Charles University in Prague

First Faculty of Medicine

Study program: Biomedical Informatics

Mgr. Pavel Šanda

Information processes in neurons

Informační procesy v neuronech

PhD Thesis

Thesis advisor: Doc. RNDr. Petr Lánský DrSc.

Institute of Physiology,
Academy of Sciences of the Czech Republic

Prague, 2012
Declaration

I declare that I have written this thesis on my own and listed all used
sources. I also declare that the work has not been used to obtain another
or the same university degree. I agree with the permanent archiving of the
electronic version of the thesis in the database project Theses.cz in order to
allow systematic similarity check of the qualification works.

Prohlášení

Prohlašuji, že jsem závěrečnou práci zpracoval samostatně a že jsem řádně

uvedl a citoval všechny použité prameny a literaturu. Současně prohlašuji,
že práce nebyla využita k získání jiného nebo stejného titulu.
Souhlasím s trvalým uložením elektronické verze mé práce v databázi
systému meziuniverzitního projektu Theses.cz za účelem soustavné kontroly
podobnosti kvalifikačních prací.

V Praze, 26.6.2012

Pavel Šanda

2
Identification Record

SANDA, Pavel. Information processes in neurons. Prague, 2012. 76 pages.

PhD Thesis. Charles University in Prague, 1st Faculty of Medicine, Dept.
of Computational Neuroscience FGU AS CR. Supervisor Petr Lansky.

Identifikační záznam

ŠANDA, Pavel. Informační procesy v neuronech. [Information processes

in neurons.] Praha, 2012. 76 stran. Dizertační práce. Univerzita Karlova
v Praze, 1. lékařská fakulta, Oddělení početních neurověd FGÚ AVČR.
Vedoucí práce Petr Lánský.

3
I would like to thank my supervisor Petr Lansky for the support he provided
during all the research. My thanks go also to Petr Marsalek for all his help.
I am also grateful to various institutions in Prague and abroad which
allowed me to stay and perform the research work.

The thesis is based on 5 published papers (detailed in the section List of

publications on page 38), the following chapters give extended introduction
and context for the whole work.
 Information processes in neurons

Contents
List of Abbreviations 6

1 Introduction 7
1.1 Level of description . . . . . . . . . . . . . . . . . . . . . . . 8

2 Integrate and Fire neuronal model 10

2.1 Types of Integrate and fire model . . . . . . . . . . . . . . . 10
2.2 The parameters of LIF . . . . . . . . . . . . . . . . . . . . . 12

3 Neural code 16
3.1 Types of neural code . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Spatial hearing . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Neural code for spatial hearing of mammals . . . . . . . . . 21
3.3.1 Low frequency ITD cues and modeling . . . . . . . . 21

4 Conclusions 24

References 27

List of publications 38

Reprints of published papers 39

List of Abbreviations
IF integrate-and-fire model

LIF leaky integrate-and-fire model

ISI Interspike interval

ITD interaural time difference

MSO medial superior olive

CN cochlear nucleus

L/MNTB lateral/medial nucleus of trapezoid body

6
 1 INTRODUCTION

1 Introduction
Starting with the work of Santiago Ramón y Cajal, neuronal cells in the
brain were recognized as independent units which communicate via the con-
tact of axons to dendrites and the body of nerve cells, and create neuronal
circuits through branching of their fibers (Ramon y Cajal, 1899), English
translation Ramon y Cajal (1995).
Simultaneously to Cajal’s morphological findings, a larger group of sci-
entists discovered the existence of action potentials (impulses, spikes) trav-
eling through nerve fibers. Since these impulses are similar in duration and
shape, this naturally lead to the binary all-or-none concept. As part of what
would be later called the neuronal doctrine, Cajal proposed unidirectional
transmission of nerve impulses from dendrites through soma to axon and
called it the law of dynamic polarization.
In the same way as Cajal postulated the neuron as a basic anatomical
unit, McCulloch and Pitts postulated the neuron as a basic unit of infor-
mation processing and used the all-or-none concept for modeling nervous
activity on the basis of logical calculus. In their seminal work a formal model
of the neuron was formulated and it was even indicated that a network of
such formal neurons is Turing-complete (McCulloch and Pitts, 1943).
With experimental research developing, such level of formal description
of information processing in a neuron was no more adequate. No later
than in 1959 it was observed that many electrical events on the membrane
are of a continuous nature and that there exists some background sponta-
neous activity (Bullock, 1959). The following decades added new findings,
which were either beyond neuronal doctrine or even contradicting it and the
whole picture became more complicated (Bullock et al., 2005). Similarly the
question where and on which scale (or level of description) information pro-
cessing takes place became problematic. For example glia cells were found
to communicate with each other via transmitters and gap junctions (Fields
and Stevens-Graham, 2002), chemical synapses between glia cells and neu-
rons were found as well (Bergles et al., 2000). Thus, it is possible that there
is parallel information processing going on at slower time scales.

7
 1 INTRODUCTION

1.1 Level of description

Another ambiguity stems from the choice of the proper level of descrip-
tion. We can distinguish the level of large neural networks, simple circuits,
individual neurons, the sub-cellular level, membranes and the underlying
biochemistry. Traditionally the community around artificial neuronal net-
works does not use a detailed description of neurons and is satisfied with ab-
stract models not much different from the original McCulloch-Pitts neuron.
This abstraction would be hardly acceptable for the community studying
the features of single neurons and their membrane for its drastic simplifica-
tion since it would be impossible to mimic many of the effects observed in
physiological experiments. However, it does not automatically follow that
a more detailed description opens up a better understanding of the system
as a whole.
One such example is Hopfield’s work on the neural network for content-
addressable memory (Hopfield, 1982). According to the critics of this paper,
the neurons should have continuous input-output relations, moreover real
neurons and circuits have integration time delays due to the capacitance
of the neuronal membrane. Therefore, the time evolution of the state of
such systems should be represented by a continuous time representation.
In his response Hopfield showed that the important properties of the orig-
inal model remain intact when these two simplifications of the model are
eliminated (Hopfield, 1984).
On the other hand, a more detailed model can completely change the
way how information processing is implemented. For example in order to
compute a certain formula from logical calculus, the classical McCulloch-
Pitts approach needs to assemble a circuit from neurons in a similar fashion
as when logical gates are assembled in modern digital computers. However,
when we stop to look at a neuron as a simple one-point integrator of in-
coming signal and make a detailed model of a branching dendritic tree, we
get a very different picture of possible computations within a single neuron
only. Decomposing the dendritic tree of the neuron into subunits (Koch
et al., 1982) shows that the combination of a specific branching topology,

8
 1 INTRODUCTION

and the precise timing of excitatory and inhibitory inputs implements the
approximation of logical gates (AND NOT, OR, AND) and even multiplica-
tive arithmetical operations. Since all logical operations can be defined via
AND and AND NOT gates, any logical operation can be synthesized by
local circuits consisting of synapses between the dendrites of two or more
neurons (Koch et al., 1983).
However, the discussion about the level of detail and precision in the
modeling of information flow inside the nervous system does not necessarily
end at the level of membrane biophysics. In a series of papers Hjelmfelt
showed that even enzymatic reactions inside cells can be interpreted as
information processing and he also showed that it is possible to construct
a universal Turing machine based on such reactions (Hjelmfelt et al., 1991;
Hjelmfelt and Ross, 1992, 1993).
While the advantages of a more detailed description are clear, there is
also a price to be paid. Firstly, more details of the model usually require
more parameters and it is often hard or even impossible to obtain such
parameters from experimental setup. Secondly, a more detailed description
might be intractable from a mathematical point of view, and no deeper
insight about the dynamics of the system can be obtained. Thirdly, when
we are interested in dynamics of large scale networks, the simulation of
detailed models can be very demanding on the computational power.
To sum up, there is no “proper” level of description unless we take into
account the aim and context of the study. In the rest of the text we will
have a look at the class of simplified models of the neuronal membrane.
Then we will inquire into possible encoding schemes of the action potential
sequences (“spike-trains”) generated by the neurons (or by their models
respectively). Finally we will focus on the problem of spatial hearing and
the different neuronal coding mechanisms used to explain animal sensitivity
of sound localization.

9
 2 INTEGRATE AND FIRE NEURONAL MODEL

2 Integrate and Fire neuronal model

2.1 Types of Integrate and fire model
Three basic types of models of a single neuron can be distinguished - digital
(all-or-none) and continuous, which can be subsequently modeled either
as a single point in a space or with more unit compartments simulating
morphology of the real neuron and its branching structure (Segev, 1992).
We shall focus on the single-point continuous models only.
Generally, the family of continuous models is described by an electrical
circuit representing the iso-potential patch of a membrane. The simplified
model of a neuron membrane as an electrical circuit consisting of capaci-
tor with leak was first proposed by Lapicque (1907), translation and review
Lapicque (2007); Brunel and van Rossum (2007). It was, however, before the
mechanisms of the action potential firing were understood and the first for-
mulations of the whole leaky-integrate-and-fire (LIF) model appeared later
(Stein, 1965; Knight, 1972).
The most simple version of the model circuit consists of capacitor C only
(representing lipid layers of the membrane) and is called perfect integrator
(Koch, 2005). Membrane voltage V (t) can be written in this case as

dV (t)
C = I(t) , (1)
dt

where I(t) represents the current from synaptic input at time t (or intra-
cellular electrode). The spiking mechanism is missing in this model and it
is usually described just as a complete reset after certain voltage threshold
S is reached; switch through which the accumulated voltage is discharged
would implement such behaviour inside electrical circuit.
If the neuronal membrane consists only of a twofold lipid layer, the
voltage would increase no matter how slowly the incoming current arrives
since the membrane functions as a perfect insulator. In biological reality
the membrane contains also proteins which form specific channels through
which various ions can flow and leak the charge. In such case the voltage
on the membrane does not integrate as above and additional resistor R

10
 2 INTEGRATE AND FIRE NEURONAL MODEL

implementing this leakage is added in parallel to capacitor in the electric

circuit thus forming leaky-integrate-and-fire (or forgetful) model. The cur-
rent through the resistor follows Ohm’s law and the equation (1) can be
rewritten as
dV (t) V (t)
C + = I(t) . (2)
dt R
Again, the voltage is integrated in subthreshold regimen (see section 2.2)
and once threshold S is reached, voltage is reset to its initial value and
integration starts anew.
Specific integrate-and-fire (IF) models which employ spike generation
as integral and emergent part of the model have been proposed, for exam-
ple by Fitzhugh (1961) who suggested simplification of the Hodgkin-Huxley
model (see below) and Nagumo et al. (1962) who formulated a correspond-
ing electronic circuit. The system is described by two coupled differential
equations and its advantage compared to the Hodgkin-Huxley model is that
the mathematical properties can be understood quantitatively.
Because of the random nature of incoming synaptic signal, a stochastic
version of integrate-and-fire model has been developed. Initial work was
done by Gerstein and Mandelbrot (1964) who used random walk to mimic
excitatory and inhibitory input pulses. The random walk counterpart of
the leaky model has been formulated by Stein (1965, 1967). Subsequently,
a continuous model can be obtained as a limit case thus obtaining Ornstein-
Uhlenbeck model (Uhlenbeck and Ornstein, 1930; Ricciardi and Sacerdote,
1979), which can be written as the equation

dV (t) V (t)
C + − Cσξ(t) = I(t) , (3)
dt R

where ξ represents white noise from the synaptic input and σ its variability.
A plethora of integrate-and-fire model variants can be found in a recent
review of Burkitt (2006a,b).
As a side note we shall remember that apart from the phenomenological
models above, more biologically plausible models exist including a channel
based biophysical description. Outstanding among there is the description

11
 2 INTEGRATE AND FIRE NEURONAL MODEL

by Hodgkin-Huxley who formulated their model after a long series of exper-

iments on the giant squid axon (Hodgkin and Huxley, 1952). Total mem-
brane current I can be written as a sum of ionic currents on the membrane
and capacitave current

dV (t)
C + IN a (t) + IK (t) + Ileak (t) = I(t) , (4)
dt

where the sodium IN a and potassium IK ionic currents are determined by

the driving potential via Ohm’s law. For a full description of the equations
left and its parameters see Hodgkin and Huxley (1952). While being closer
to the physiological reality, the Hodkgin-Huxley model cannot be analyzed
analytically, has lots of parameters and its simulations require high compu-
tational resources. On the other hand, simplified IF models are both easier
to analyze and simulate, fewer parameters are needed to be estimated from
the data.

2.2 The parameters of LIF

Because IF models have been widely employed in neuronal modeling, there
is always question whether these models are sophisticated enough to mimic
the behaviour of biological neurons (Feng, 2001). Despite of wide usage of
IF models in the theoretical literature there has never been an attempt to
check IF models accuracy and assumption against intracellular recordings of
the membrane voltage (the traditional attempts were more focused on the
interspike intervals (ISI) statistics obtained from extracellular recordings).
In a couple of papers we compared in-vivo intracellular neuronal recordings
and stochastic variant of LIF model (3)

dV (t)
= −β(V (t) − x0 ) + µ + σξ(t), V (0) = x0 (5)
dt

with few modifications, namely we assume that input µ is constant in time

and the process starts at the resting level x0 . The (constant) parameters
C, R form the so-called membrane time constant τ = RC, which we use
1
here inversely as β = RC to conform with notation used later in our work

12
 2 INTEGRATE AND FIRE NEURONAL MODEL

-50

-55

Depolarization [mV]
-60

-65

-70

-75
0.1 0.2 0.3 0.4 0.5 0.6
Time [s]

Figure 1: Three basic regimes of neuronal firing. From top to bottom

we successively see the schematic evolution of a membrane potential in
the suprathreshold/threshold/subthreshold regimen. Spike generation upon
reaching the threshold S (thick solid line) is not considered. Thin dotted
lines converge to asymptotic mean depolarization. We can see that in case
of subthreshold regimen additional noise is the only way how the membrane
potential can reach the threshold.

and let the reader directly delve into it. As in other LIF models spike is not
intrinsic part of description and membrane potential is reset to x0 whenever
V (t) reaches the threshold S.
Parameters of such model are traditionally divided into those, which
depend on membrane properties (β – inverse of the membrane time constant,
S – threshold, x0 – resting level) and those, which depend on the input signal
(µ – mean signal, σ – signal variability).
The asymptotic mean depolarization derived from (5) as E(V (∞)) =
x0 + µ/β determines three regimes of neuronal firing (see Fig. 1):

• the subthreshold regimen (µ/β  S − x0 ) with Poissonian firing. As

the asymptotic mean depolarization does not reach the threshold the
firing depends on the noise and without it the neuron would remain
silent.

13
 2 INTEGRATE AND FIRE NEURONAL MODEL

• the threshold regimen (µ/β ≈ S − x0 ), where the distribution of ISIs

is positively skewed and resembles for example Gamma distribution.

• the suprathreshold regimen (µ/β  S −x0 ), where the firing is almost

regular and ISI histogram resembles normal distribution. The noise
plays a limited role in this range of parameters.

Initially, in Lansky et al. (2006), we obtained parameters and compared

different estimation methods for the spontaneous part1 of the recordings
and checked basic assumptions of the model. We found, that in general the
data are consistent with the model. The spontaneous part is in subthreshold
noise-driven regimen and ISIs are exponentially distributed, which suggests
they are generated in accordance with Poisson process. There were, how-
ever, also inconsistencies with the model. The model assumes that spectra of
the input signal should be flat, while we found characteristic hump around
the frequency of 2200 Hz in data, which was subsequently eliminated by
filtering. Next, the subthreshold regime requires the asymptotic depolar-
ization far below the threshold. In our case the asymptotic threshold is
below the threshold, but it is less than two-standard-deviations envelope.
The model also assumes a fixed value of the reset depolarization x0 , which
did not hold, the effect of this discrepancy for the model performance is
negligible, however.

In the second analysis (Lansky et al., 2010), we compared activity and

estimated parameters for spontaneous and stimulated part of the recordings
and discussed their firing regimes. The simulation based on the estimated
parameters fits well with the course of the membrane depolarization (see
Fig. 2). The parameters dependent on the input signal µ and σ were larger
than in spontaneous part and the overall firing regimen is suprathreshold as
expected. Despite the assumptions that x0 is not dependent on input signal
our data shows that its value is actually influenced by the stimulation.
To summarize, external stimulation affects input parameters and thus
1
The recordings contained spontaneous activity of the neurons in auditory pathway
and stimulated activity triggered by acoustic signal.

14
 2 INTEGRATE AND FIRE NEURONAL MODEL

5 3

4
2
3
Depolarization [mV]

Depolarization [mV]
1
2

1 0

0
-1
-1
-2
-2

-3 -3
0 200 400 600 800 1000 0 2 4 6 8 10 12 14
Time [ms] Time [ms]

Figure 2: The accuracy of the model simulation compared with the recorded
data. The simulation of the model (5) used parameters estimated from the
data. The middle line shows the mean difference of the membrane potential
between simulations and experimental recordings, while the surrounding
lines are 2 standard-deviations envelope. In ideal case we would obtain flat
line at 0. Left panel is comparison of the spontaneous part of the record,
right panel shows comparison of (acoustically) stimulated part of the record.
Different time axis is due to the fact that stimulation brings the neuron to
the suprathreshold regime where interspike intervals are much shorter.

15
 3 NEURAL CODE

time evolution of membrane voltage in (5) and it has direct consequence

on the ISI statistics, since the stimulated intervals will be generally much
shorter.
The currently dominating opinion is that for neural coding is important
just occurrence of the spike (in time), neither its shape nor the membrane
evolution between spikes. From this point of view, the evolution details of
the membrane potential are not important, what matters is the time when
the potential reaches the threshold for firing and the question is whether
such simplified models can reliably predict occurrence of spiking times. The
ability to predict spiking behaviour has been shown for particular types of
neurons in this class of models (Kistler et al., 1997; Keat et al., 2001; Jolivet
et al., 2006; Kobayashi et al., 2009). It needs to be emphasized that in
order to strengthen the predictive power of spiking models, the dynamics of
threshold was included in the IF description, see (Jolivet et al., 2008a,b).

3 Neural code
3.1 Types of neural code
Let us now abstract from the underlying spiking mechanisms and focus only
on the resulting sequence of spikes. Such spike train is often considered
as a code through which information is conveyed across the neural system.
Beginning with the pioneering work of Lord Adrian who showed the relation
between the frequency of spikes of frog muscle receptors and the stretching
the muscle with different weights (Adrian and Zotterman, 1926), the idea
of firing rate code became the textbook model of neural coding.
This idea suggests that there is continuous function f (t) according to
´ t +∆t
which resulting spike train is produced, more precisely p = t00 f (t)dt
determines the number of spikes expected to occur in the interval [t, t + t0 ].
In case ∆t is sufficiently small so that p ≤ 1 we can interpret p as a prob-
ability that the spike occurs in [t, t + t0 ] . Of course such rate function is
not directly visible and we experimentally observe only realization of spikes
sampled from f . The conventional way of estimating such background rate

16
 3 NEURAL CODE

function is recording multiple trials with identical stimuli and average the
rate over all recorded trials into post-stimulus time histogram (PSTH, Ger-
stein and Kiang (1960)). Another approach which has the advantage of
creating smooth function even for single recorded trial is kernel smoothing
(Nawrot et al., 1999).2 In this method the spike train is convolved with
kernel of a particular shape (usually gaussian one) and (band)width. Com-
mon problem of both PSTH and kernel based method is ad-hoc choice of
the histogram bin-size/kernel bandwidth which results in different firing
rate estimates. In series of papers Shimazaki and Shinomoto (2007, 2010)
suggested method for determining optimal width, based on the assumption
that the underlying spike generation process is Poissonian. We employed
the kernel optimization method for the analysis of experimental recordings
of olfactory neurons and found it computationally demanding. In order
to speed up the bandwidth determination we analyzed the algorithm and
transformed it into the parallel one which thanks to its speed-up in the
supercomputer environment allowed online interactive work with the data.
The details are covered in Šanda (2010), the analysis of the data was not
yet published.
The relationship between the mean firing rate and stimulus intensity
has been established (Fuortes, 1958; Matthews, 1964; Grüsser et al., 1968),
more recent examples of firing rate codes are from the monkey visual cor-
tex neurons, where the spike count correlates with motion discrimination
performance (Britten et al., 1992; Celebrini and Newsome, 1994) or stim-
ulus identity (Desimone et al., 1984). For some time it was assumed that
such code captures all relevant information because the firing-rate coding
scheme was robust against unreliability and noise observed in neuronal ac-
tivity. Later decades brought experimental findings which showed that what
was often considered as erratic behaviour (McCulloch, 1959) was rather mis-
understanding of the code transmitted (Barlow, 1972) and neurons can be
thought as a reliable unit of transmission – as an outstanding example Sakitt
2
There is a group of methods employing Bayesian inference for the firing rate estima-
tion which we will not pursue here. For recent review and comparison of various methods
see Cunningham et al. (2009).

