggestions truly help improve the manuscript.

an
References

Arbib, M., 2003. The Handbook of Brain Theory and Neural Networks. MIT
Press.
M
Brunel, N., 2000. Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons. J Comp Neurosci 8, 183–208.
Brunel, N., Hansel, D., 2006. How noise affects the synchronization properties
of recurrent networks of inhibitory neurons. NeCo 18, 1066–1110.
d

Buice, M., Chow, C., 2013. Dynamic finite size effects in spiking neural networks.
te

PLoS Comput. Biol. 9 (1), e1002872.

Buonocore, A., Caputo, L., D’Onofrio, G., Pirozzi, E., 2015a. Closed-form solu-
tions for the first-passage-time problem and neuronal modeling. Ricerche di
p

Matematica 64 (2), 421–439.

ce

Buonocore, A., Caputo, L., Nobile, A. G., Pirozzi, E., 2015b. Restricted
Ornstein-Uhlenbeck process and applications in neuronal models with pe-
riodic input signals. J Comput Appl Math 285, 59–71.
Buonocore, A., Caputo, L., Pirozzi, E., Carfora, M. F., 2013. On the dynamics
Ac

of a couple of mutually interacting neurons. In: Moreno-Dı́az, R., Pichler, F.,

Quesada-Arencibia, A. (Eds.), Computer Aided Systems Theory - EURO-
CAST 2013. Vol. 8111 of Lecture Notes in Computer Science. Springer Berlin
Heidelberg, pp. 36–44.
Buonocore, A., Caputo, L., Pirozzi, E., Carfora, M. F., 2014. Gauss-diffusion
processes for modeling the dynamics of a couple of interacting neurons. Math
Biosci Eng 11 (2), 189–201.

18

Page 18 of 19
 Carfora, M. F., Pirozzi, E., 2015. Stochastic modeling of the firing activity of
coupled neurons periodically driven. In: de Leon, M., Feng, W., Feng, Z.,
Lopez-Gomez, J., Lu, X., Martell, J., Parcet, J., Peralta-Salas, D., Ruan, W.
(Eds.), Dynamical Systems, Differential Equations and Applications, AIMS

t
Conference 2015. AIMS Proceedings. American Institute of Mathematical Sci-

ip
ences, pp. 195–203.
D’Onofrio, G., Pirozzi, E., 2015. On two-boundary first exit time of Gauss-

cr
diffusion processes: closed-form results and biological modeling. Lecture Notes
of Seminario Interdisciplinare di Matematica 12, 111–124.
Forcaud, N., Brunel, N., 2002. Dynamics of the firing probability of noisy

us
integrate-and-fire neurons. Neural Computation 14, 2057–2110.
Galves, A., Locherbach, E., 2016. Modeling networks of spiking neurons as
interacting processes with memory of variable length. J Soc Fran Stat 157 (1),

an
17–32.
Greenwood, P. E., Ward, L. M., 2016. Stochastic Neuron Models. Springer In-
ternational Publishing.
M
Hansson, B. S., 1995. Olfaction in lepidoptera. Experientia 51, 1003–1027.
Lansky, P., Ditlevsen, S., 2008. A review of the methods for signal estimation
in stochastic diffusion leaky integrate-and-fire neuronal models. Math Biosci
Eng 99 (4), 253 – 262.
d

Ostojic, S., Brunel, N., Hakim, V., 2009. Synchronization properties of networks
of electrically coupled neurons in the presence of noise and heterogeneities. J
te

Comp Neurosci 26, 369 – 392.

Politi, A., Luccioli, S., 2010. Dynamics of networks of leaky-integrate-and-fire
p

neurons. In: Estrada, E., Fox, M., Higham, D., Oppo, G. (Eds.), Network
Science: Complexity in Nature and Technology. Springer-Verlag London, pp.
ce

217–242.
Postnova, S., Finke, C., Jin, W., Schneider, H., Braun, H., 2010. A computa-
tional study of the interdependencies between neuronal impulse pattern, noise
effects and synchronization. J Physiol 104 (3 – 4), 176 – 189.
Ac

Rospars, J. P., Gremiaux, A., Jarriault, D., Chaffio, A., Monsempes, C., Deisig,
N., Anton, S., Lucas, P., Martinez, D., 2014. Heterogeneity and convergence
of olfactory first-order neurons account for the high speed and sensitivity of
second-order neurons. PLoS Comput Biol 10 (12), e1003975.
Sirovich, R., Sacerdote, L., Villa, A. E. P., 2010. Cooperative behavior in a
jump diffusion model for a simple network of spiking neurons. Math Biosci
Eng 11 (2), 385 – 401.
Soula, H., Chow, C., 2007. Stochastic dynamics of a finite-spize spiking neural
network. Neural Computation 19, 3262–3292.

19

Page 19 of 19