17
 3 NEURAL CODE

(1972) showed that even single spike in certain critical sensory neuron can
lead to conscious experience.
In an influential workshop report by Perkel and Bullock (1968) it was
concluded that one should not expect universal coding principle indepen-
dent of the context and an extensive list of possible neuronal codes was
given. Experimental evidence of codes which depend on precise timing
of individual spikes started to appear at that time (Segundo et al., 1963;
Chung et al., 1970; Richmond et al., 1987). Moreover, it was shown that
spiking mechanism can be very reliable (Bryant and Segundo, 1976; Mainen
and Sejnowski, 1995). Subsequently the term temporal coding was coined
for the situation when precise timing of spikes matters, however, a precise
definition is missing and the term may be used to refer to different concepts.
One important concept is that of synchronous firing across neurons at
the same time leading to the coherent firing of spatially distributed neu-
rons (Bialek et al., 1991), possibly connected to oscillations (Eckhorn et al.,
1988; Gray et al., 1989). Although a vast body of work focuses on the visual
system, this type of coding has been found both in auditory (deCharms and
Merzenich, 1996) and in olfactory systems (Wehr and Laurent, 1996; Lau-
rent et al., 1996; Perez-Orive et al., 2002). Compared to rate coding where
fundamental operation would be the temporal integration, basic operation
for this type of code would be coincidence detection for spikes coming from
different inputs. Such summation based on the activity of synchronized in-
puts would be more effective than code based on the firing rate (Singer and
Gray, 1995).
Another concept of temporal coding is represented by the notion that
specific time intervals between spikes may code some information, proposed
as early as in Lorente de Nó (1939), experimentally confirmed in Strehler
and Lestienne (1986); similar coding via temporal patterns has been found
as well (Eskandar et al., 1992). Yet another specific type of this code can
be based on the first spikes and its latencies (VanRullen and Thorpe, 2002).
There is also recent evidence of the so called sparse coding. It refers to
the idea that sensory information can be encoded by only a small number of
neurons within population or sparse activity of a single neuron (Olshausen

18
 3 NEURAL CODE

and Field, 2004). One notable example is from the primary auditory cortex
where DeWeese et al. (2003) showed that a neuron can produce a single
spike in a response to sound stimuli with very high reliability. Except ex-
perimental evidence in other sensory systems (Vinje and Gallant, 2000;
Perez-Orive et al., 2002), there is also a body of theoretical work showing
that sparse coding increases the capacity of associative memories, makes the
representation of signal easier to transmit and is efficient in terms of energy
consumption (Levy and Baxter, 1996). The relation between dense spatial-
temporal code and sparse code has been discussed in Theunissen (2003),
from another point of view (Földiák and Young, 1995) sparse coding can be
seen as a compromise between global activity of a whole neuron population
and single grandmother-like cell (Gross, 2002).
Another type of coding given by topographic position of a neuron is not
characterized by the type of neural firing activity but by the spatial position
of the active neuron in the brain tissue alone. An example of such “code” is
the mapping human anatomy in motor cortex (Penfield and Boldrey, 1937;
Nakamura et al., 1998) or representation of sound source spatial azimuth
in nucleus laminaris in birds (Carr and Konishi, 1988).
From the point of view of the whole network more codes can be used
simultaneously (Huxter et al., 2003) and even single spike train can encode
multiple features, for example Keat et al. (2001) shows how three differ-
ent features – what, when, how much, are assembled into waveform shape,
precise latency and firing rate of action potentials. Another example is our
proposal of a neural circuit computing sound azimuth in mammals (Sanda
and Marsalek, 2012) which uses a combination of topographic code (auditory
nerve fibers are sensitive to narrow range of frequencies), time coding (co-
incidence detection of appropriate spikes from ipsi- and contra-later fibers)
and rate code (capturing the final azimuth), details are depicted in section
3.3.1.
After introducing basic types of neural code we will focus on a particular
topic of binaural hearing, where most types of the aforementioned codes
occur in parallel.

19
 3 NEURAL CODE

3.2 Spatial hearing

In order to determine the direction of a sound source the neural circuit
needs to compute azimuth and elevation of the sound source given then
input signal. The important auditory cues are interaural time differences
(ITD3 ), interaural intensity differences (IID) and spectral content of the sig-
nal. While the visual information can severely modulate auditory responses
and is important in the development and calibration of the (auditory) space
representations, the neural circuit in auditory pathway responsible for spa-
tial hearing is capable of highly accurate localization predictions based on
the auditory cues only (King, 2009). Thus we will focus on the compu-
tational processing of auditory cues occurring in initial parts of auditory
pathway.
Historically two mechanisms of localization in the horizontal plane were
proposed by Békésy (von Békésy, 1930; van Bergeijk, 1962) and Jeffress
(1948). Von Békésy model assumes neurons in unspecified brain nucleus
on which fibers from left and right ear converge. The first arriving signal
from left/right (L/R) defines tuning of the neuron (channel) to be L or
R respectively. L/R signals traveling through the nucleus tune the whole
population of the neurons and higher centers integrate the number of L/R-
tuned neurons. In parallel more intensive stimuli from one side is able to
excite larger population of neurons tuned to that side. The final azimuth is
determined by the ratio of L/R tuned neurons.
Later and more widespread concept of delay lines (Jeffress, 1948) as-
sumes array of neurons each acting as a coincidence detector for the signal
from left and right side. The axonal fibers have systematically different
lengths so that additional time needed for the action potential to traverse
fiber from one side exactly compensates ITD. Thus each neuron is tuned to
narrow sector of the azimuthal space and the array of such neurons create
a whole topographical map of the azimuthal space (see Fig. 3).
The research on barn owls (Carr and Konishi, 1990) (and similarly later
3
Sound arrives at different times on the left and right ear. This difference defines ITD
and is dependent on the sound source position and the head size.

20
 3 NEURAL CODE

in chicken (Overholt et al., 1992; Köppl and Carr, 2008) and emu (MacLeod
et al., 2006)) convincingly showed that Jeffress delay lines are employed in
birds and although it is known that binaural hearing evolved independently
in different species (birds, reptiles and mammals in particular) the Jeffress
model became textbook model for binaural hearing. Last decade, however,
brought controversy over the mechanism of binaural hearing in mammals.

3.3 Neural code for spatial hearing of mammals

Experimenting and theorizing about the mechanism of (human) spatial
hearing dates back to the psychoacoustic works of Thompson (1882) and
Rayleigh (1907) who formalized the duplex theory. In this theory low fre-
quency sounds are localized by ITD while high frequency sounds by IID. In
general contours this distinction holds even nowadays, anatomical and phys-
iological findings confirmed this distinction on physiological and anatomical
level (Yin, 2002; Tollin, 2003). Additionally, research showed that spectral
cue is actively used for sound source elevation (Davis et al., 2003; Oertel
and Young, 2004). There seems to be agreement on the general mechanism
of IID processing (roughly speaking subtraction of excitatory signal from
ipsilateral side and inhibitory signal from contralateral side (Covey et al.,
1991)). The mechanisms of ITD processing particularly in low frequencies
is a matter of discussions and poses open problem upon which will we focus
from now on.

3.3.1 Low frequency ITD cues and modeling

While the anatomical evidence for delay lines in birds is solid, there is a
weak anatomical evidence for delay lines in mammals. Moreover there are
contradictions in physiological recordings on small mammals, which show
rather broadly tuned neurons (channels) and bring back the attention to
the Békésy concept (McAlpine et al., 2001; McAlpine and Grothe, 2003).4
4
Another distinct concept of cochlear traveling wave was proposed to account for the
existing experimental data, but we would not pursue any details here (Joris et al., 2006).

21
 3 NEURAL CODE

Figure 3: Different concepts explaining ITD sensitivity in spatial hearing.

The place code proposed by Jeffress consists of topographical ordered set
of neurons (top-right). Each neuron functions as coincidence detector and
receives ipsi- and contralateral axonal input. As we move along the array of
neurons the axonal delay from ipsilateral side gets longer and vice-versa for
the contralateral side. Because the neurons work as coincidence detectors
the difference in axonal length must exactly counter-balance ITD delay. In
consequence each ITD value activates (ideally) single neuron. On top-left
we can see the profile of neuronal firing (ITD tuning-curves) — for two dis-
tinct ITD values A1 , A2 two different neurons maximally fire at short range
around A1 /A2 respectively. On contrary the concept of rate code suggests
that neurons do have approximately similar broad response curves (bottom-
left) and thanks to specific excitation/inhibition mechanism (bottom-right,
we show mammalian anatomical wiring at MSO neuron) the firing response
of binaural neuron gradually changes as we move along different ITD values.
The difference between two azimuths A1 , A2 is represented by the different
firing rates of the neuron. Because critical encoding is employed in the slope
part of the tuning curve (in our scheme the part between the A1 and A2 ) it
is sometimes called also slope-encoding.

22
 3 NEURAL CODE

In series of experiments the key role of synaptic inhibition on the first

binaural neuron in MSO was elucidated (Brand et al., 2002; Grothe, 2003;
Pecka et al., 2008). The key finding is that synaptic inhibition on the
critical neuron shifts maximum of the broad ITD curve out of physiological
range relevant for the animal. It was suggested that azimuth is in this case
encoded in the slope part within physiological range of the ITD curve (see
Fig. 3) by the firing rate of binaural neuron. Marsalek and Lansky (2005)
took this concept and proposed stochastic model for the spike interaction
in this first binaural neuron.
We continued in this direction and simulated such stochastic circuit in
(Sanda and Marsalek, 2012). Our aim was twofold. First, to build circuit
representing the whole auditory pathway up to the first binaural neuron,
endow it with small number of parameters and then explore the parameter
space in order to understand the circuit dynamics. Second, to compare
the performance of such neuron with known results in psychophysics by
employing the concept of ideal observer located at the final stages of the
circuit — such observer measures time consumed by the circuit until reliable
azimuth estimate is reached.
There are currently two areas of spatial hearing research which usu-
ally do not coincide much. Either electro-physiological recording of critical
(MSO) neurons on small rodents which provides knowledge about the shape
of ITD tuning curves and let us theorize about the neural code used. Or
psychophysical experiments on human subjects, which provide us informa-
tion about spatial accuracy and measurements about minimal time needed
for solid azimuth estimate. Providing electro-physiological recording from
human subjects or psychological estimates of sound location from small ro-
dents is difficult task from obvious reasons. Our model connects these two
separated fields of research.
The first part of the model we proposed can be seen as formalization
of available anatomical and physiological data, while the second “observer”
part directly connects its result with psychophysics.
As a short summary the model is able to reproduce the shape and po-
sition of ITD tuning curves known from experiments as well as inhibition

23
 4 CONCLUSIONS

related results causing tuning curve shift. We identified the main parame-
ter responsible for this shift as coincidence window width. Next, we found
that certain amount of jitter actually improves efficiency of the circuit and
explored more thoroughly impact of jitter on time efficiency of the circuit in
Šanda (2011). Thanks to the observer module and data from psychophysics
we estimated value of minimal number of parallel circuits needed to repro-
duce psychoacoustic experiments.
The actual number of parallel fibers and their convergence on criti-
cal neurons is not exactly known and technical difficulties of physiological
recordings in MSO neurons do not provide sufficient experimental data to
decide the underlying mechanism even after decades of research. Mathe-
matical modeling can thus elucidate restrictions of suggested mechanisms
which can be checked by additional experiments.

4 Conclusions
It is widely believed that information processing in and between neurons is
mediated by action potentials (spikes) traveling along the neuronal mem-
brane. This process can be described at different levels — starting from
detailed biochemical models of membrane, continuing to its phenomeno-
logical models (integrate-and-fire models being the typical example) and
ending with very abstract models, in which only spike times are considered.
One particular description was chosen — stochastic LIF model — and
compared with in-vivo intracellular activity of neuron (such analysis has
not been done before, only either extracellular or in-vitro data are usually
available). We estimated parameters of the LIF model and tested in numer-
ical simulations (based on the estimated parameters) how model predictions
correspond to the real neuron. Additionally we characterized the difference
between spontaneous and acoustically stimulated behavior of the neuron.
To conclude, it was found that

• generally, the data are consistent with the model

• in the spontaneous part of the record:

24
 4 CONCLUSIONS

– asymptotic depolarization is below threshold and neuron is in

subthreshold noise-driven regimen
– ISIs distribution suggests Poisson process
– there are also inconsistencies between the data and model (or its
assumptions):
∗ spectra of the input signal should be flat, while we found one
dominant frequency
∗ the reset depolarization is not a fixed value
∗ experimental and modeled ISI distributions do not coincide
exactly

• in the stimulated part of the record:

– firing regimen is suprathreshold

– parameters representing input take larger values in stimulated
part of the recording
– experimental and modeled ISI distribution coincide
– inconsistencies with the model assumptions:
∗ reset depolarization is influenced by stimulation

As expected, we found that stimulation brings neuron into suprathreshold

regimen which causes the average ISI to be shorter than it is in the spon-
taneous neuronal firing. This observation brings us directly to the more
abstract level of description, where we analyze spike trains without consid-
ering exact membrane voltage between the spikes. In other words, we ask
how the external stimulus is encoded in the spike train emitted by neurons.
There are many such neuronal codes described in the literature and we
focused on the open problem of neural mechanisms responsible for spatial
hearing in mammals. Several theories (which are actually a mixture of
different codes) explaining the experimental findings were proposed and
we modeled a specific variant of so called slope-encoding model. Stochastic
neuronal circuit mimicking auditory pathway up to the first binaural neuron
was constructed. Considering this circuit we were able to:

25
 4 CONCLUSIONS

• reproduce of standard results found in experiments (ITD tuning curves)

• identify the role of parameters responsible for ITD timing shift (coin-
cidence window)

• describe parameter responsible for efficiency of coding (jitter)

• estimate minimal number of parallel circuits needed to reproduce re-

sults obtained psychoacoustic experiments on binaural hearing.

26
 REFERENCES

References
E.D. Adrian and Y. Zotterman. The impulses produced by sensory nerve-
endings: Part 2. The response of a single end-organ. Journal of Physiol-
ogy, 61:151–171, 1926.

H.B. Barlow. Single units and sensation: a neuron doctrine for perceptual
psychology. Perception, 1:371–394, 1972.

D.E. Bergles, J.D.B. Roberts, P. Somogyi, and C.E. Jahr. Glutamatergic

synapses on oligodendrocyte precursor cells in the hippocampus. Nature,
405:187–191, 2000.

W. Bialek, F. Rieke, R.R. De Ruyter Van Steveninck, and D. Warland.

Reading a neural code. Science, 252:1854–1857, 1991.

A. Brand, O. Behrend, T. Marquardt, D. McAlpine, and B. Grothe. Precise

inhibition is essential for microsecond interaural time difference coding.
Nature, 417:543–547, 2002.

K.H. Britten, M.N. Shadlen, W.T. Newsome, and J.A. Movshon. The anal-
ysis of visual motion: a comparison of neuronal and psychophysical per-
formance. Journal of Neuroscience, 12:4745–4765, 1992.

N. Brunel and M.C.W. van Rossum. Lapicque’s 1907 paper: from frogs to
integrate-and-fire. Biological Cybernetics, 97:337–339, 2007.

H.L. Bryant and J.P. Segundo. Spike initiation by transmembrane current:

a white-noise analysis. Journal of Physiology, 260:279–314, 1976.

T.H. Bullock. Neuron doctrine and electrophysiology. Science, 129:997–

1002, 1959.

T.H. Bullock, M.V.L. Bennett, D. Johnston, R. Josephson, E. Marder, and

R.D. Fields. The neuron doctrine, redux. Science, 310:791–793, 2005.

A.N. Burkitt. A review of the integrate-and-fire neuron model: I. Homoge-

neous synaptic input. Biological Cybernetics, 95:1–19, 2006a.

27
 REFERENCES

A.N. Burkitt. A review of the integrate-and-fire neuron model: II. Inhomo-

geneous synaptic input and network properties. Biological Cybernetics,
95:97–112, 2006b.

C. E. Carr and M. Konishi. Axonal delay lines for time measurement in the
owl’s brainstem. Proceedings of the National Academy of Sciences USA,
85:8311–8315, 1988.

C.E. Carr and M. Konishi. A circuit for detection of interaural time dif-
ferences in the brain stem of the barn owl. Journal of Neuroscience, 10:
3227–3246, 1990.

S. Celebrini and W.T. Newsome. Neuronal and psychophysical sensitivity to

motion signals in extrastriate area MST of the macaque monkey. Journal
of Neuroscience, 14:4109–4124, 1994.

S.H. Chung, S.A. Raymond, and J.Y. Lettvin. Multiple meaning in single
visual units. Brain, Behaviour and Evolution, 3:72–101, 1970.

E. Covey, M. Vater, and J.H. Casseday. Binaural properties of single units

in the superior olivary complex of the mustached bat. Journal of Neuro-
physiology, 66:1080–1094, 1991.

J.P. Cunningham, V. Gilja, S.I. Ryu, and K.V. Shenoy. Methods for estimat-
ing neural firing rates, and their application to brain-machine interfaces.
Neural Networks, 22:1235–1246, 2009.

K.A. Davis, R. Ramachandran, and B.J. May. Auditory processing of spec-

tral cues for sound localization in the inferior colliculus. Journal of the
Association for Research in Otolaryngology, 4:148–163, 2003.

R.C. deCharms and M.M. Merzenich. Primary cortical representation of

sounds by the coordination of action-potential timing. Nature, 381:610–
613, 1996.

R. Desimone, T.D. Albright, C.G. Gross, and C. Bruce. Stimulus-selective

properties of inferior temporal neurons in the macaque. Journal of Neu-
roscience, 4:2051–2062, 1984.

28
 REFERENCES

M.R. DeWeese, M. Wehr, and A.M. Zador. Binary spiking in auditory

cortex. Journal of Neuroscience, 23:7940–7949, 2003.

R. Eckhorn, R. Bauer, W. Jordan, M. Brosch, W. Kruse, M. Munk, and

H.J. Reitboeck. Coherent oscillations: A mechanism of feature linking in
the visual cortex? Biological Cybernetics, 60:121–130, 1988.

E.N. Eskandar, B.J. Richmond, and L.M. Optican. Role of inferior tempo-
ral neurons in visual memory. I. Temporal encoding of information about
visual images, recalled images, and behavioral context. Journal of Neu-
rophysiology, 68:1277–1295, 1992.

J. Feng. Is the integrate-and-fire model good enough? – a review. Neural

Networks, 14:955–975, 2001.

R.D. Fields and B. Stevens-Graham. New insights into neuron-glia commu-

nication. Science, 298:556–562, 2002.

R. Fitzhugh. Impulses and physiological states in theoretical models of

nerve membrane. Biophysical Journal, 1:445–466, 1961.

P. Földiák and M.P. Young. Sparse coding in the primate cortex. In Arbib
M.A., editor, The handbook of brain theory and neural networks, pages
895–898. MIT Press, Cambridge, MA, 1995.

M.G. Fuortes. Electric activity of cells in the eye of Limulus. American

Journal of Ophthalmology, 46:210–223, 1958.

G.L. Gerstein and N.Y.S. Kiang. An approach to the quantitative analysis

of electrophysiological data from single neurons. Biophysical Journal, 1:
15–28, 1960.

G.L. Gerstein and B. Mandelbrot. Random walk models for the spike ac-
tivity of a single neuron. Biophysical Journal, 4:41–68, 1964.

C.M. Gray, P. König, A.K. Engel, and W. Singer. Oscillatory responses

in cat visual cortex exhibit inter-columnar synchronization which reflects
global stimulus properties. Nature, 338:334–337, 1989.

29
 REFERENCES

C.G. Gross. Genealogy of the "Grandmother Cell". The Neuroscientist, 8:

512–518, 2002.

B. Grothe. New roles for synaptic inhibition in sound localization. Nature

reviews neuroscience, 4:540–550, 2003.

O.J. Grüsser, U. Grüsser-Cornehls, and MD Licker. Further studies on the

velocity function of movement detecting class-2 neurons in the frog retina.
Vision Research, 8:1173–1185, 1968.

A. Hjelmfelt and J. Ross. Chemical implementation and thermodynamics

of collective neural networks. Proceedings of the National Academy of
Sciences USA, 89:388–391, 1992.

A. Hjelmfelt and J. Ross. Mass-coupled chemical systems with computa-

tional properties. Journal of Physical Chemistry, 97:7988–7992, 1993.

A. Hjelmfelt, E.D. Weinberger, and J. Ross. Chemical implementation

of neural networks and Turing machines. Proceedings of the National
Academy of Sciences USA, 88:10983–10987, 1991.

A.L. Hodgkin and A.F. Huxley. A quantitative description of membrane

current and its application to conduction and excitation in nerve. Journal
of Physiology, 117:500–544, 1952.

J.J. Hopfield. Neural networks and physical systems with emergent col-
lective computational abilities. Proceedings of the National Academy of
Sciences USA, 79:2554–2558, 1982.

J.J. Hopfield. Neurons with graded response have collective computational

properties like those of two-state neurons. Proceedings of the National
Academy of Sciences USA, 81:3088–3092, 1984.

J. Huxter, N. Burgess, and J. O’Keefe. Independent rate and temporal

coding in hippocampal pyramidal cells. Nature, 425:828–832, 2003.

L.A. Jeffress. A place theory of sound localization. Journal of Comparative

and Physiological Psychology, 41:35–39, 1948.

30
 REFERENCES

R. Jolivet, A. Rauch, HR Lüscher, and W. Gerstner. Integrate-and-Fire

models with adaptation are good enough: predicting spike times under
random current injection. Advances in Neural Information Processing
Systems, 18:595–602, 2006.

R. Jolivet, R. Kobayashi, A. Rauch, R. Naud, S. Shinomoto, and W. Ger-

stner. A benchmark test for a quantitative assessment of simple neuron
models. Journal of Neuroscience Methods, 169:417–424, 2008a.

R. Jolivet, F. Schürmann, T.K. Berger, R. Naud, W. Gerstner, and A. Roth.

The quantitative single-neuron modeling competition. Biological cyber-
netics, 99:417–426, 2008b.

P.X. Joris, B. Van de Sande, D.H. Louage, and M. Van Der Heijden. Bin-
aural and cochlear disparities. Proceedings of the National Academy of
Sciences USA, 103:12917–12922, 2006.

J. Keat, P. Reinagel, R.C. Reid, and M. Meister. Predicting every spike: A

model for the responses of visual neurons. Neuron, 30:803–817, 2001.

A.J. King. Visual influences on auditory spatial learning. Philosophical

Transactions of the Royal Society B: Biological Sciences, 364:331–339,
2009.

W.M. Kistler, W. Gerstner, and J.L. Hemmen. Reduction of the Hodgkin-

Huxley equations to a single-variable threshold model. Neural Computa-
tion, 9:1015–1045, 1997.

B.W. Knight. Dynamics of encoding in a population of neurons. Journal of

General Physiology, 59:734–766, 1972.

R. Kobayashi, Y. Tsubo, and S. Shinomoto. Made-to-order spiking neuron

model equipped with a multi-timescale adaptive threshold. Frontiers in
Computational Neuroscience, 3, 2009.

C. Koch. Biophysics of computation: information processing in single neu-

rons. Oxford University Press, USA, 2005.

31
 REFERENCES

C. Koch, T. Poggio, and V. Torres. Retinal ganglion cells: a functional

interpretation of dendritic morphology. Philosophical Transactions of the
Royal Society of London. B, Biological Sciences, 298:227–263, 1982.

C. Koch, T. Poggio, and V. Torre. Nonlinear interactions in a dendritic

tree: localization, timing, and role in information processing. Proceedings
of the National Academy of Sciences USA, 80:2799–2802, 1983.

C. Köppl and C.E. Carr. Maps of interaural time difference in the chicken’s
brainstem nucleus laminaris. Biological cybernetics, 98:541–559, 2008.

P. Lansky, P. Sanda, and J. He. The parameters of the stochastic leaky

integrate-and-fire neuronal model. Journal of Computational Neuro-
science, 21:211–223, 2006.

P. Lansky, P. Sanda, and J. He. Effect of stimulation on the input param-

eters of stochastic leaky integrate-and-fire neuronal model. Journal of
Physiology-Paris, 104:160–166, 2010.

L. Lapicque. Recherches quantitatives sur l’excitation électrique des nerfs

traitée comme une polarisation. Journal de Physiologie et de Pathologie
Générale, 9:620–635, 1907.

L. Lapicque. Quantitative investigations of electrical nerve excitation

treated as polarization. Biological Cybernetics, 97:341–349, 2007.

G. Laurent, M. Wehr, and H. Davidowitz. Temporal representations of

odors in an olfactory network. Journal of Neuroscience, 16:3837–3847,
1996.

W.B. Levy and R.A. Baxter. Energy efficient neural codes. Neural Compu-
tation, 8:531–543, 1996.

R. Lorente de Nó. Transmission of impulses through cranial nerve nuclei.

Journal of Neurophysiology, 2:402–464, 1939.

32
 REFERENCES

K.M. MacLeod, D. Soares, and C.E. Carr. Interaural timing difference cir-
cuits in the auditory brainstem of the emu (Dromaius novaehollandiae).
Journal of comparative neurology, 495:185–201, 2006.

Z.F. Mainen and T.J. Sejnowski. Reliability of spike timing in neocortical

neurons. Science, 268:1503–1506, 1995.

P. Marsalek and P. Lansky. Proposed mechanisms for coincidence detection

in the auditory brainstem. Biological Cybernetics, 92:445–451, 2005.

P.B. Matthews. Muscle spindles and their motor control. Physiological

Reviews, 44:219–288, 1964.

D. McAlpine and B. Grothe. Sound localization and delay lines - do mam-

mals fit the model? Trends in Neurosciences, 26:347–350, 2003.

D. McAlpine, D. Jiang, and A.R. Palmer. A neural code for low-frequency

sound localization in mammals. Nature Neuroscience, 4:396–401, 2001.

W.S. McCulloch. Agatha Tyche of nervous nets-the lucky reckoners. In

Mechanization of Thought Processes: Proceedings of a Symposium held
at the National Physical Laboratory, volume 2, pages 611–634. HMSO,
London, 1959.

W.S. McCulloch and W. Pitts. A logical calculus of the ideas immanent in

nervous activity. Bulletin of Mathematical Biology, 5:115–133, 1943.

J. Nagumo, S. Arimoto, and S. Yoshizawa. An active pulse transmission

line simulating nerve axon. Proceedings of the IRE, 50:2061–2070, 1962.

A. Nakamura, T. Yamada, A. Goto, T. Kato, K. Ito, Y. Abe, T. Kachi, and

R. Kakigi. Somatosensory homunculus as drawn by MEG. Neuroimage,
7:377–386, 1998.

M. Nawrot, A. Aertsen, and S. Rotter. Single-trial estimation of neuronal

firing rates: from single-neuron spike trains to population activity. Jour-
nal of Neuroscience Methods, 94:81–92, 1999.

33
 REFERENCES

D. Oertel and E.D. Young. What’s a cerebellar circuit doing in the auditory
system? Trends in Neurosciences, 27:104–110, 2004.

B.A. Olshausen and D.J. Field. Sparse coding of sensory inputs. Current
Opinion in Neurobiology, 14:481–487, 2004.

E.M. Overholt, E.W. Rubel, and R.L. Hyson. A circuit for coding interaural
time differences in the chick brainstem. Journal of neuroscience, 12:1698–
1708, 1992.

M. Pecka, A. Brand, O. Behrend, and B. Grothe. Interaural time difference

processing in the mammalian medial superior olive: the role of glycinergic
inhibition. Journal of Neuroscience, 28:6914–6925, 2008.

W. Penfield and E. Boldrey. Somatic motor and sensory representation in

the cerebral cortex of man as studied by electrical stimulation. Brain, 60:
389–443, 1937.

J. Perez-Orive, O. Mazor, G.C. Turner, S. Cassenaer, R.I. Wilson, and

G. Laurent. Oscillations and sparsening of odor representations in the
mushroom body. Science, 297:359–365, 2002.

D.H. Perkel and T.H. Bullock. Neural coding. Neurosciences Research

Program Bulletin, 6:221–350, 1968.

S. Ramon y Cajal. Textura del sistema nervioso del hombre y los vertebrados.
Imprenta y Librería de Nicolás Moya, Madrid, 1899.

S. Ramon y Cajal. Histology of the nervous system of man and vertebrates.

Oxford University Press, NY, 1995.

L. Rayleigh. On our perception of sound direction. Philosophical Magazine,

13:214–232, 1907.

L.M. Ricciardi and L. Sacerdote. The Ornstein-Uhlenbeck process as a

model for neuronal activity. Biological Cybernetics, 35:1–9, 1979.

34
 REFERENCES

B.J. Richmond, L.M. Optican, M. Podell, and H. Spitzer. Temporal en-

coding of two-dimensional patterns by single units in primate inferior
temporal cortex. I. Response characteristics. Journal of Neurophysiology,
57:132–146, 1987.

B. Sakitt. Counting every quantum. Journal of Physiology, 223:131–150,

1972.

P. Šanda. Speeding up the algorithm for finding optimal kernel bandwidth

in spike train analysis. European Journal for Biomedical Informatics, 6:
73–75, 2010.

P. Šanda. Jitter effect on the performance of the sound localization model

of medial superior olive neural circuit. European Journal for Biomedical
Informatics, 7:51–54, 2011.

P. Sanda and P. Marsalek. Stochastic interpolation model of the medial

superior olive neural circuit. Brain Research, 1434:257–265, 2012.

I. Segev. Single neurone models: oversimple, complex and reduced. Trends

in Neurosciences, 15:414–421, 1992.

J.P. Segundo, G.P. Moore, L.J. Stensaas, and T.H Bullock. Sensitivity of
neurones in Aplysia to temporal pattern of arriving impulses. Journal of
Experimental Biology, 40:643–667, 1963.

H. Shimazaki and S. Shinomoto. A method for selecting the bin size of a

time histogram. Neural Computation, 19:1503–1527, 2007.

H. Shimazaki and S. Shinomoto. Kernel bandwidth optimization in spike

rate estimation. Journal of Computational Neuroscience, 29:171–182,
2010.

W. Singer and C.M. Gray. Visual feature integration and the temporal
correlation hypothesis. Annual Review of Neuroscience, 18:555–586, 1995.

R.B. Stein. A theoretical analysis of neuronal variability. Biophysical Jour-

nal, 5:173–194, 1965.

35
 REFERENCES

R.B. Stein. Some models of neuronal variability. Biophysical Journal, 7:

37–68, 1967.

B.L. Strehler and R. Lestienne. Evidence on precise time-coded symbols and

memory of patterns in monkey cortical neuronal spike trains. Proceedings
of the National Academy of Sciences USA, 83:9812–9816, 1986.

F.E. Theunissen. From synchrony to sparseness. Trends in Neurosciences,

26:61–64, 2003.

S.P. Thompson. On the function of the two ears in the perception of space.
Philosophical Magazine, 13:406–416, 1882.

D.J. Tollin. The lateral superior olive: A functional role in sound source
localization. Neuroscientist, 9:127–143, 2003.

G.E. Uhlenbeck and L.S. Ornstein. On the theory of the Brownian motion.
Physical Review, 36:823–841, 1930.

W.A. van Bergeijk. Variation on a theme of Bekesy: a model of binaural

interaction. Journal of the Acoustical Society of America, 34:1431–1437,
1962.

R. VanRullen and S.J. Thorpe. Surfing a spike wave down the ventral
stream. Vision Research, 42:2593–2615, 2002.

W.E. Vinje and J.L. Gallant. Sparse coding and decorrelation in primary
visual cortex during natural vision. Science, 287:1273–1276, 2000.

G. von Békésy. Zur theorie des hörens. über das Richtungshören bei einer
Zeitdefferenz oder Lautstärkenungleichheit der beiderseitigen Schallein-
wirkungen. Physikalische Zeitschrift, pages 824–835, 1930.

M. Wehr and G. Laurent. Odour encoding by temporal sequences of firing

in oscillating neural assemblies. Nature, 384:162–166, 1996.

36
 REFERENCES

T. Yin. Neural mechanisms of encoding binaural localization cues in the

auditory brainstem. In Fay R.R. Oertel D., Popper A.N., editor, Integra-
tive functions in the mammalizan auditory pathway, pages 99–159. New
York: Springer-Verlag, 2002.

37
List of publications
Reviewed journals with impact factor
1. Lansky, P. and Sanda, P. and He, J., The parameters of the stochastic
leaky integrate-and-fire neuronal model, Journal of Computational
Neuroscience, 21:211–223, 2006. (Journal IF: 2.325)

2. Lansky, P. and Sanda, P. and Weiss, M., Modeling the influence

of non-adherence on antibiotic efficacy: application to ciprofloxacin,
The International Journal of Clinical Pharmacology and Therapeu-
tics, 45:438–447, 2007. (Journal IF: 1.189)

3. Lansky P. and Sanda, P. and He J., Effect of stimulation on the in-

put parameters of stochastic leaky integrate-and-fire neuronal model,
Journal of Physiology - Paris, 104:160–166, 2010. (Journal IF: 3.030)

4. Sanda P. and Marsalek P., Stochastic interpolation model of the me-

dial superior olive neural circuit, Brain Research, Brain Research,
1434:257–265, 2012. (Journal IF: 2.623)

Reviewed journals without impact factor

5. Šanda P., Speeding up the Algorithm for Finding Optimal Kernel
Bandwidth in Spike Train Analysis, European Journal for Biomedical
Informatics, 6:73–75, 2010.

6. Šanda P., Jitter Effect on the Performance of the Sound Localization

Model of Medial Superior Olive Neural Circuit, European Journal for
Biomedical Informatics, 7:51–54, 2011.

38
Reprints of published papers
The research papers 1, 3-6 represent the main part of the thesis, the preced-
ing text gives extended introduction and context for the whole work. The
research paper 2 is out of the thesis scope and is not included in the thesis.
The reprints of the papers are attached below.

39
J Comput Neurosci
DOI 10.1007/s10827-006-8527-6

The parameters of the stochastic leaky integrate-and-fire neuronal

model
Petr Lansky · Pavel Sanda · Jufang He

Received: 24 November 2005 / Revised: 23 March 2006 / Accepted: 28 March 2006 / Published online: 28 July 2006


C Springer Science + Business Media, LLC 2006

Abstract Five parameters of one of the most common neu- characterized by the Poissonian firing. This is in a complete
ronal models, the diffusion leaky integrate-and-fire model, agreement with the observed interspike interval data.
also known as the Ornstein-Uhlenbeck neuronal model, were
estimated on the basis of intracellular recording. These pa- Keywords Leaky integrate-and-fire model .
rameters can be classified into two categories. Three of them Ornstein-Uhlenbeck neuronal model . Parameters
(the membrane time constant, the resting potential and the fir- estimation . Spontaneous firing
ing threshold) characterize the neuron itself. The remaining
two characterize the neuronal input. The intracellular data
were collected during spontaneous firing, which in this case Introduction
is characterized by a Poisson process of interspike intervals.
Two methods for the estimation were applied, the regression Application of mathematical methods in neuroscience is
method and the maximum-likelihood method. Both methods based on construction of models aiming to mimic real ob-
permit to estimate the input parameters and the membrane jects. The models range from phenomenological mathemat-
time constant in a short time window (a single interspike in- ical models to very detailed biophysical models. From a
terval). We found that, at least in our example, the regression biophysical point of view, the models of a single neuron
method gave more consistent results than the maximum- reflect the electrical properties of its membrane via elec-
likelihood method. The estimates of the input parameters tric circuit description. Such circuit models can be written
show the asymptotical normality, which can be further used in terms of differential equations for the membrane volt-
for statistical testing, under the condition that the data are col- age. Reducing these models, we can obtain integrate-and-
lected in different experimental situations. The model neu- fire types of model, which are reviewed in detail in most
ron, as deduced from the determined parameters, works in computational neuroscience monographs (Tuckwell, 1988;
a subthreshold regimen. This result was confirmed by both Koch, 1998; Dayan and Abbot, 2001; Gerstner and Kistler,
applied methods. The subthreshold regimen for this model is 2002). These models are sometimes criticized for their too
drastic simplification of reality (e.g., Segev, 1992). Simulta-
neously, the opposite opinion appears. For example, Kistler
Action Editor: Nicolas Brunel
et al. (1997) claim that the integrate-and-fire model with a
P. Lansky (
) · P. Sanda properly selected threshold, after reduction of the Hodgkin-
Institute of Physiology,
Huxley four dimensional model, predicts 90 percent of the
Academy of Sciences of the Czech Republic,
Prague, Czech Republic spikes correctly. Independently from this discussion, we ob-
e-mail: [email protected] serve that the number of papers devoted to the integrate-and-
fire model, or at least employing it, is very high.
J. He
The simplest “realistic” neuronal model is the determin-
Department of Rehabilitation Sciences,
The Hong Kong Polytechnic University, istic leaky integrate-and-fire model (Lapicque model, RC-
Hung Hom, Kowloon, Hong Kong circuit). It assumes that the membrane depolarization can

Springer
 J Comput Neurosci

be described by a circuit with a generator, a resistor and a in vivo conditions with known input to the system. Estima-
capacitor in parallel. It has to be stressed that while the elec- tion methods from in vitro voltage recordings for known in-
trical representation is related to a small isopotential patch of put were presented by Stevens and Zador (1998), Rauch et al.
neuronal membrane, the mathematical variable (the voltage) (2003), Le Camera et al. (2004), Jolivet et al. (2006), Paninski
reflects an abstract representation of a complete neuron. This et al. (2004). None of these papers treats comparison of the
is another simplification based on neglecting the spatial prop- Ornstein-Uhlenbeck model with in vivo spontaneous activ-
erties of a neuron. There are attempts to overcome this situ- ity. The likely reason is that, using the model, only interspike
ation (e.g., Pinsky and Rinzel, 1994; Rodriguez and Lansky, intervals (ISIs) were usually predicted and thus the attempts
2000) but still the single-point models dominate most of the to identify the model parameters were based on observation
applications. Due to the simplicity of the deterministic leaky of ISIs. Such a task is enormously complicated and leads to
integrate-and-fire model, the action potential generation is rather difficult numerical and mathematical problems (Inoue
not an inherent part of the model as in more complex models et al., 1995; Shinomoto et al., 1999; Ditlevsen and Lansky,
and a firing threshold has to be imposed. The model neuron 2005).
fires whenever the threshold is reached and then the voltage We aimed to study the estimation methods in the Ornstein-
is reset to its initial value. This means that in the electrical cir- Uhlenbeck model, their stability and reproducibility. In the
cuit representation a switch is added to the circuit. The reset first Section we summarize the properties of the model. Then
following the threshold crossing introduces a strong nonlin- the methods for the estimation of its parameters are given
earity into the model. For a constant input the model neu- and details of data acquisition presented. Simultaneously,
ron remains silent, never reaching the threshold (subthresh- the assumptions of the model are tested. Finally the param-
old regimen), or fires at constant intervals (suprathreshold eters of the model are estimated and the obtained results
regimen). are discussed. We restricted the study on a single neuron
The experimental data recorded from very different neu- under spontaneous activity conditions. To extend the results
ronal structures and under different experimental conditions on several neurons and different experimental conditions is
suggest a presence of stochastic variables in neuronal activ- possible, but beyond the scope of this article.
ity. We may assume that there is a random component, gener-
ally regarded as noise, contained in the incoming signal. The
other source of noise can be the neuron itself where a random
component is added to the signal. Unfortunately, there is no Model and its properties
clear distinction between noise contained in the signal and
the system noise. A phenomenological way how to introduce The Ornstein-Uhlenbeck model of membrane depolarization
stochasticity into the deterministic leaky integrate-and-fire is formally given by the stochastic differential equation,
model is simply by assuming an additional noise term. If
the noise is not further specified, but assumed to be Gaussian dX(t) = (−β(X (t) − x0 ) + µ)dt + σ dW(t) , X (0) = x0 , (1)
and white, then the model is well known in physical literature
as an Ornstein-Uhlenbeck model (e.g., Gardiner, 1982) and where dW represents increments of a standard stochastic
this model has been widely used in neuroscience literature Wiener process (Brownian motion), and β > 0 characterizes
(Tuckwell, 1988; Koch, 1998; Dayan and Abbot, 2001; Ger- the spontaneous decay of the membrane depolarization in
stner and Kistler, 2002). An alternative way to end up with the absence of input to the resting level x0 . The drift coeffi-
the Ornstein-Uhlenbeck model is by diffusion approximation cient µ reflects the local average rate of displacement due to
of the model with discontinuous trajectories (Stein, 1965). the neuronal input and local variability is represented by the
An advantage of this approach is that a direct interpreta- infinitesimal variance σ (the variability of the neuronal in-
tion of the parameters appearing in the Ornstein-Uhlenbeck put). The spikes are not an intrinsic part of the model but are
neuronal model is available (Lansky, 1997). generated when the membrane depolarization X(t) reaches
Models without specified parameters remain only a tool for the first time the firing threshold S, which is an additional
for qualitative comparison and thus finding methods for esti- parameter. Then, the depolarization is reset to the resting
mation is equally important as model construction. The lack level, x0 , and the process of input “integration” starts anew.
of methods for parameters identification had been noticed for We should keep in mind that also the reset level, x0 , repre-
a long period (e.g., Tuckwell and Richter, 1978; Brillinger sents an additional parameter of the model. Thus the model
and Segundo, 1979). In general, the traditional approaches is fully described by Eq. (1) with its five parameters: β, µ, σ ,
were more frequently focused on interspike interval (ISI) S and x0 . As said, the ISIs are identified in model (1) with the
distribution. Keat et al. (2001) as well as Paninski et al. first-passage times of the process X(t) across the boundary
(2004) developed methods based on extracellular recordings S,

Springer
J Comput Neurosci

T = inf (t > 0, X (t) ≥ S > x0 ) . (2) where h denotes the time step of simulation, Xi (i = 1,2,
. . .) are the simulated values of the process, and  i are in-
Due to the complete reset in defining ISI by Eq. (2) and dependent and normally distributed random variables, √ i∼
due to the constant input µ, the ISIs form a renewal process, N(0, h). The increments  i in (6) can be replaced by ± h
which means that ISIs are independent and identically dis- selecting these values with equal probability 1 /2 , which
tributed random variables. Formula (1) can be rewritten in a substantially decreases the simulation time (Tuckwell and
form often seen in engineering applications using the term Lansky, 1997). This was the procedure applied to simu-
white noise, late the membrane depolarization in this study. Apparently,
the parameters β, σ , µ and x0 have to be determined for
dX(t) the simulation procedure. If the ISIs are to be simulated,
= −β(X (t) − x0 ) + µ + σ ξ (t) , X (0) = x0 , (3)
dt then in addition, the firing threshold S is required. As men-
tioned, the spikes in the model are generated when the mem-
with the same interpretation of the parameters as above,
brane depolarization X(t) reaches, for the first time, the fir-
only the white noise ξ (t) is a formal derivative of the Wiener
ing threshold S. While the simulation of the trajectories X
process with respect to time. For a fixed time t, X(t) given
contains no systematic bias, it is not true for the simula-
by (1) or (3) is a Gaussian random variable. In absence of
tion of the first passage times (Lansky and Lanska, 1994).
the threshold S and if σ tends to zero, we can solve the
It is systematically overestimated and this effect has to be
differential Eq. (3). The solution is identical with the mean
minimized.
value of the stochastic depolarization given by Eq. (1)
Two basic types of data can be used for the identification
µ of the parameters appearing in Eq. (1). In the first of them
E(X (t)) = x0 + (1 − exp(−βt)) (4) only the ISIs are available, which means the realizations of
β
the random variable defined by Eq. (2). If this is the case, then
and the variance of X(t) is the situation is complicated and the solution can be achieved
only under some additional assumptions. For example, it has
σ2 to be assumed that the firing threshold and the resting level
Var(X (t)) = (1 − exp(−2βt)) . (5)
2β are known. In the second situation, which is investigated
here, the membrane depolarization is recorded between the
The position of the asymptotic depolarization E(X( ∞ )) generation of spikes. To specify the firing threshold and reset
= x0 + µ/β determines regimes of firing of the Ornstein- level seems to be a simpler task than to estimate the remaining
Uhlenbeck model. For µ/β  S − x0 , the suprathreshold parameters of the model. We should simply record what was
regimen, the firing is almost regular and ISI histogram re- the reset after the end of an action potential and what was the
sembles normal distribution. The noise plays a limited role in final value of the depolarization when it started. However, we
this range of parameters. For µ/β ≈ S − x0 , the distribution will see that the situation is not so simple and also these two
of ISIs is positively skewed and resembles Gamma distribu- parameters need to be estimated. A method of estimating the
tion. In the subthreshold regimen, µ/β  S − x0 the firing remaining parameters was proposed more than two decades
becomes Poissonian. Here, the noise plays a crucial role and ago (Lansky, 1983). Thus the novelty of this paper is mainly
without it the neuron would remain silent. This last regimen in application of the method to real intracellular data. For an
is important for this study, as will be seen. Of course, the extensive methodological review of estimation methods in
signs “  ” and “  ” are relative to the asymptotic vari- stochastic diffusion processes, for which Eq. (1) is a special
ance Var(X( ∞ )) = σ 2 /2β. More details on the Ornstein- case, see Prakasa Rao (1999).
Uhlenbeck neuronal model can be found, for example, in The aim of this article is primarily determination of the
Tuckwell (1988) or in Ricciardi and Lansky (2003). values of the parameters β, σ and µ. The question is whether
The description of the process via Eq. (3) is apparently an these parameters are stable over a long period or whether
intuitive extension of the deterministic approach. Its advan- they vary in short time ranges. Whereas σ and µ are input
tage is in giving a method for a computer simulation of the parameters and thus are assumed to change whenever the
process sample trajectories (Kloeden and Platen, 1992). The input to a neuron has changed, model (1) assumes that β is
simplest discrete-time approximation of (3) is a stochastic a property of the membrane (in the same way as S and the
analogue of the Euler scheme for ordinary differential equa- reset level) and these three intrinsic parameters should be sta-
tions, ble. However, these are only assumptions which have never
been confirmed. Thus, initially we estimate the parameters
X i+1 = X i − β(X i − x0 )h + µh + σ i , X 0 = x0 , (6) separately for each ISI.

Springer
 J Comput Neurosci

Methods (b) Estimation from several interspike intervals

(a) Estimation from a single interspike interval In this situation, if we assume that the parameters remain
stable over several ISIs we can use the extension of three
The records of the depolarization within single ISI permit estimates as were formally proposed by Lansky (1983). That
us to estimate β, µ and σ . Theoretically also two additional method takes into account the length of ISIs and in some
parameters S and x0 could be determined, but as we will see, sense shorter ISIs contribute to the estimates less than longer
for that purpose more realizations of the ISIs are necessary. ones. Here we use a slightly different approach. We estimate
Let us assume that in one ISI the membrane depolarization the parameters for each ISI separately. Then, the global esti-
Xi = xi is sampled at N + 1 points (i = 0, . . ., N) at steps mates are representative over ISI counts not the total length
h at times ti = ih (the notation is complicated for non equal of the record. Further, and it is the main reason, in this way
sampling step but the results are analogous). Then the for- we also get some information about the dependency of the
mulas for the estimation of the parameters by the maximum values on the lengths of ISI and their position in the record.
likelihood method are The global record can be characterized by representative val-

N −1
N −1 ues of β M , µM and σ M , in our case we use medians (denoted
1 j=0 x 2j − j=0 x j+1 x j + (x N − x0 )x by index M) of the estimated values.
β̂ =
N −1 , (7)
If several ISIs are available, then in addition to the param-
j=0 x j + x N
h 2 2

eters mentioned in the previous Section, also the threshold S

and the reset value x0 can be estimated. For this purpose we
x N − x0 simply use the medians of the values observed for each ISI.
µ̂ = + β̂ x̄ (8)
T
Animal preparation for the intracellular recordings
and
Guinea pigs served as subjects for the intracellular recording
N −1
1  experiments. Anaesthesia was initially induced with pento-
σ̂ = (x j+1 − x j + x j h β̂ − h µ̂)2 , (9)
T j=0 barbital sodium (Nembutal, Abott, 35 mg/kg, ip) and main-
tained by supplemental doses of the same anaesthetic (about

N 5–10 mg/kg/hr) during the surgical preparation and record-
where x = N1 j=0 x j , T = Nh. These formulas are
ing. Throughout the recording, an electrocorticograph was
discrete-time variants of the formulas based on the assump-
monitored to assess the level of anaesthesia. The subject was
tion that the depolarization is continuously recorded in be-
mounted in a stereotaxic device following the induction of
tween the spikes.
anaesthesia. A midline incision was made in the scalp and
Formula (4) suggests that the method of moments can also
a craniotomy was performed to enable vertical access to the
be used. Then, we minimize the functional
MGB in the right hemisphere (He, 2003; Xiong et al., 2003;
N Yu et al., 2004). The head was fixed with two stainless steel
µ
L(β, µ) = (x j − x0 − (1 − exp(−β j h)))2 (10) bolts to an extended arm from the stereotaxic frame using
j=1
β acrylic resin. The left ear was then freed from the ear bar,
so that the subject’s head remained fixed to the stereotaxic
with respect to the parameters β and µ by a regression device without movement.
method. It is obvious that efficiency of this method depends Cerebrospinal fluid was released at the medulla level
on the distance of x0 from the asymptotic depolarization through an opening at the back of the neck. The animal was
µ/β. An increase of β relatively to h also handicaps the artificially ventilated. Both sides of the animal’s chest were
method. opened, and its body was suspended to reduce vibrations of
Another method for estimate of the noise amplitude is the brain caused by intra-thoracic pressure. The experimen-
tal procedures were approved by the Animal Subjects Ethics
N −1
1  Sub-Committee of The Hong Kong Polytechnic University.
σ̂ = (x j+1 − x j )2 . (11) A glass-pipette as the recording electrode, filling it with
T j=0
0.5M KCl (pH 7.6, 0.05M Tris HCl buffer) was used. The
resistance of the electrode ranged between 40–90 M . The
This estimate follows from theoretical results established electrode was advanced vertically from the top of the brain
by Feigin (1976). Comparing Eqs. (9) and (11), we can see by a stepping motor (Narishige, Tokyo, Japan). After the
that for h → 0 in (9) we end up with Eq. (11). We will electrode was lowered to a depth of 4–5 mm, the cortical
compare all these estimation methods. exposure was sealed using low-melting temperature paraffin.

Springer
J Comput Neurosci

When the electrode was near or in the target area, it was

slowly advanced at 1 or 2 µm per step.

Data collection

Upon penetrating the membrane of a cell, the electrode de-

tected the negative membrane potential. After amplification,
the membrane potential as well as the auditory stimulus
were stored in the computer with the aid of commercial
software (AxoScope, Axon). The direct current (DC) level
of the recording electrode was frequently checked and set
to zero during the experiments. The DC level after each
recording was used to compensate for the membrane poten-
tial of some neurons. Neurons showing a resting membrane
potential lower than −50 mV and spontaneous spikes (if
any) of larger than 50 mV were included in the present
study. Single neuron data were selected for this article.
The membrane potential was recorded (in 100 mV) with
time step h = 0.00015 [s] = 0.15 [ms], for period 0−501
[s]. Accompanying the values of the membrane potential
is the stimulus level. For the purpose of this study we se-
lected only ISIs which were entirely outside the stimulation
period.

Detection of spikes and determination of S and x0

The parameter estimation method is based on the observation

of the membrane depolarization between spikes. Therefore Fig. 1 Example of data used for the estimation of the parameters. (a)
The spikes are not initiated at the same values and the end is not uniquely
the spikes have to be removed from the data but it is not
defined. (b) Schematic illustration of ISI initial point detection (for
entirely obvious which part of the records can be included in details see text). Gray line is the moving average over six observations
the estimation procedure. At first, we detect the spikes and given by black line
then we judge their beginnings and their ends. In this way
ISIs are fixed. The least problematic is spike detection. The
Its start is fixed at the moment when the depolarization
level for this purpose was experimentally chosen at the value
reaches the value of −65.5 [mV] for the first time after
of −35.5 [mV] (note that this is not the firing threshold S,
spike generation. Its end is the time point 0.01005 [s] after
but a value to detect spikes in continuous sampling of the
its beginning. In this region the minimum depolarization is
voltage), see Fig. 1(a).
sought for (see Fig. 1(b)).
From visual inspection of the data it is clearly difficult
Defining the end of each ISI was not so problematic and
to decide where exactly to start and to end the spikes, and
we took the point 0.01005 [s] before detected spike, i.e., be-
hence to decide which data to include in the parameter es-
fore the voltage reaches −35.5 [mV]. For threshold determi-
timation procedure. It follows from this inspection that for
nation, we took the last point with decreasing depolarization
detected spikes the width of the spike as well as the voltage
before the spike detection (in other words, the depolarization
where ISI starts, are not always the same. This is in con-
only increases to the top of the spike after this point).
trast with the assumptions of model (1). The consequences
are summarized in the Discussion. Determining x0 by the
minimum voltage after detected spike failed due to the large Model assumptions—Frequency analysis
fluctuations of these values. The final solution, which was
adopted, was that all data were transformed by a moving When we tried the maximum-likelihood estimates directly
average (over 6 values) and the minimum in “the valley” from the raw data we got the results which are illustrated in
after a spike is considered to be start of an ISI. This pro- Fig. 2 in a typical example.
cedure was confirmed by the following analysis (see next It obvious that the simulated trajectory differs from the
Section). To find this minimum, at first “the valley” has to be recorded one in several respects. The former reaches the
defined. steady-state much faster and the amplitude of the noise is

Springer
 J Comput Neurosci

Table 1 Example of dependence of maximum likelihood estimates

on number of steps in moving average procedure

Number of steps Estimate σ Estimate µ Estimate β

1 0.071 1.954 170.80

2 0.038 0.785 62.17
3 0.024 0.454 33.03
4 0.017 0.370 24.90
5 0.014 0.355 23.06
6 0.011 0.343 21.58
7 0.011 0.331 20.53
8 0.009 0.329 20.10
9 0.007 0.317 19.04
10 0.007 0.310 18.59

Fig. 2 Example of membrane potential trajectories, experimental

(gray) and simulated (black) using the parameters estimated by the
maximum likelihood method from the original data

Fig. 4 Example of membrane potential trajectories, experimental

(light gray), filtered (dark gray) and simulated (black) using the param-
eters estimated from the filtered data. Filtering decreased the amplitude
of the noise and made it similar to that of the simulated signal
Fig. 3 Example of spectral decomposition of the data
val (20.4201–20.54955) [s], statistical regression gives µ =
much higher. Further, it seems that the noises is not of the 0.341 [V/s] and β = 21.036 [1/s]. The results of the max-
same type. This suggests that the assumptions of the model imum likelihood estimate after computing moving averages
should be checked. The assumptions imply that the spectra are given in Table 1.
of the data should not contain a dominant frequency (white From this table we can see that, the moving average over
noise contains all the frequency components equally). We about six steps removes the discrepancy. Subsequently we
performed a spectral decomposition of several parts of the simulated the model again using the parameters estimated
data and always found dominant frequency at 2180 [Hz]. from the signal after the filtering (see Fig. 4. and compare
An example of spectral density obtained from the data is in with Fig. 3).
Fig. 3. In this example there is also high peak at 1600 [Hz]. It can be seen that the high-frequency noise present in
The source of this high frequency noise is not clear and to the experimental data has been removed and appears neither
avoid its influencing the results, all the values of the mem- in the filtered nor simulated trajectory. After the signal was
brane depolarization were transformed by a moving aver- filtered, the estimates of the parameters for each detected ISI
age over a time window of six steps. To eliminate this high were calculated according to formulas (7–9). From now on,
frequency noise we tried two strategies (moving averages, by “the data” we mean the filtered original data.
averages over non overlapping time windows). The second The differences in estimates of σ using formula (9) or
strategy appeared as inferior to the moving averaging. The (11) were negligible. For example, on the interval (20.4201,
√
success was judged from the fit of the estimates to those 20.54955); we found σ̂ = 0.0115 [V/ s] using (11), and
√
obtained from the regression of the data to an exponential σ̂ = 0.0114 [V/ s] using (9). Therefore for the noise am-
function (4), see formula (10). In the example, on the inter- plitude only estimate (9) was used.

Springer
J Comput Neurosci

Fig. 5 Dependency of the sum of the ISIs length on the serial number Fig. 7 Three curves: gray—data; fluctuating black—simulated model
of the ISI. The dotted line corresponds to the constant firing rate with parameters estimated by maximum likelihood method from the
data; smooth black—mean value (model without noise). The horizontal
line represents the estimated threshold S

the simulated depolarizations were plotted. An example is in

Fig. 7.
From the simulations made with estimated parameters,
it appears that the estimated asymptotical voltages, µ/β,
coincide well with the data. On the other hand, in the
first part of the trajectories, the real data is a bit faster in
reaching the asymptote than simulated trajectory. This vi-
sual impression was confirmed by the following method
which aimed to check the fit of the model to the data.
For each ISI we have a vector of values of depolarization
Fig. 6 Histogram of ISIs together with the corresponding density of x i = (xi0 , xi1 , . . . , xin ) a corresponding vector of depolar-
exponential distribution normalized on the number of ISIs ization, y i = (yi0 , yi1 , . . . , yin ), obtained from simulating
Eq. (1) by formula (6) using the estimated parameters. The
Results differences z i = x i − y i were calculated and their averages
and standard deviations evaluated. These results are illus-
Parameters of the input and the membrane time constant trated in Fig. 8 for both estimation methods. The main differ-
ence between the methods is in the period just after the spike
Using the above described procedure we identified all ISIs generation. This could be due to a violation of the model
and estimated the parameters for each of them. In total, 312 assumptions (for example, hyperpolarization) and possibly
ISIs were analyzed and before estimating the parameters of the method of moments could be more robust against this
the model, we applied simple standard statistical procedures violation.
on them. The ISIs appear stable in time (see Fig. 5), which Apparently the regression method works better. In both
means that there is no trend in their length. cases there is a systematic hump after the origin, but for the
The corresponding statistical characteristics are median regression method it is much smaller.
0.585 [s], average 0.872 [s] and coefficient of variation 0.883. An important question is dependency of the estimated
The shape of histogram of ISIs (Fig. 6) suggests that the ISIs parameters on the length of ISIs. The only dependency we
are generated in accordance with the exponential distribu- can expect that, if the input to neurons changes with the ex-
tion. Kolmogorov-Smirnov test does not reject the hypoth- periment, then µ could get smaller for longer ISI. Otherwise,
esis of exponentiality, at 5% a significance level. Also the µ and β should keep stable and independent on the length of
other test for normality of the estimates are at 5% significance ISI. The results are illustrated in Fig. 9. We can see, that the
level. results obtained by the regression method are independent of
The parameters of model (1) were estimated by both meth- ISI (corr (µ̂, ISI) = 0.016 and corr (β̂, ISI) = 0.136) which
ods for all the ISIs and by using the estimates in schema (6) is not the case of the estimates obtained by the maximum

Springer
 J Comput Neurosci

Fig. 8 Average difference between the data and simulated trajec-

tory. (a) estimates obtained by maximum likelihood (b) estimates ob-
tained by regression method. The region around the curves indicates
± 2 ∗ standard deviations

likelihood method (corr (µ̂, ISI) = 0.663 and

corr (β̂, ISI) = 0.668).
Similarly, we investigated the estimated value of σ and
found it independent of the ISI length, (corr (σ̂ , ISI) =
−0.099). As is apparent from the above analyses, the re-
gression method was superior to the maximum likelihood
method for estimation of µ and β (see Figs. 8 and 9),
Fig. 9 Dependency of the estimated parameters (vertical axis) on the
so the values from the regression are considered further
length of ISI (horizontal axis). In (Fig.9(a), and (b) are results obtained
on. by regression method, (Fig.9(c), and (d) those obtained by the maximum
In Fig. 10 are presented histograms of estimated values likelihood method
of the input parameters µ, σ and of the inverse time con-
stant β. We can see that the distributions are rather broad Finally we calculated the central characteristics of the
and quite symmetric resembling Gaussian density. This sug- estimated parameters. Median value of the noise amplitude
√
gests the asymptotic normality of the estimates which can was σ̂ M = 0.013505 [V/ s]. Median values of µ and β were
be used for future statistical inference. The Kolmogorov- by regression method, µ̂ M = 0.2846 [V/s], β̂ M = 25,8042
Smirnov test rejected normality of µ̂ and β̂, but not for σ̂ . [1/s], for the maximum likelihood method the obtained val-
The reason for the rejection in the case of µ̂ and β̂ were ues were µ̂ M = 0.4606 [V/s], β̂ M = 43,5068 [1/s]. Due to
the outliers on the right hand side of the histograms (see the symmetry of histograms, the averages and the medi-
Fig. 10). After their removal the estimates also fit the normal ans were practically the same. We should notice that both
distribution. methods give almost identical asymptotic depolarization

Springer
J Comput Neurosci

Fig. 11 Histograms of the membrane parameters. (a) initial membrane

depolarization, (b) firing threshold depolarization

out from the parameters estimation. The normality of both

determined parameters, x0 and S, was rejected.
To illustrate the relationship between the firing threshold
and the asymptotic depolarization we compare Fig. 12 and
Fig. 11(b). Namely we can see that the estimated parameters
predict subthreshold firing type, as from Fig. 12 we can see
Fig. 10 Histograms of the estimated parameters. (a) the drift, (b) the
noise amplitude, (c) the inverse membrane time constant
that the asymptotic depolarization is below the threshold
even with respect to its variation.
Finally, to show that the neuron, as the estimation of the
µ/β, for regression 0.0110 [V] and for maximum likelihood parameters suggests, is in the subthreshold (noise-driven)
0.0106 [V]. We can calculate the membrane time constant, regimen, we compare directly the theoretical asymptotic de-
which from the regression method yields after inverting the polarization and the corresponding firing threshold (Fig. 13).
estimate of β, the value of 38.8 ms. We can see that the difference between the threshold and
the asymptotic depolarization is almost always positive
Reset, threshold and asymptotic depolarization (Fig. 13(a)). If we investigate some kind of two-standard-
deviations envelope around the asymptotic depolarization,
For each ISI we estimated the initial value of the depolar- then we get below zero (Fig. 13(b)). The possible reasons
ization after a spike and the firing threshold (the last value for this result are presented in Discussion.
before the spike is generated), Fig. 11.
We see from comparison of Fig. 11(a) and (b) that the
initial values were more variable than the thresholds S. The
reason may be that the spike was not in principle generated Discussion
at the time when the voltage was at its highest level during
the ISI and this will be discussed later. The median value This is mostly a methodological attempt to estimate the pa-
of the initial depolarization is x0 = −73.92 [mV]. The rameters of the Ornstein-Uhlenbeck neuronal model from
median threshold value is S = −61.0 [mV]. It implies that, the intracellular recordings and for an unknown input. We
in average, the firing threshold is about 13 [mV] above the have to realize that only a single neuron in a single record was
initial depolarization. It is larger than approximately 11 [mV] analyzed and thus the results are more a methodological il-
which is the level of asymptotic depolarization as it comes lustration of how to deal with the problem than a statistically
complete analysis.

Springer
 J Comput Neurosci

tally observed many times and on very different neuronal

structures (e.g., Eggermont et al., 1993; Jones and Jones,
2000; Lin and Chen, 2000; Tateno et al., 2002). In sensory
neurons, we showed in frogs (Rospars et al., 1994) and rats
(Duchamp-Viret et al., 2005) that the spontaneous activity
of olfactory receptor neurons can be described by a Poisson
process. Theoretical arguments for the fact that the sponta-
neous activity is a low rate Poisson process can be found
in Laughlin (2001). The spontaneous activity can be inter-
preted as the summation of two processes: (i) an intrinsic
process which implies firing due to the noise, and (ii) an ex-
trinsic process, which induces firing due to the uncontrolled
occurrence of effects either from the environment or other
neurons. Of course, Poisson process, as any other model, is
an approximation which can be always questioned (Koyama
and Shinomoto, 2005), in this case for example by the exis-
tence of the absolute refractory period, but such objections
are marginal at this level of description.
From the point of view of the Ornstein-Uhlenbeck neu-
ronal model, the Poisson process corresponds to the situation
Fig. 12 Histograms of the asymptotic properties calculated from the in which the signal is so weak that the asymptotic depolar-
parameter estimates. (a) asymptotic depolarization, (b) asymptotic stan- ization is far below the firing threshold, which has to be true
dard deviation of the depolarization with respect to the amplitude of noise. In other words, for
model (1) and (2) this type of activity is predicted if the firing
threshold S is far above the asymptotic depolarization, µ/β,
given by Eq. (4) with respect the asymptotic variance (5),
which is controlled by σ . We can see from Fig. 13(a), that the
asymptotic depolarization is below the threshold. However,
when we plot the asymptotic depolarization increased for two
asymptotic standard deviations, we get above the threshold.
This could break the exponentiality of ISIs distribution ob-
served in our data. The reason why this has not happened may
be that not every crossing of the voltage has induced a spike,
in other words, that we had not recorded exactly the trigger
zone depolarization. Namely, two-compartment version of
the model would be closer to reality. Similar conclusion can
be found in Jolivet et al. (2006).

Model assumptions

There are many assumptions of the model which are dis-

putable and can be true only in idealization. First of all, it
Fig. 13 Histogram of the pair wise differences between the es- is the constancy of the input parameters µ and σ over each
timated firing thresholds and the asymptotic √ depolarizations. (a) ISI. We selected the spontaneous activity as an experimen-
Ŝ − (x̂0 + µ̂/β̂), (b) Ŝ − (x̂0 + µ̂/β̂) − 2σ/ 2β tal material, for which such a constancy can be expected.
Another assumption is the Gaussian white noise on the neu-
Spontaneous activity ronal input. This we found was not true for our data and
the high-frequency noise was detected and eliminated. The
We selected for statistical treatment only the unstimulated source of this high frequency noise remains unclear to us.
activity of the neuron. The detected ISIs are exponentially Next assumption, violated in our observations, was the first-
distributed which suggests that they are generated in ac- passage time inducing the generation of a spike. Probably,
cordance with a Poisson process. The Poisson process as a we registered the membrane depolarization in other location
model of spontaneous activity of a neuron was experimen- than the spikes are generated. The traditional assumption of

Springer
J Comput Neurosci

the model is the fixed value of the reset depolarization (for neurons investigated by Inoue et al. (1995) there were ei-
the variants with random initial value see Lansky and Smith, ther in sleeping animals or in bird watching state, but there
1989; Christodoulou and Bugmann, 2000). This assumption is no apparent difference in the activity or the parameters.
represents an oversimplification which was noticed as early We should keep in mind that the firing rates were higher
as by Stevens (1964). We found that this assumption was also than in our case and none of the neurons resembled Pois-
violated, but for the model performance the effect is negligi- sonian firing. There the firing threshold and the membrane
ble. Despite that any assumption of the model can be made time constants were selected a priory, S was set 15 [mV]
questionable, as it is only a model, we may conclude that the above the reset value and the time constant was taken equal
data are consistent with the model. This was not the case in to 5 [ms]. The threshold obtained in this study (approx. 13
Stevens and Zador (1998) where model (1–2) in absence of [mV]) is not so different from the value assumed by these
noise was fitted to response of cortical neurons in vitro to authors. On the other hand, the membrane time constant we
injection of constant current. There the voltage approached estimated was about seven times larger than that used by
the threshold as a concave curve, in contrast to Eq. (4), and them. This fact corresponds well to the fact that our values
our data. The authors of the paper solved the discrepancy of the drift parameter were estimated lower than in the cited
by employing time-varying membrane resistance and time- paper. The estimated µ ranges in their paper from −6 to
varying membrane time “constant”. In contrast to Stevens 3 [mV/ms], whereas in this study we obtained the values
and Zador (1998), here the investigated trajectories shown in much smaller range, around 0.28 [mV/ms]. Finally, the
the convex shapes, but it may happen that when a neuron is most striking at the first sight seems to be the difference
stimulated, this property is met. in the amplitude of noise, which usually was found much
larger than in this paper. However, it is indirectly clear that
Intrinsic parameters the values in Inoue et al. (1995) are in different units than
in this paper ([mV/ms]). After rescaling, the values the dif-
Tuckwell and Richter (1978), who pioneered the estimation ference becomes less apparent. It is an open question if the
of the parameters in the stochastic neuronal models, clas- comparison of our results with Inoue et al. (1995) can be con-
sified the model parameters into two classes—intrinsic and sidered as a discrepancy, or if their results are so widespread
input. Unfortunately, their results are based on different as- that ours can be seen included in theirs. A possible source
sumptions and thus not comparable with ours. In model (1–2) of discrepancy between these two papers is probably dif-
are three intrinsic parameters, β, S and x0 , and for different ference in applied method and the overall activity of the
neurons have been reported different values of them, even neurons.
without modeling concept. However, the intrinsic parame- Another attempt to compare the Ornstein-Uhlenbeck
ters were not in the center of our interest despite they play model with ISI data was done by Shinomoto et al. (1999)
their important role for the model performance. Jolivet et al. with a negative result. Their method is not based on direct es-
(2006) analyzed spike response model which was claimed to timation of the model parameters but on studying mutual re-
be equivalent to the leaky integrate-and-fire model. However, lationship between coefficient of variation and the skewness
their approach is not oriented on estimating the parameters coefficient. They concluded that the model is not adequate
µ and σ , but it is aimed on spike train prediction. The time- to account for the spiking in cortical neurons. These authors
dependent threshold is considered in their paper, but the did not estimate the membrane time constant and considered
constant value is achieved in about 10 ms, which would not for it several optional values. This complicates the compar-
change our results as no ISI shorter of 10 ms was observed. In ison of their results with ours. However, they claimed that
their Fig. 1(a) we can see, that the threshold reaches the value the reason for inconsitency of their data with the Ornstein-
between −60 and −70 mV which corresponds to our results. Uhlenbeck process was mainly due to anomalous long ISIs.
In general, the values of intrinsic parameters obtained in this In our case the parameters were found in the subthreshold
article are consistent with the values found in literature. region and thus the long ISIs represent no problem for the
fit of the data to the model. On the other hand, in the case of
Input parameters the stimulated activity such a situation can arise.
Both these attempts (Inoue et al., 1995; Shinomoto et al.,
Completely different situation comes with the input param- 1999) were based on ISIs statistics. It means that for estima-
eters. The attempts to estimate them were up to now rare tion of the parameters a sample of several hundreds of ISIs is
and based on additional assumptions. Inoue et al. (1995) an- necessary. Our method permits to estimate the input parame-
alyzed spontaneous activity of the mesencephalic reticular ters in a short time window (in a single ISI). It appeared that
formation neurons on the basis of ISIs. It should be noted the regression method was superior to the maximum like-
that the term spontaneous activity has in central neurons dif- lihood. There might be several reasons for this effect. The
ferent meaning than in the sensory systems. In the case of first one can be the above mentioned violation of the model

Springer
 J Comput Neurosci

assumptions and thus that the maximum-likelihood may be References

more sensitive to these discrepancies between the model and
the data. The second reason is that the maximum likelihood Brillinger DR, Segundo JP (1979) Empirical examination of the
estimates are discretized version of continuous sampling the- threshold model of neuron firing. Biol. Cybern. 35: 213–220.
Christodoulou C, Bugmann G (2000) Near poisson-type firing produced
ory. The asymptotic depolarization µ/β was estimated very
by concurrent excitation and inhibition. BioSystems 58: 41–48.
well by both methods, better than the parameters µ and β Dayan P, Abbot LF (2001) Theoretical Neuroscience. MIT Press,
separately. The reason is that the membrane potential was Cambridge, MA.
almost permanently at the asymptotic level and in this situ- Ditlevsen S, Lansky P (2005) Estimation of the input parameters in
the Ornstein-Uhlenbeck neuronal model. Phys. Rev. E 71: Art.
ation the estimation of individual parameters is less precise No. 011907.
(Lansky et al., 1988). The distinction on input and intrin- Duchamp-Viret P, Kostal L, Chaput M, Lansky P, Rospars J-P (2005)
sic parameters fails for more realistic leaky integrate-and- Patterns of spontaneous activity in single rat olfactory receptor
fire models (e.g., Lansky and Lanska, 1987; Richardson and neurons are different in normally breathing and tracheotomized
animals. J Neurobiol.
Gerstner, 2005). There the membrane time constant becomes Eggermont JJ, Smith GM, Bowman D (1993) Spontaneous burst firing
input dependent. The change is only formal at the level of in cat primary auditory-cortex—Age and depth dependency and
description applied in this article. its effect on neural interaction measures. J. Neurophysiol. 69:
1292–1313.
Feigin P (1976) Maximum likelihood estimation for stochastic
Noise processes—A martingale approach. Adv. Appl. Probab. 8:
712–736.
Gardiner CW (1983) Handbook of Stochastic Methods for Physics,
The results suggest that our neuron was firing in the noise Chemistry and the Natural Sciences. Springer, Berlin.
activated regimen, in other words, that in the absence of the Gerstner W, Kistler W (2002) Spiking Neuron Models. Cambridge
noise it would remain silent. This corresponds very well to University Press, Cambridge.
the fact that the driving signal is small and the neuron fires He J (2003) Slow oscillation in non-lemniscal auditory thalamus. J
Neurosci. 23: 8281–8290.
only due to the stochastic fluctuation of the membrane de- Inoue J, Sato S, Ricciardi LM (1995) On the parameter estimation for
polarization. Theoretical prediction of the Poissonian firing diffusion models of single neurons’ activities. Biol. Cybern. 73:
in the subthreshold regime of the Ornstein-Uhlenbeck neu- 209–221.
ronal model is well known for a long time (Nobile et al., Johnson DH (1996) Point process models of single-neuron discharges.
J. Comput. Neurosci. 3: 275–300.
1985) and here the prediction and data estimation fits per- Jolivet R, Rauch A, Lüscher H-R, Gerstner W (2006) Integrate-and-fire
fectly together. The values of the estimated noise amplitude models with adaptation are good enough: Predicting spike times
(Fig. 10(b)) seems to be quite small, but this is only an illu- under random current injection. In: Y Weiss, B Schölkopf, J Platt,
sion as what has to be considered comes out of Eq. (6), and eds. Advances in Neural Information Processing Systems 18,
MIT Press, Cambridge MA.
it is the asymptotic standard deviation of the depolarization,
√ Jones TA, Jones SM (2000) Spontaneous activity in the statoacoustic
σ 2β. ganglion of the chicken embryo. J. Neurphysiol. 83: 1452–1468.
Keat J, Reinagel P, Reid RC, Meister M (2001) Predicting every spike:
A model for the responses of visual neurons. Neuron 30: 803–
817.
Conclusions Kistler WM, Gerstner W, van Hemmen JL (1997) Reduction of the
Hodgkin-Huxley equations to a single-variable threshold model.
Neural Comput. 9: 1015–1045.
We estimated the parameters of the Ornstein-Uhlenbeck neu- Kloeden PE, Platen E (1992) Numerical Solution of Stochastic
ronal model in spontaneous neuronal activity. The achieved Differential Equations. Springer, Berlin.
results are consistent with the conclusions which can be ob- Koch C (1999) Biophysics of Computation: Information Processing in
tained from the statistical analysis of the ISIs. The neuron Single Neurons. Oxford University Press, Oxford.
Koyama S, Shinomnoto S (2005) Empirical Bayes interpretation of
fires in subthreshold regimen and thus the activity is Poisso- random point events. J. Phys. A: Math. Gen. 38: L531–L537.
nian. The advantage of the applied method is that it permits La Camera G, Rauch A, Luscher HR, Senn W, Fusi S (2004) Minimal
to judge quantitatively the input to the neuron within a single models of adapted neuronal response to in vivo-like input
ISI. This property will appear more important in presence of currents. Neural Comput. 16: 2101–2124.
Lansky P (1983) Inference for the diffusion models of neuronal
stimulation and comparison of neuronal activity under dif- activity. Math. Biosci. 67: 247–260.
ferent conditions. Lansky P (1997) Sources of periodical force in noisy integrate-and-fire
models of neuronal dynamics. Phys. Rev. E 55: 2040–2043.
Lansky P, Lanska V (1987) Diffusion approximation of the neuronal
Acknowledgments The authors thank to P.W.F. Poon for initiating model with synaptic reversal potentials. Biol. Cybern. 56:19–26.
and permanent support for this work and to P.E. Greenwood for stimu- Lansky P, Giorno V, Nobile AG, Ricciardi LM (1998) A diffusion
lating discussions. This work was supported by Academy of Sciences neuronal model and its parameters. In: LM Ricciardi, ed. Pro-
of the Czech Republic Grant (Information Society, 1ET400110401), ceedings of International Workshop Biomathematics and related
AV0Z50110509, Center for Neurosciences LC554. Computational Problems. Kluwer, Dordrecht.

Springer
J Comput Neurosci

Lansky P, Lanska V (1994) First-passage-time problem for simulated enhanced phase-locking in a leaky integrate-and-fire model of a
stochastic diffusion processes. Comp. Biol. Med. 24: 91–101. neuron. Phys. Rev. E 62: 8427–8437.
Lansky P, Smith CE (1989) The effect of a random initial value in Rospars J-P, Lansky P, Vaillant J, Duchamp-Viret P, Duchamp A
neural 1st-passage-time models. Math. Biosci. 93: 191–215. (1994) Spontaneous activity of first- and second-order neurons in
Laughlin SB (2001) Energy as a constraint on the coding and processing the olfactory system. Brain Res. 662: 31–44.
of sensory information. Curr. Opin. Neurobiol. 11: 475–480. Segev I (1992) Single neurone models: Oversimple, complex and
Lin X, Chen SP (2000) Endogenously generated spontaneous spiking reduced. TINS 15:414–421.
activities recorded from postnatal spiral ganglion neurons in vitro. Shinomoto S, Sakai Y, Funahashi S (1999) The ornstein-uhlenbeck
Developmental Brain Res. 119: 297–305. process does not reproduce spiking statistics of neurons in
Nobile AG, Ricciardi LM, Sacerdote L (1985) Exponential trends of prefrontal cortex. Neural Comp. 11: 935–951.
Ornstein-Uhlenbeck 1st-passage-time densities. J. Appl. Prob. Stein RB (1965) A theoretical analysis of neuronal variability. Biophys.
22: 360–369. J. 5: 173–195.
Paninski L, Pillow J, Simoncelli E (2005) Comparing integrate-and- Stevens CF, Zador AM (1998) Novel Integrate-and-fire-like Model of
fire models estimated using intracellular and extracellular data. repetitive firing in cortical neurons. Proceedings of the 5th Joint
Neurocomputing 65: 379–385. Symposium on Neural Computation, UCSD, La Jolla CA.
Paninski L, Pillow JW, Simoncelli EP (2004) Maximum likelihood Stevens CF (1964) Letter to the editor. Biophys. J. 4: 417–419.
estimation of a stochastic integrate-and-fire neural encoding Tateno T, Kawana A, Jimbo Y (2002) Analytical characterization of
model. Neural Comput. 16: 2533–2561. spontaneous firing in networks of developing rat cultured cortical
Pinsky PF, Rinzel J (1994) Intrinsic and network rhythmogenesis in a neurons. Phys. Rev. E 65: Art. No. 051924.
reduced Traub model for CA3 neurons 1: 39–60. Tuckwell HC (1988) Introduction to Theoretical Neurobiology.
Prakasa Rao BLS (1999) Statistical inference for diffusion type Cambridge Univ. Press, Cambridge.
processes. Arnold, London. Tuckwell HC, Lansky P (1997) On the simulation of biological
Rauch A, La Camera G, Luscher HR, Senn W, Fusi S (2003) diffusion processes. Comput. Biol. Med. 27: 1–7.
Neocortical pyramidal cells respond as integrate-and-fire neurons Tuckwell HC, Richter W (1978) Neuronal interspike time distribution
to in vivo-like input currents. J. Neurophysiol. 90: 1598–1612. and the estimation of neurophysiological and neuroanatomical
Ricciardi LM, Lansky P (2003) Diffusion models of neuronal activity. parameters. J. theor. Biol. 71: 167–183.
In: MA Arbib, ed. The Handbook of the Brain Theory and Neural Xiong Y, Yu YQ, Chan YS He J (2003) An in-vivo intracellular study
Networks, (2nd edn.) MIT Press, Cambridge, MA. of the auditory thalamic neurons. Thalamus Related Sys. 2: 253–
Richardson MJE, Gerstner W (2005) Synaptic shot noise and con- 260.
ductance fluctuations affect the membrane voltage with equal Yu YQ, Xiong Y, Chan YS He JF (2004) Corticofugal gating of
significance Neural Comput. 17: 923–947. auditory information in the thalamus: An in vivo intracellular
Rodriguez R, Lansky P (2000) Effect of spatial extension on noise recording study. J. Neurosci. 24: 3060–3069.

Springer
 Author's personal copy

Journal of Physiology - Paris 104 (2010) 160–166

Contents lists available at ScienceDirect

Journal of Physiology - Paris

journal homepage: www.elsevier.com/locate/jphysparis

Effect of stimulation on the input parameters of stochastic leaky

integrate-and-fire neuronal model
Petr Lansky a,*, Pavel Sanda a, Jufang He b
a
Institute of Physiology, Academy of Sciences of the Czech Republic, Prague, Czech Republic
b
Department of Rehabilitation Sciences, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong

a r t i c l e i n f o a b s t r a c t

Keywords: The Ornstein–Uhlenbeck neuronal model is specified by two types of parameters. One type corresponds
Interspike interval to the properties of the neuronal membrane, whereas the second type (local average rate of the mem-
Ornstein–Uhlenbeck
brane depolarization and its variability) corresponds to the input of the neuron. In this article, we esti-
Intracellular recording
mate the parameters of the second type from an intracellular record during neuronal firing caused by
Estimation of parameters
stimulation (audio signal). We compare the obtained estimates with those from the spontaneous part
of the record. As predicted from the model construction, the values of the input parameters are larger
for the periods when neuron is stimulated than for the spontaneous ones. Finally, the firing regimen of
the model is checked. It is confirmed that the neuron is in the suprathreshold regimen during the
stimulation.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction techniques to analyze them. We can see from this review that with
exception of a few theoretical attempts, the comparison of the
Spiking activity is the basic mode of the information transfer models with data has been rather neglected. Only recently the phe-
within the nervous system. The sequences of spikes (action poten- nomenological models are confronted with experimental data, for
tials) are generated and sent along an axon to other neurons. These a review see Lansky and Ditlevsen (2008).
action potentials are considered to be none-or-all events with Stochastic diffusion integrate-and-fire neuronal model (the
shapes irrelevant for the information they convey. Formally, they Ornstein–Uhlenbeck process) describes the membrane potential
are taken in the limit as Dirac delta function and the complete se- as a continuous-time stochastic process. Along the introduction
quence of them as a realization of a stochastic point process. The of this model there have been given many arguments why this
application of the theory of stochastic point processes in descrip- treatment of the integrate-and-fire model is appropriate. Leakage
tion of spike trains is very common and the phenomenological of the neuronal membrane, it means the current which flows
models of single neurons predicting properties of these point pro- through the membrane due to its passive properties, was one of
cesses are often investigated (Tuckwell, 1988; Gerstner and Kistler, the first specification of the integrate-and-fire neuron model. It is
2002; Dayan and Abbott, 2001). The models are build to generate a crucial property of the integrate-and-fire models and thus it is
interspike intervals (ISIs) and they are often based on the first-pas- inherent for practically all the variants of the model. Generaliza-
sage-time principle for so called integrate-and-fire models which tions were recently introduced aiming to improve flexibility of
mimic accumulation of the incoming signal and the final genera- the model and its predictive power (Clopath et al., 2007; Jolivet
tion of the spike is replaced by instantaneous reset of the generator et al., 2006).
to the initial level (Brunel and van Rossum, 2007). The model investigated in this paper has parameters of two
Attempts to compare the experimental data with the models is types. The first are the parameters which can be measured by indi-
very common for so called biophysical models of neurons, but con- rect electrophysiological methods, deduced from the properties of
sidering the phenomenological models, they are more frequently other neurons or from measuring the membrane potential fluctua-
compared qualitatively, and the researchers are satisfied if they tions. If these parameters are known, one can check how well the
perform in a similar way to the real neurons. Burkitt (2006) re- model predicts spiking activity under the condition of an input
viewed the integrate-and-fire neuron models and mathematical identical with the input to a real neuron. The second set of param-
eters, investigated in this paper, is identified with the signals
impinging upon the neuron. Knowledge of these parameters can
be used either to deduce unknown signal coming to a neuron or
* Corresponding author. Address: Institute of Physiology, Academy of Sciences,
Videnska 1083, 142 20 Prague 4, Czech Republic. to check whether we are able to read correctly an artificially deliv-
E-mail address: [email protected] (P. Lansky). ered signal.

0928-4257/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jphysparis.2009.11.019
 Author's personal copy

P. Lansky et al. / Journal of Physiology - Paris 104 (2010) 160–166 161

To estimate the input signal either the membrane depolariza-  The subthreshold regimen ðl=b  S  x0 Þ with Poissonian firing.
tion or ISIs have to be at our disposal. As mentioned, the previous As the asymptotic mean depolarization does not reach the
attempts to identify the model parameters were based on observa- threshold the firing depends on the noise and without it the
tion of ISIs. Such a task is complicated and leads to rather difficult neuron would remain silent. It was shown in our previous article
numerical and mathematical problems (Inoue et al., 1995; Shinom- Lansky et al. (2006) that the spontaneous part of the recorded
oto et al., 1999; Ditlevsen and Lansky, 2005, 2006, 2007; Koyama data fits mostly with subthreshold regimen characterization.
and Kass, 2008; Mullowney and Iyengar, 2008). In the study Lansky  The threshold regimen ðl=b  S  x0 Þ, where the distribution of
et al. (2006) estimation of five basic parameters the Ornstein– ISIs is positively skewed and resembles for example Gamma
Uhlenbeck model was studied for the membrane depolarization distribution.
data. The whole estimation procedure was based on the spontane-  The suprathreshold regimen ðl=b  S  x0 Þ, where the firing is
ous-firing part of the intracellular recording. The primary aim of almost regular and ISI histogram resembles normal distribution.
this study is to extend the results on the stimulation part of the re- The noise plays a limited role in this range of parameters.
corded signal.
As already mentioned in Section 1 two basic types of data can
be used for the identification of the parameters of the Ornstein–
2. Model and its properties
Uhlenbeck model. If only the ISIs are available, the methods appli-
cable in this situation are reviewed in Lansky and Ditlevsen (2008).
The deterministic leaky integrate-and-fire model (Lapicque
In the second situation, which is investigated here, the membrane
model, RC-circuit) can be derived from the assumptions that the
depolarization is recorded in between the spike generation. Some
membrane depolarization is described by a circuit with a genera-
methods for estimation of the parameters under this sampling
tor, a resistor and a capacitor in parallel. It has to be stressed that
were compared for this type of data in Lansky et al. (2006) and
while the electrical representation is related to a small isopotential
now we apply only the selected ones.
patch of neuronal membrane, the voltage in the model reflects an
abstract representation of a complete neuron; usually described as
depolarization at the trigger zone. The trigger zone serves as a ref- 3. Methods
erence point and all the other properties of the neuron are inte-
grated into it. 3.1. Data collection
The Ornstein–Uhlenbeck model of membrane depolarization is
a stochastic variant of the RC-circuit model and can be formally de- 3.1.1. Animal preparation and data preprocessing
scribed by a stochastic differential equation, Guinea pigs served as subjects for the intracellular recording
experiments. Throughout the recording, an electrocorticograph
dXðtÞ was monitored to assess the level of anaesthesia. A midline inci-
¼ bðXðtÞ  x0 Þ þ l þ rnðtÞ; Xð0Þ ¼ x0 ; ð1Þ
dt sion was made in the scalp and a craniotomy was performed to en-
able vertical access to the MGB in the right hemisphere (He, 2003;
where b reflects the spontaneous decay of the membrane depolar-
Xiong et al., 2006; Yu et al., 2004). The experimental procedures
ization to the resting level x0 ; l represents the local average rate
were approved by the Animal Subjects Ethics Sub-Committee of
of displacement due to the neuronal input, r reflects the variability
The Hong Kong Polytechnic University. Upon penetrating the
of the membrane potential due to the neuronal input and nðtÞ
membrane of a cell, the electrode detected the negative membrane
stands for Gaussian white noise. The spikes are not an intrinsic part
potential. After amplification, the membrane potential as well as
of model (1) but are generated when the membrane depolarization
the auditory stimulus were stored in the computer with the aid
XðtÞ reaches the firing threshold S for the first time. So, S is an addi-
of commercial software (AxoScope, Axon). Single neuron data were
tional parameter. After firing, the depolarization is reset to the rest-
selected for this article. The membrane potential was recorded
ing level, x0 , and the process of input ‘‘integration” starts anew. Also
with time step h ¼ 0:00015 s ¼ 0:15 ms , for period 0–501 s. For
the reset level, x0 , represents a parameter of the model. The model is
further processing we make basic noise-filtering of the membrane
fully described by five parameters: b; l; r, S and x0 , which specify
potential by a moving average over six values, see Lansky et al.
Eq. (1) together with its initial and boundary condition. More details
(2006) for details. The acoustic signal used for the stimulation of
on the Ornstein–Uhlenbeck neuronal model can be found, for exam-
neuron has duration of 0.1 s and the series of the acoustic signals
ple, in Tuckwell (1988) or Burkitt (2006). The parameters intro-
divided the record into the sections of stimulated and spontaneous
duced above can be divided in two categories:
parts. We compare the stimulated parts of the record with the
unstimulated ones.
 Parameters depending on the membrane properties – b being
the inverse of the membrane time constant, threshold S and
resting level x0 . 3.1.2. Detection of spikes
 Parameters depending on the input signal – l representing the We detected and selected 86 stimulated parts in the record,
mean signal and r characterizing its variability. In this study which were used for the analysis. An example of a data can be seen
we identify these parameters. As we focus on the changes in in Fig. 1.
them brought by the stimulation, we are interested in compar- A typical neuronal response varies between two and three
ing the values of these two parameters in spontaneous and stim- spikes during the stimulation. In the spontaneous part, the spikes
ulated parts of the data. appear rarely. As can be seen in Fig. 2 there is a substantial change
of depolarization course after the second spike within the period of
There can be posed serious questions whether b; S and x0 are stimulation.
independent of the input. Actually, from the experimental data Because of this fact we need to distinguish between the first and
we directly found that, at least, x0 changes in the presence of stim- the following ISIs within one stimulation. For this purpose we di-
ulation and we further take this fact into account. vide the periods of stimulation according to the number of ISIs in
The position of the asymptotic mean depolarization them. We got three periods of stimulation without any complete
EðXð1ÞÞ ¼ x0 þ l=b, as seen from Eq. (1) determines three basic re- ISI, 68 periods of stimulation containing just one ISI, 14 periods
gimes of firing of the Ornstein–Uhlenbeck model: of stimulation containing two ISIs and one period of stimulation
 Author's personal copy

162 P. Lansky et al. / Journal of Physiology - Paris 104 (2010) 160–166

-0.04

-0.045
valley detection level
-0.05

Depolarization [V]
-0.055 start-valley

-0.06 threshold

-0.065

-0.07

end-valley end-valley
-0.075
181 181.1 181.2 181.3 181.4 181.5 181.6 x0
Time [s] -0.08
182.63 182.64 182.65 182.66 182.67
Fig. 1. Example of the acoustic signal (shaded curve) and the membrane depolar- Time [s]
ization (black curve). There are no spikes in the absence of stimulation and three
spikes during the stimulation. Fig. 3. ISI detection. We can see three spikes defining two ISIs within period of
stimulation. The values of the membrane potential from x0 to the end of the valley
are used for the parameters estimation. See text for details.
-0.05

of the last point with decreasing depolarization before the next

Stimulation
-0.055 spike or the last point when the threshold S is crossed before next
spike. Both cases can be seen in Fig. 3.
Depolarization [V]

Threshold S itself is not searched for in periods of stimulation,

-0.06
as the heuristic method used in the spontaneous part failed. Here
the depolarization often goes from the valley after spike straightly
-0.065 (almost linearly) towards next spike without crossing any specific
point, which can be clearly marked as the threshold (see Fig. 2).
Disregarded 2. ISI Model (1) considers the threshold as the membrane property and
-0.07 ISI thus we take its value S ¼ 61 mV as derived from the spontane-
Disregarded
1. ISI ISI
ous part of the data (Lansky et al., 2006). In the same way, for
the inverse of the membrane time constant we fix b ¼ 25:8 1=s
-0.075
222.61 222.63 222.65 222.67 222.69 as it was estimated from the spontaneous activity.
Time [s] After detecting the valley we can detect the corresponding ISI.
We identify reset potential x0 as the minimal voltage value in the
Fig. 2. An example of membrane depolarization course during the period of valley and it is the only intrinsic parameter determined in this pa-
stimulation containing two ISIs.
per, see Section 4.2.3. The ISI is defined as interval between the
time when XðtÞ ¼ x0 and the end of the valley. The period from
the spike to the beginning of ISI can be identified as the refractory
containing three ISIs. It is obvious from this distribution, that for
period. For a recent review of methods for determination of refrac-
statistical evaluation only the first ISIs (68 + 14 + 1) can be used
tory period from ISI data, see Hampel and Lansky (2008). Here we
(the second and the third ISIs are too rare).
do not investigate this problem, but we complement the picture by
In the spontaneous part we got 312 complete ISIs and because
giving the values derived from the depolarization but not the ISIs,
the spike frequency in this part was substantially lower the shapes
see Section 4.2.4. We should note that the current procedure ap-
of membrane depolarization were similar and there was no need to
plied to the data is slightly different from that used in our previous
distinguish serial number of ISIs in spontaneous part. Note that the
paper. In this study the valley has to be defined in a more complex
comparison between stimulated and spontaneous records is based
way as the lengths of ISIs are very short and the shapes of mem-
on ISIs which are completely within these specific periods and thus
brane potential are much more variable (see details in Fig. 3).
the ISIs which are partly in both periods are disregarded (see
Fig. 2).
3.3. Parameters estimation

3.2. Precise detection of ISI There are two parameters of model (1) driven by the incoming
signal to the neuron – l and r. In Lansky et al. (2006), the regres-
The membrane potential trajectories between spikes take a sion method appeared to be more appropriate than the likelihood
shape of valleys in the stimulated parts of the record (see Fig. 2). method for estimating the parameter l. To apply the regression
Firstly, we formulate a heuristic procedure how to detect these val- method we minimize the functional
leys in the record. As the estimation of the parameters is based on
XN  2
the membrane potential trajectory which is entirely outside the l
LðlÞ ¼ xj  x0  ð1  expðbjhÞÞ ð2Þ
spikes, we have to determine carefully from what time and up to b
j¼1
which time we consider the data. Then, these two time instances
implicitly define the corresponding ISI. with respect to the parameter l, where h stands for time step and xj
For the detection of spike we fix the voltage level at 35.5 mV are individual measured values of membrane depolarization for the
as in the previous study (Lansky et al., 2006). For the beginning total time T ¼ Nh. What is minimized in Eq. (2) is the distance be-
of the valley we pose the valley-detection level at 50 mV. For tween observed values of the membrane potential and the mean
the detection of the end of the valley we look for the maximum depolarization in the absence of the threshold. There are two draw-
 Author's personal copy

P. Lansky et al. / Journal of Physiology - Paris 104 (2010) 160–166 163

backs of the method. At first, it is the fact, that the predicted values 4.2. Comparison of the parameters
are those expected in the absence of the threshold. We investigated
this fact deeply in Bibbona et al. (2008) and the effect on the estima- 4.2.1. Parameter l
tion is not substantial. At second, the observed values of the mem- It can be clearly seen in Fig. 5 that the estimates of l have much
brane depolarization are not independent realizations of random broader distribution with median around 1.1 V/s in the stimulated
variables and it restricts the conclusions made on the basis of the part than in the spontaneous part, where the distribution is located
regression. Nevertheless, in our experience, the method is an around the value of 0.3 V/s. The fact that for stimulated part all the
acceptable compromise between the tractability and efficiency. estimated values of l are higher than in the spontaneous one cor-
For estimation of the noise amplitude we use the formula ob- responds with the interpretation of l in the Ornstein–Uhlenbeck
tained by the maximum likelihood method model. The lower variability of l in the spontaneous regime corre-
sponds to the fact that their values are lower, see Table 1, but the
1 X
N 1
relative variability, as reflected by CV, is higher for the spontaneous
c
r2 ¼ ^ Þ2 ;
ðxjþ1  xj þ xj hb  hl ð3Þ
T j¼0
data.

and also the formula established by Feigin (1976), which is inde- 4.2.2. Parameter r
pendent of the other parameters estimation: In Fig. 6 we can see that the situation for r is analogous to the
case of l. The estimate of the variance r has broader distribution
1 X
N1 pffiffi
c
r2 0 ¼ ðxjþ1  xj Þ2 : ð4Þ with median around 0:026 V= s in the stimulated parts than in
T j¼0 the spontaneous parts, where r has more narrow range with sig-
pffiffi
nificantly higher peak around the value of 0:013 V= s. The results
In this way for each ISI a pair of estimated l
^ i; r
^ i is computed.
are also documented in Table 2, where we can realize that Eq. (3)
gives slightly lower values of estimates than Eq. (4).
4. Results and discussion In correspondence with Lansky and Sacerdote (2001) there is
higher r for the stimulated ISIs. This is the first experimental con-
We investigate the effect of stimulation in two directions. At firmation of our results originally achieved on entirely theoretical
first, we compare simulations of model (1) based on the estimated basis. To get a better picture of the relationship between l and r
parameters with the experimental record. Secondly, we compare different levels of the stimulation would be necessary.
the stimulated ISIs and their respective l; r parameters to ISIs
and the parameters from the spontaneous part of the record. 4.2.3. Reset potential x0
It is apparent from the data and illustrated in Fig. 2 that the re-
4.1. Model and data comparison set value is influenced by the stimulation. This can be interpreted
in such a way, that the accumulation of the incoming signal takes
As already stated, we have 83 ISIs which are the first ones and place during the reset of the membrane potential, resp., during the
completely contained in the stimulation periods. For i-th ISI refractory period. Obviously, for the second ISI within one stimula-
ði ¼ 1; . . . ; 83Þ we have a vector of values of depolarization tion period, there are additional reasons why x0 changes. Thus the
xi ¼ ðxi0 ; xi1 ; . . . ; xin Þ, where n depends on the length of the ISI. Cor-
responding vector of mean depolarization, yi ¼ ðyi0 ; yi1 ; . . . ; yin Þ was
8
obtained from simulating Eq. (1) using the estimated parameter l ^i
and ri ¼ 0. The differences zi ¼ xi  yi were calculated and their 7
average and standard deviation evaluated. The results are shown
6
in Fig. 4. We can see that model based on the estimated parameters
does not show any systematic error in the course of membrane 5
depolarization.
Density

Spontaneous
4

3
0.003
2

1 Stimulated
0.002

0
0.001 -0.5 0 0.5 1 1.5 2 2.5
Depolarization [V]

0 Fig. 5. Distribution of the estimates of the parameter l: left curves – spontaneous,

right curves
ðlmÞ2
– stimulated. Dotted lines are fitted normal distributions f ðlÞ ¼

p1ffiffiffiffi e 2s2 ; m
stim ¼ 1:158; sstim ¼ 0:304 and mspont ¼ 0:283; sspont ¼ 0:091.
s 2p
-0.001

-0.002 Table 1
Descriptive statistics for the estimates of parameter l.
-0.003 l^ ½V=s Stimulated Spontaneous
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014
Min 0.6046 0.04665
Time [s] Max 2.0840 0.92737
Median 1.1061 0.28460
Fig. 4. Comparison of experimental data with the model. The central solid curve is Mean 1.1580 0.28324
an average of the differences zi , dotted curves indicate 2  standard dev iations . CV 0.2645 0.32156
See text for details.
 Author's personal copy

164 P. Lansky et al. / Journal of Physiology - Paris 104 (2010) 160–166

400 Table 3
Descriptive statistics for parameter x0 .
350
x^0 ½mV Stimulated Spontaneous
300 Min 74.92 77.25
Max 67.08 67.33
250 Spontaneous Median 70.58 73.92
Density

Mean 70.52 73.90

200 CV 0.024 0.023

150
Stimulated
100
1400
50
1200
0
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
σ 1000

Fig. 6. Distribution of the estimates of the parameter r: left curves – spontaneous, 800

Density
right curvesðlmÞ2
– stimulated. Dotted lines are fitted normal distributions f ðlÞ ¼

p1ffiffiffiffi e 2s2 ; m
stim ¼ 0:0264; sstim ¼ 0:0032 and mspont ¼ 0:0135; sspont ¼ 0:0012.
s 2p
600

400

Table 2
200
Descriptive statistics for the estimates of parameter r.
pffiffi
^0
r ½V= s Stimulated Spontaneous Stimulated (r
^ from Eq. (3)) 0
1 1.5 2 2.5 3 3.5 4
Min 0.01843 0.01043 0.01576
Max 0.03686 0.01681 0.03166 Time [ms]
Median 0.02646 0.01351 0.02262
Mean 0.02640 0.01348 0.02215 Fig. 8. Distribution of the time from the peak of the spike to the resting potential,
CV 0.12218 0.08310 0.12566 169 (=3 + 2  68 + 2  14 + 2  1 – two first spikes taken, if possible) spikes were
regarded.

250 Table 4
Descriptive statistics for the time to reach the resting potential.

Time to reach the resting potential [ms]

200
Spontaneous
Stimulated Min 1.35
Max 3.6
150 Median 1.8
Density

Mean 1.9
CV 0.21
100

starts on the top of the spike and calculate the time till the first lo-
50
cal minimum in the membrane depolarization after the spike. This
procedure has the advantage that we can take both the first and
0 the second spikes in stimulated region for statistical evaluation –
-80 -78 -76 -74 -72 -70 -68 -66 -64
as it follows from the previous section, it is not generally possible
Asymptotic depolarization [mV]
to define x0 after the second spike. The density estimation of the
Fig. 7. Distribution of the reset potential x0 : left curve – spontaneous, right curve – time till x0 is reached is in Fig. 8 and summary statistics in Table 4.
stimulated. These results practically coincide with the estimation of the
refractory period based on ISIs only, which was estimated around
3 ms (Hampel and Lansky, 2008), taking into account that the
last parameter which we determine is x0 . We can see in Fig. 7 that refractory period should be longer than the time to the minimum
the shapes of the densities of the estimates do not differ and the depolarization.
range remains around 10 mV for both cases. The median value is
lower in the case of spontaneous record. As mentioned, this can
be explained by the fact, that the input signal contributes to the 4.2.5. Firing regimen
membrane potential during the refractory period. See Table 3 for As mentioned in Section 2, the key issue for the behavior of
the descriptive statistics of x0 . model (1) is the mutual position between S  x0 and l=b. To check
the threshold regime we compute S  x0 and l ^ =b, where for l
^ we
take the vector of estimated values, S, b are medians of the esti-
4.2.4. Time to reach the resting potential mates taken from the spontaneous part of the data and x0 from
We also measure how much time it takes for the membrane po- the corresponding part (spontaneous or stimulated). In Fig. 9 we
tential to reach the resting level after the spike. It is not a param- see that asymptotic mean depolarization is always higher (median
eter of model (1), but such knowledge could help to judge how is 43 mV) than threshold value in the case of stimulated data,
realistic the model is. In particular we take the interval which which is in a good agreement with our expectation.
 Author's personal copy

P. Lansky et al. / Journal of Physiology - Paris 104 (2010) 160–166 165

200 Table 5
Stimulated Descriptive statistics for ISIs.
180 Spontaneous
ISIs [ms] Stimulated Spontaneous
160
Min 3.900 88.5
140 Max 24.750 5090.4
120 Median 8.400 584.6
Density

Mean 9.033 871.9

100 CV 0.345 0.883

80

60
sity of the spontaneous ISIs suggests that they are generated in
40
accordance with the exponential distribution (Kolmogorov–Smir-
20 nov test does not reject the hypothesis of exponentiality at 5% sig-
0 nificance level), which may imply the Poissonian firing regime. The
0 10 20 30 40 50 60 70 shape of the density for the stimulated ISIs suggests gamma distri-
Asymptotic depolarization [mV]
bution (Kolmogorov–Smirnov test does not reject the hypothesis of
Fig. 9. Estimations of the densities of the asymptotic depolarization with regard to
gamma distribution at 5% a significance level).
the threshold: left curve – spontaneous, right curve – stimulated. The values The descriptive statistics of the ISIs are in Table 5. We can see
(vertical lines) of S  x0 ¼ 13 mV for the spontaneous case and 9.5 mV for that there is a clear distinction between the spike frequency for
stimulated case divides the supra and subthreshold regimen. stimulated part fstim ¼ 110 Hz and for spontaneous part fspon ¼
1:14 Hz, for median the values are fstim ¼ 119 Hz; f spon ¼ 1:71 Hz.

4.3. ISI distribution

4.3.1. Comparison of experimental and theoretical distribution
Using the methods described above we got 83 first ISIs com- To compare experimental ISI distributions with the Ornstein–
pletely within the stimulated parts and 312 ISIs in the spontaneous Uhlenbeck first-passage-time distribution, we simulated model
parts. In Fig. 10 we compare their distributions. The empirical den- (1) with estimated parameters. The results are in Fig. 11.

1.4 250
Spontaneous Stimulated
1.2
200
1

0.8 150
Density

Density

0.6
100

0.4
50
0.2

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.005 0.01 0.015 0.02 0.025 0.03
ISI length [s] ISI length [s]

Fig. 10. Distributions of ISIs lengths: the left panel – spontaneous ISIs, the dotted line represents the fitted exponential ðf ðxÞ ¼ kekx ; k ¼ 1:14Þ and gamma
a bx
ðf ðxÞ ¼ xa1 bCeðaÞ ; a ¼ 1:56; b ¼ 1:79Þ distribution; the right panel – stimulated ISIs, the dotted line represents the fitted gamma distribution ða ¼ 9:45; b ¼ 1043:3Þ.

7 250
Spontaneous Stimulated
6
200
5

150
4
Density

Density

3
100

2
50
1

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.005 0.01 0.015 0.02 0.025 0.03
ISI length [s] ISI length [s]

Fig. 11. Distributions of the first-passage-times obtained from Eq. (1) with parameters estimated from the trajectories: the left panel – spontaneous, the right panel –
stimulated. The solid lines are estimated densities of experimentally obtained ISIs (histograms), the dotted lines are estimated densities from the simulation of first-passage-
times of model (103 runs).
 Author's personal copy

166 P. Lansky et al. / Journal of Physiology - Paris 104 (2010) 160–166

The fit of the observed ISIs to the model is perfect for the stim- Feigin, P., 1976. Maximum likelihood estimation for continuous-time stochastic
processes. Advances in Applied Probability 8, 712–736.
ulated activity. It is not the case for the spontaneous activity,
Gerstner, W., Kistler, W., 2002. Spiking Neuron Models. Cambridge University Press,
where from the model we obtained distribution which is more nar- New York.
row and shifted in direction of short ISIs. At this moment we are Hampel, D., Lansky, P., 2008. On the estimation of refractory period. Journal of
not able to explain this discrepancy. Neuroscience Methods 171, 288–295.
He, J., 2003. Slow oscillation in non-lemniscal auditory thalamus. Journal of
Neuroscience 23, 8281–8290.
Acknowledgments Inoue, J., Sato, S., Ricciardi, L., 1995. On the parameter estimation for diffusion
models of single neuron’s activities. Biological Cybernetics 73, 209–221.
Jolivet, R., Rauch, A., Luscher, H., Gerstner, W., 2006. Integrate-and-fire models with
AV0Z50110509, Center for Neurosciences LC554, Grant Agency adaptation are good enough: predicting spike times under random current
of the Academy of Sciences of the Czech Republic – Project injection. In: Weiss, Y., Schölkopf, B.P.J. (Eds.), Advances in Neural Information
IAA101120604. Processing Systems, Vol. 18. MIT Press, Cambridge MA, pp. 595–602.
Koyama, S., Kass, R., 2008. Spike train probability models for stimulus-driven leaky
integrate-and-fire neurons. Neural Computation 20, 1776–1795.
References Lansky, P., Ditlevsen, S., 2008. A review of the methods for signal estimation in
stochastic diffusion leaky integrate-and-fire neuronal models. Biological
Bibbona, E., Lansky, P., Sacerdote, L., Sirovich, R., 2008. Errors in estimation of the Cybernetics 99, 253–262.
input signal for integrate-and-fire neuronal models. Physical Review E 78, Lansky, P., Sacerdote, L., 2001. The Ornstein–Uhlenbeck neuronal model with
011918. signal-dependent noise. Physics Letters A 285, 132–140.
Brunel, N., van Rossum, M., 2007. Lapicque’s 1907 paper: from frogs to integrate- Lansky, P., Sanda, P., He, J., 2006. The parameters of the stochastic leaky integrate-
and-fire. Biological Cybernectics 97, 1–3. and-fire neuronal model. Journal of Computational Neuroscience 21, 211–223.
Burkitt, A., 2006. A review of the integrate-and-fire neuron model: I. Homogeneous Mullowney, P., Iyengar, S., 2008. Parameter estimation for a leaky integrate-and-fire
synaptic input. Biological Cybernetics 95, 1–19. neuronal model from ISI data. Journal of Computational Neuroscience 24, 179–
Clopath, C., Jolivet, R., Rauch, A., Lüscher, H., Gerstner, W., 2007. Predicting neuronal 194.
activity with simple models of the threshold type: adaptive exponential Shinomoto, S., Sakai, Y., Funahashi, S., 1999. The Ornstein–Uhlenbeck process does
integrate-and-fire model with two compartments. Neurocomputing 70, 1668– not reproduce spiking statistics of neurons in prefrontal cortex. Neural
1673. Computation 11, 935–951.
Dayan, P., Abbott, L., 2001. Theoretical Neuroscience. MIT Press, Cambridge, CA. Tuckwell, H., 1988. Introduction to Theoretical Neurobiology. Cambridge University
Ditlevsen, S., Lansky, P., 2005. Estimation of the input parameters in the Ornstein– Press, New York.
Uhlenbeck neuronal model. Physical Review E 71, 11907. Xiong, Y., Yu, Y., Fujimoto, K., Chan, Y., He, J., 2006. An in vivo intracellular study of
Ditlevsen, S., Lansky, P., 2006. Estimation of the input parameters in the Feller auditory thalamic neurons. Thalamus and Related Systems 2, 253–260.
neuronal model. Physical Review E 73, 61910. Yu, Y., Xiong, Y., Chan, Y., He, J., 2004. Corticofugal gating of auditory information in
Ditlevsen, S., Lansky, P., 2007. Parameters of stochastic diffusion processes the thalamus: an in vivo intracellular recording study. Journal of Neuroscience
estimated from observations of first-hitting times: application to the leaky 24, 3060–3069.
integrate-and-fire neuronal model. Physical Review E 76, 41906.
 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5

Available online at www.sciencedirect.com

www.elsevier.com/locate/brainres

Research Report

Stochastic interpolation model of the medial superior olive

neural circuit

Pavel Sandaa, c, d,⁎, Petr Marsalekb, c, d

a
Institute of Physiology, Academy of Sciences of the Czech Republic, Videnska 1083, 142 20, Praha 4, Czech Republic
b
Institute of Pathological Physiology, First Medical Faculty, Charles University of Prague, U Nemocnice 5, 128 53, Praha 2, Czech Republic
c
Max Planck Institute for the Physics of Complex Systems, Noethnitzer Str. 38, 011 87, Dresden, Germany
d
Faculty of Biomedical Engineering, Czech Technical University of Prague, Nam. Sitna 3105, 272 01, Kladno, Czech Republic

A R T I C LE I N FO AB S T R A C T

Article history: This article presents a stochastic model of binaural hearing in the medial superior olive
Accepted 19 August 2011 (MSO) circuit. This model is a variant of the slope encoding models. First, a general
Available online 27 August 2011 framework is developed describing the elementary neural operations realized on spike
trains in individual parts of the circuit and how the neurons converging onto the MSO
Keywords: are connected. Random delay, coincidence detection of spikes, divergence and
Coincidence detection convergence of spike trains are operations implemented by the following modules: spike
Directional hearing generator, jitter generator, and coincidence detector. Subsequent processing of spike
Interaural time delay trains computes the sound azimuth in the circuit. The circuit parameters that influence
Sound azimuth efficiency of slope encoding are studied. In order to measure the overall circuit performance
Interpolation model the concept of an ideal observer is used instead of a detailed model of higher relays in the
auditory pathway. This makes it possible to bridge the gap between psychophysical
observations in humans and recordings taken of small rodents. Most of the results are
obtained through numerical simulations of the model.
© 2011 Elsevier B.V. All rights reserved.

1. Introduction Karino et al., 2011). The interpolation model differs from the
model proposing an array of delay lines (Jeffress, 1948). In the Jef-
A model of the mammalian MSO circuit which localizes the fress model, the sound direction is encoded using a spatial code
direction of the sound source in the horizontal plane (azimuth) which includes the position of peak neural activity within the
at low frequency range is presented here. This model is based array. In the interpolation model, the binaural neuron encodes
upon the interpolation of the direction in question using the the sound direction using its firing rate. Firing of this neuron
firing rate code of the MSO binaural neurons. This is proposed constitutes the ITD (interaural time difference) curve. The
as an alternative mechanism for low frequency sound locali- sound direction is read out by interpolation from this curve.
zation in mammals, as the precise mechanism in mammals McAlpine et al. (2001), Brand et al. (2002), Grothe (2003) and
is not known to date (Dean et al., 2008; Joris et al., 1998, 2006; others have proposed an ITD readout mechanism which en-

⁎ Corresponding author at: Institute of Physiology, Academy of Sciences of the Czech Republic, Videnska 1083, 142 20, Praha 4, Czech Republic.
Fax: +420 241 062 488.
E-mail address: [email protected] (P. Sanda).
Abbreviations: AN, auditory nerve; AP, action potential; EE, excitatory–excitatory; EI, excitatory–inhibitory; EPSP, excitatory post-
synaptic potentials; IPSP, inhibitory post-synaptic potentials; ITD, interaural time difference; MSO, medial superior olive

0006-8993/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.brainres.2011.08.048
258 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5

codes the sound direction through the maximal slope of the abstracted in this paper. In vitro experiments conducted in
response function. This function has the ITD as its argument the search for the subcellular mechanism responsible for the
and the firing rate as its output. In line with the terminology sub-millisecond precision show the plurality of underlying
used previously, we can call this mechanism slope encoding. biophysical mechanisms. To name only two of these one par-
Our interpolation model is a variant of this slope encoding ticular mechanism which contributes to setting the duration
mechanism. One of the crucial experimental findings is that of the coincidence detection window is implemented by the
abolishing of inhibitory inputs by strychnine moves the max- voltage-activated potassium current at the postsynaptic
imum slope out of the operational range (Brand et al., 2002). membrane (Mathews et al., 2010). Another mechanism by
Based on this observation, the slope encoding model has, as which the order of the IPSP and EPSP is distinguished is the
one of its key elements, an inhibitory input to the MSO. post-inhibitory rebound mechanism (Brand et al., 2002).
Our model reproduces the MSO computations at the level We also use the ergodic assumption of indistinguishable time
of abstract neural circuit description. Biophysical interactions and population ensemble encoding of the signals by spike trains.
of the excitatory and inhibitory postsynaptic potentials (EPSPs The ergodic assumption in physics was originally introduced in
and IPSPs) at the postsynaptic membrane are not modeled in statistical mechanics of gases by Boltzmann in the nineteenth
detail. Concerning the interaction of the IPSP with the EPSP, century. According to this assumption the average of a selected
their time order is important. The IPSP must precede the quantity is the same for the time average and for the statistical
EPSP to exceed the threshold (Grothe, 2003; Pecka et al., ensemble average (Papoulis, 1991). In other words, it assumes
2008). The PSPs from the two sides have to meet within a that the same result is obtained when observing a process over
given coincidence detection time window. The window size a long period of time as when sampling a given number of
is one of the key parameters in our model. We stress this as- shorter independent realizations of the same process. This ap-
pect of the model because the rise and decay time constants proach is applied in the neural coding description as follows.
of postsynaptic voltage dependent currents determine the co- To receive information about the sound source direction, one
incidence detection window size. The subcellular details are can either wait for a given period of time and measure all the

Fig. 1 – Schemes of the MSO circuit anatomy and the interpolation model. The anatomical scheme (left) shows branches of the
binaural circuit (only one half of the symmetrical circuit is shown). L1, L2,…, R1, R2 are left and right neurons. The direction
from the auditory periphery to the auditory center is oriented top-down. Each neuron is denoted by its nucleus abbreviation
(expanded in the left-hand column). Synapses are either excitatory (E) or inhibitory (I). The model scheme (right) employs
neural circuit simplifications without losing the implementation of the key mechanisms.
 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5 259

spikes in that period on an individual neuron, or one can observe input frequency fin are generated, it can be observed that the
a given number of neurons over a shorter time period and firing rate grows and the ITD curve changes from flat in
achieve the same precision. The number of spikes required for lower frequencies to steep in higher frequencies (Fig. 3, left).
such a measurement was estimated using the Bernoulli process Similar rise of the firing rate with rising input frequency is ob-
statistical model in Marsalek and Lansky (2005). served with the use of more detailed Meddis model, with
We propose a simple model where the basic processing markedly lower output firing rates (Fig. 3, right). Furthermore,
stages of the neural circuit have been replaced by four modules the distance between the ITD peaks becomes smaller with in-
operating in succession (see Fig. 1): a spike generator that repre- creasing fin as the input sound period decreases. The curve off-
sents the input from the auditory nerve (AN), a jitter generator set around the DITD = 0 does not shift while changing fin.
that adds jitter and delays, occurring in the pathway up to the As jitter decreases, the firing rate in the ITD curves in-
MSO, a coincidence detector that represents the first binaural creases accordingly. See Fig. 4, right. This is due to the fact
neuron in the MSO, and an ideal observer module. The detailed that the smaller jitter helps spikes to occur in the same coin-
description of these modules can be found in Section 4. cidence window. For TJ → 0, we would obtain a rectangular
pulse, which is unusable for interpolation. If the jitter be-
comes larger than the sound period, this would render the re-
2. Results sponse flat (not shown, effects of higher jitter values on ITD
curve are thoroughly demonstrated by simulations in Supple-
2.1. Effects of inhibition mentary information). Thus, the feasible jitter values are
bounded on both sides. Changes in jitter TJ values in the jitter
For the basic set of parameters (see Table 1), we obtained the generator do not affect the ITD curve shift.
ITD curve shown in Fig. 2. It was checked that disabling the
asymmetric rule in the coincidence detector affects the output 2.2.1. Window width and ITD curve shift
of the circuit. We consider this to be our counterpart of inhibi- The ITD curve shift in our model is produced by the asymmetric
tion suppression in the physiological model (see Section 2.3). coincidence rule but its particular size is affected by the param-
The result was as expected: in the purely excitatory case, i.e., eters of the circuit. Therefore each parameter of our circuit was
with no inhibition, there is a higher firing frequency and, due studied to determine which parameters influence the ITD peak
to the symmetric processing of excitatory spikes, the curve shift. As a result of the simulations, it was possible to observe
peaks at DITD = 0. In all other simulations we used the asymmet- that the parameter that changes the peak shift considerably is
ric coincidence detection rule only. the coincidence detection window width wCD. See Fig. 4. Other
parameters of our model did not affect the peak shift signifi-
2.2. Effects of varying time parameters cantly. See vertical lines in Figs. 3, 4.

After testing the basic parameter set, let us move to exploring 2.3. Ideal observer
the parameter space, beginning with the timing jitter and
input sound frequency. When the ITD curves for increasing The ideal observer module is connected to the output of the
coincidence detector. For interpolation from the firing rate
back to the DITD value a previously fitted ITD function is
140 Excitatory used. The ITD readout slope can be fitted to a sine function:
Inhibitory
y ¼ 12 Að1 þ sinðBt þ CÞ þ DÞ; where A is the multiplicative con-
130
stant setting up the maximum firing rate, B is dependent on
the input frequency, C is the phase shift and D is the sponta-
Firing rate [AP/s]

120
neous activity. See Fig. 5.
110 The shape of this function used for interpolation deter-
mines how precise the computation of DITD encoded in the in-
100 coming spike train is. As shown in previous sections, two
features are important. The first feature is the unambiguous
90
assignment of the unique DITD value from the firing rate. It is
the asymmetric rule together with wCD that plays an important
80
role in determining the readout curve whose fit is shown in
-0.6 -0.3 0 0.3 0.6 Fig. 5. The second feature is the slope of the interpolation curve
ITD [ms] obtained as the difference ΔITD = ΔDITD/Δt and output range of
the ITD curve. Both are decisive for accuracy (or speed) of ITD es-
Fig. 2 – ITD curves for the basic set of parameters. This shows timate by the observer. As the output range increases, we obtain
a detailed view of the peak with inhibition shifted out of a more precise azimuth estimate, which is influenced by all the
the point, where DITD = 0 as compared to the peak without parameters, with jitter TJ being the main factor, as shown above.
inhibition. The relative distances of the two curves (30–50 Hz)
are given in Grothe (2003). Further, the EE interaction has a 2.3.1. Asymptotic times
higher maximum firing rate than the EI interaction. The Up to now we have looked only at the circuit output in the
basic ITD curve has shift of 0.25 ms, which is an emergent terms of the firing rate. At the end, the information contained
property with no free parameter tuning. in the output spike train is used in higher stages of the
260 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5

0.3 0.3
250 25

200 20
Firing rate [AP/s]

Firing rate [AP/s]

150 15

100 10

50 5

0 0
-6 -3 0 3 6 -6 -3 0 3 6
ITD [ms] ITD [ms]

Fig. 3 – ITD response curves for different input frequencies fin directly delivered to the circuit (left-hand panel) and delivered
pffiffiffi
through the Meddis model (right-hand panel). For increasing the sound frequencies with step 2 (50, 71, 100, 141, 200, 283 Hz),
the maximum firing also increases. In both cases, the period of the ITD curve decreases as the input frequency increases. If
cochlear filtering is used, it can be seen that the width of a single peak increases as the input frequency decreases. Secondly,
the output firing is approximately ten times lower with the Meddis model involved. For each sound frequency shown, the
Meddis model was used with the same characteristic frequency. For cut-off frequencies see Supplementary information.

auditory pathway to estimate the azimuth. Since we do not towards the source DITD value when basic jitter value (TJ = 1
model higher stages of auditory processing, we use the ob- ms) is applied (left), and a comparison of the asymptotic
server module, which sequentially collects information from times for different parameter sets (right). In all cases shown
the circuit output and predicts resulting ITD from the infor- here we set DITD = 0. A comparison between a pure tone and
mation given. Naturally, the longer the spike train is mea- a noisy input produced by the Meddis model was added to
sured for, the more accurate the estimate of the source's the Supplementary information.
DITD which can be assessed. Thus in order to investigate the The time to reach the asymptotic values TA, which range
overall estimate performance of our model circuit, we fed it from 0.5 s up to 5 s (see Fig. 6), would be too long compared
with a longer input and estimated the time TA needed for to the typical reaction time. The human minimum integration
the ideal observer to achieve accuracy 4°, in the head-on di- time is discussed in Middlebrooks and Green (1991) as a range
rection (Mills, 1972). Fig. 6 shows fast asymptotic convergence of values: while a stimulus of 50 ms duration is detected with

0.3
140
140
1.86
120 TJ = 0.5 ms
120 600µs

100
Firing rate [AP/s]

100 300µs
Firing rate [AP/s]

1.56
80 80

TJ = 1.5 ms
60 60

40 1.26 40

20 20
0.06 0.36 0.66 0.96

0 0
0 0.5 1 1.5 -3 0 3
ITD [ms] ITD [ms]

Fig. 4 – Dependence of ITD peak shift on the coincidence detection window width wCD (left-hand panel) and different values of
jitter TJ (right-hand panel). Peaks do not exceed 140 Hz due to the input frequency fin = 140 Hz in the basic set of parameters. In
the left-hand panel a shift of the peak towards higher ITDs can be seen as the window size increases from 60 μs towards 2 ms
(with step 300 μs). The two top left intervals show physiologically relevant range in a domestic cat (300 μs) and in a human
(600 μs), respectively. From around wCD = 0.6 ms, and above that value, the maximum slope is observed and the shape of
the ITD function in physiologically relevant range does not change. For further notes about physiological relevance of this
observation see the Discussion section. Right-hand panel: larger jitter causes lower output dynamic range and wider curves.
The presence of high jitter values renders the response curve flat. The response increases as the magnitude of jitter decreases
towards zero.
 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5 261

110
minimum time required by human listeners to determine the
100 DITD, when the entire brainstem circuit is working in parallel
90 and cooperating on the ITD estimation. When the basic set of
parameters is used we obtain that approximately 4 parallel neu-
80
Firing rate [AP/s]

rons would be needed in order to achieve the 150 ms integration

70
time.
60

50
3. Discussion
40

30 3.1. Type of neuronal coding

20 ITD curve
Fitted curve
10
Jeffress (1948) proposed a theory of how sound is localized by
-0.3 -0.2 -0.1 0 0.1 0.2 0.3 a particular neural circuit. According to this concept, neural
ITD [ms] representation divides the whole azimuthal space into dis-
tinct sections. Each section is handled by a specific neuron
Fig. 5 – The interpolation ITD curve in the physiologically and the fact that the sound originates from a given direction
relevant range of values. The ITD curve is shown together is represented by higher firing activity of the corresponding
with the fitted sinus function used for interpolation in the neuron. Jeffress suggested that this circuit is anatomically re-
ideal observer. This function is fixed for the set of parameters alized by an array of delay lines. Indeed, experimental evi-
which would correspond to the fixed tuning of the neural dence supporting this hypothesis was later found in the
circuit for selected input signals. In this example, we used circuit of the nucleus laminaris of birds by Carr and Konishi
the ITD curve, which results from simulation of the circuit for (1988).
the basic set of parameters. Note that the period of this slope The responses of higher nuclei such as the colliculus inferior
corresponds to the fundamental sound period. in both birds, Wagner et al. (1987, 2007) and mammals, Joris et al.
(1998, 2006) have been systematically studied until recently.
errors, typical azimuth integration time ranges from 150 to This is because the colliculus inferior is more accessible to re-
300 ms of sound duration. However, the pathway from the cordings and its binaural response is analogous to that of the
AN to the MSO is highly parallel and consists of up to a few nucleus laminaris or the MSO, respectively. In the MSO, which
hundred neurons. When the ergodic hypothesis is used, as dis- is a mammalian anatomical counterpart of the avian circuit,
cussed in Section 3.4, TA can be divided by the number of neu- there is little supporting evidence for the existence of a structure
rons. The time thus obtained corresponds to the theoretical similar to such an array of delay lines, Grothe (2003), but see also

0.65
0.00014 Mean
0.0001
Mean ± SD TJ = 5 ms
0.00012
Physiological accuracy
0.0001 ITD of stimulus o
1e-05 +2
8e-05
Mean+SD
ITD [s]
ITD [s]

6e-05 1e-06
wCD = 3 ms
TA
4e-05 Mean

2e-05 1e-07
+2o
DITD=0
−2o 1e-08
TJ = 1 ms
-2e-05
Mean−SD wCD = 0.6 ms
-4e-05
1e-09
0.01 0.1 1 10 100 0.01 0.1 1 10 100
Time [s] Time [s]

Fig. 6 – Time required to reach a reliable estimate (with a given precision) of azimuth TA by a single binaural neuron. The x-axis
shows time and the y-axis shows the ITD. Note the logarithmic scales used. In the left-hand panel, the middle solid line represents
the mean value obtained for 1000 simulation runs, while the dotted lines show standard deviation. The three dotted horizontal
lines show DITD original source and margins of azimuth difference ± 2° respectively. In the right-hand panel, the top line shows
mean convergence when using larger jitter (TJ = 5 ms), the middle line shows how convergence improves when using a wider
window (wCD = 3 ms), which is different from the basic parameter set, and the bottom line shows faster convergence for the basic
set of parameters (TJ = 1 ms). The horizontal lines show the precision obtained in the psychoacoustical experiments. At the point in
time TA (TA = 0.65 s for basic set of parameters) when the mean value crosses the precision line, enough information has been
obtained for our model to be able to report the azimuth with average precision corresponding to that of human listeners ([− 2° ; 2°]).
262 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5

Joris et al. (1998). This raises the question of whether such a type values (Jercog et al., 2010). Using smaller jitter values proved
of neural coding is also to be found in the mammalian brain much more effective in increasing the output range and
stem. Various alternatives have been proposed, for example, steepness of the ITD curve. As noted in Section 4.3, on the
delays of the cochlear traveling wave (Jennings and Colburn, one hand, jitter plays a positive role for more subtle differen-
2010; Joris et al., 2006; Shamma et al., 1989). Previous experi- tiation of azimuthal space, but on the other hand, higher jitter
ments have shown that ITD peaks occur out of the relevant reduces coding efficiency. This indicates the possible exis-
physiological range (McAlpine et al., 2001) and further reproduc- tence of an optimal jitter value which merits further research.
tions of mammalian experiments similar to those have led to Asymptotic behavior of the sound direction “ideally ob-
the development of a theory showing that the mammalian lo- served” by the model was also investigated. This gives us
calization mechanism is different from that of the avian. The overall precision and time performance for the whole circuit.
importance of the role of synaptic inhibition was also experi- It could be seen that an ideal observer located on a single
mentally discovered by Brand et al. (2002). The shift of the ITD fiber performs in the range of seconds, which is acceptable
curve peak is somewhat counter-intuitive when we assume considering that the auditory pathway works in parallel
that the Jeffress model is employed, and thus the slope (rate) using tens to hundreds of neurons. Using an ergodic assump-
coding model has been proposed instead. In this paper a circuit tion for the ideal observer, we obtain a hundreds of milli-
based on the slope coding scheme is studied. seconds time range. This allows us to connect physiological
and psychoacoustical observations, which are normally two
3.2. Asymmetric rule for coincidence detection separate areas of research.
Besides the regular input, the Meddis (2006) model of the audi-
The excitatory and inhibitory effects in the sound localization tory periphery was used to compare the performance of the two
circuit have been extensively discussed in experimental liter- input models. Generally, the Meddis model causes a fall in the
ature. Classical electrophysiological works by Goldberg and firing rate delivered as the input to our model, compared to the
Brown (1969) and others have distinguished between the EE regular spiking input we used. In addition, only phase-locked
and EI type of binaural unit response. The EE (excitatory– ex- spikes will be registered by coincidence detection. This reduces
citatory) abbreviation labels units where the sound level in- both the available dynamic range of values of the ITD curve and
crement at both sides has an excitatory effect, while EI the computational performance of the circuit. The question re-
stands for the excitatory effect of one side and inhibitory ef- mains how closely this model corresponds to the real input to
fect of the other side. However, this is a phenomenological the MSO, since the model produces only the spike train occurring
classification and does not imply the synaptic mechanisms. at the auditory fiber. The cochlear nucleus can improve the
In the mammalian MSO circuit, other inhibitory inputs be- phase-locking due to summation of more AN fiber inputs.
sides those from the MNTB (Medial Nucleus of Trapezoid The timing jitter also plays a role in deteriorating the phase-
Body) are involved. Each sensory pathway, including the audi- locking precision. Most of the auditory processing stages blur
tory pathway, also contains local inhibitory interneurons and the phase locking. The comparison with the Meddis model indi-
often other inhibitory circuits, like, for example the indirect cates that introducing more processing stages into the circuit
pathway of the basal ganglia. In all these synaptic circuits, does not necessarily blur the spike trains beyond the state
the role of inhibition is to dynamically set up the gain of the when the circuit would no longer function and would no longer
circuit via feedback connections. Other inhibitory synapses possess a functioning coincidence detection mechanism. In-
in the MSO circuit contribute to setting the gain, adaptation, deed, Reed et al. (2002) have shown that the phase locking in
and perhaps to further unknown functions of the circuit. a model neuron within a neural circuit can deteriorate, can be
Our model follows anatomical connections of the feedfor- sharpened, or can stay unaltered, depending upon the connec-
ward sound localization pathway. Although throughout the tion divergence and other simple properties of the neuron as a
paper excitatory and inhibitory synapses are referred to relay within the sensory pathway.
according to their anatomical and functional classification,
we have shown that instead of synapse polarity, it is possible 3.4. Ergodic assumption
to use a timing rule for coincidence detection. Thus, we have
been able to describe the circuitry independent of the synapses The performance of a single binaural neuron has been shown
polarity in more abstract and simple terms. We have also for several parameters and inputs. Time (TA) to reach the esti-
demonstrated that our results are in line with the experimen- mate of DITD by single neuron is higher than in psychoacoustic
tal results related to inhibition blocking (Brand et al., 2002; experiments. It is known that the auditory pathway uses AN
Pecka et al., 2008). fibers and subsequent neurons in parallel so the brainstem
contains several parallel circuits cooperating on the computa-
3.3. Properties of the circuit tion of azimuth. We used the ergodic hypothesis to divide TA
by the number of parallel neurons. This hypothesis used in in-
Having studied the parameters that affect important proper- terpretation of Fig. 6 states that the same result is obtained
ties of the ITD curve, we have found that the parameter signif- when counting spikes for a long time as when sampling
icantly affecting peak position is the coincidence window them across a given number of parallel fibers. This is also con-
width. One consequence of shifting the peak by employing a sistent with the actual ITD processing in the MSO reported in
wider window is a steeper ITD function in the physiological several species, which produced the same ITD tuning curves
range of ITDs. Bringing steepness to its maximum value by several repeats of sound stimulus, as shown, for instance
would require wCD ≈ 0.6 ms which is higher than the observed by Goldberg and Brown (1969).
 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5 263

The exact number of circuits which work in parallel is dif- small random delay value defined as D = TJ(B(a, b) − 0.5), where
ficult to estimate. It concerns not only the number of the bin- B(a, b) is a random variable from the beta distribution with pa-
aural MSO neurons but also the number and connectivity of rameters a = 2, b = 4 and TJ is the timing jitter magnitude.
the spherical bushy cells of the cochlear nucleus. About 7000 Values of a = 2 and b = 4 are the smallest nontrivial values
of these cells connect to the MSO in a domestic cat. Since resulting in a nonzero skewness of the beta probability densi-
their axons branch, the average number of connections to ty function, as discussed in Drapal and Marsalek (2010). Note
the MSO is difficult to obtain with the current methodology the positive role of this type of noise in the circuit — it is be-
(P.X. Joris, personal communication). Next, the exact spectral cause of this noise that more refined azimuth precision is
content of incoming acoustic stimuli affects the number of in- achieved.
volved fibers due to the tonotopic organization. For example
the broadband stimuli would be expected to perform better 4.3. Coincidence detector
than pure tone stimuli due to higher number of parallel fibers
transmitting the signal. All these problems were not studied Whenever two spikes are detected within the short time win-
deeper in the current paper but they pose interesting open dow, the coincidence detector neuron fires. In the classical
questions for further work. Jeffress (1948) model, coincidence detection is realized by the
excitatory spikes converging on a binaural neuron from both
3.5. Stochastic coding of ITD sides simultaneously in a short time succession. The mere
fact of coincidence determines the azimuth defined by the
Finally, we would like to point out the use of the built-in random position of the neuron in a delay line. The interpolation model
number generator in our model. By means of these stochastic also introduces a short time window for detecting consequent
simulations we have obtained realizations of a random variable, spikes, while distinguishing between two EPSPs and a pair of
namely of the timing jitter. The results obtained for the ideal one EPSP and one IPSP. The preceding application of the timing
observer are shown together with their statistical description. jitter is of essential importance to the circuit here as it allows
A previous paper by Marsalek and Lansky (2005) presented a a more subtle distribution of coincidences. Instead of a binary
stochastic model of a coincidence detector and theoretically answer about two spikes being in the coincidence window, a
calculated average detection times. In this paper a further step probabilistic variable is obtained which, in turn, allows a finer
has been taken, as the entire MSO model circuit has been studied. distribution of recognized ITD values. Compared to the “all or
The precision of the ITD detection assessed by our model re- none” output of the Jeffress model, this means that one single
quires further verification by future experiments employing psy- binaural neuron is already capable of determining DITD. How-
choacoustical data. ever a longer spike train is needed to achieve the necessary pre-
cision of the estimate.
The important property of the ITD curve is its peak shift as
4. Experimental procedures observed in an experimental study of mammals (Brand et al.,
2002). First, this leads to a shift of possible ITD values inside
4.1. Spike generator the physiologically relevant range. Second, it allows the un-
ambiguous interpolation of a particular firing rate to a single
In some simulations a constant frequency spike train was ITD value. This is in contrast to the situation where the ITD
used as the input, with two spikes generated in each cycle — curve has a peak at DITD = 0 and the concave shape allows
one to one side, the other delayed by the ITD (this variable is two possible ITDs to be selected. We show that in order to
denoted by DITD) to the other side. achieve this ITD curve shift, we need an “asymmetric” rule
While the regular spiking input is suitable for analysis of cir- of EPSP and IPSP coincidence detection, allowing incoming
cuit parameter properties, a more biologically plausible model of spikes to trigger an output spike only if the inhibitory spike pre-
spike generation was also included. Specifically, the first part of cedes the excitatory spike. This corresponds to the detailed bio-
the computational model of the auditory periphery as described physics of the PSPs. Pharmacological suppression of the
in Meddis (2006) was used. Here, the complex acoustic stimuli inhibition is equivalent in our model to disabling this asymmet-
are passed through successive stages of processing: stapes, bas- ric rule and letting the symmetric coincidence detection operate
ilar membrane, inner hair cell receptor potential, presynaptic on the incoming EPSPs only. By “asymmetry” we mean that the
calcium currents and transmitter release events, AN (auditory order of excitatory and inhibitory spikes matters — in contrast
nerve) synapse and finally AN spiking response, including re- to the excitatory only situation.
fractory effects. The output – spike train in the auditory nerve
fiber with selected characteristic frequency – is subsequently 4.4. Ideal observer module
used as an input to our circuit. An extensive overview of all the
parameters can be found in the appendix under Meddis (2006). Compared to the Jeffress model where the azimuth is deter-
This model was then used without further modifications. mined by the position of the firing neuron in the delay line,
the interpolation model relies on the value of a firing rate.
4.2. Jitter generator The firing rate is realized by random variations in the input
spike timings, and one way to eliminate this random compo-
In the next processing stage, jitter is generated and added to nent is to take an average of the firing rates over a given
the spike times, simulating natural noise occurring in the time. Because the input sound sequences may only last for a
neural spike transmission. Each spike is shifted in time by a short period of time, we would like to know what length of
264 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5

time intervals are needed in order to obtain an azimuth esti- Table 1 – The basic set parameters.
mate. Our aim here is not to model the higher stages of spatial Parameter Value
sound localization, but rather to estimate the information
TJ 1 ms
that is available after the MSO processing stage and to connect
wCD 600 μs
the physiologically relevant information (firing rate) with ex-
fin 140 Hz
perimental azimuth measurements. T 500 s
The concept of the ideal observer, originally also known as
the ideal receiver, is frequently used in the signal detection
theory and psychophysics (Tanner, 1961). By implementing fundamental frequency. (We get 1/wCD = 1667 Hz, therefore it
the ideal observer module, we used this concept to describe must be fin < 1667 Hz.) This latter assumption is also frequently
a mechanism which represents the function of the next pro- mentioned in other models. The subcellular experimental
cessing stages of the auditory pathway. It should be noted studies also measured values of relevant postsynaptic
that this is not strictly an ideal observer analysis. Unlike pre- current- and voltage-dependent ionic current rise time. The
vious modules, this module does not process the data. In- values measured give window width in the sub-millisecond
stead, it simply extracts the available information encoded range close to this value (Jercog et al., 2010). Representative
in the output spike train from the binaural neuron, and in par- fundamental sound frequency in the following results is
ticular it reads the sound source direction from the MSO neu- fin = 140 Hz. Besides the fundamental frequency of spike gener-
ron. This is done via the interpolation of the firing rates in ator fin, we set the duration of most simulations to 500 s in
the ITD curve. This lookup (calibration) function must be order to reach a steady state. We select the parameters
known beforehand and it can be approximated by a sine func- above as a basic set, summarized in Table 1. An extensive list
tion RITD ≈ Rmaxsin2πfinDITD. The meaning of the parameters of of circuit parameters can be found in the Supplementary in-
this approximation are listed in the Results section. formation. In all the simulations we use this basic set of the
Since the ideal observer has complete access to perceptual parameters unless stated otherwise.
information, the azimuth perception can be obtained. In psy- The physiological range of DITD in humans when localizing
chophysics this precision is called the just noticeable differ- sound source in one quadrant (0–90°) ranges from 0 to 600 μs.
ence, and in azimuthal sound localization in the head-on This can be calculated from the sound speed in the air and the
direction is 4° (angular degrees). This has been observed with- distance between the ears. Differences in the ear distance of
in a decade variation amongst distant animal species, in smaller animals have to be taken into account when inter-
humans, Mills (1972), guinea pigs, McAlpine et al. (1996) barn preting the data recorded in various laboratory animals.
owls, Moiseff and Konishi (1981) and even in parasitic flies Supplementary materials related to this article can be
Ormia ochracea, Miles et al. (1995). Here, different azimuths found online at doi:10.1016/j.brainres.2011.08.048.
are distinguished as distinct points on the ITD curve along
its slope. Thanks to interpolation the just noticeable differ-
ence is obtained from the difference between two spike train
Acknowledgments
firing rates elicited by two stimuli (azimuths) φ1 and φ2. The
firing rates of spike trains in response to stimuli φ1 and φ2
Supported by research initiatives MSM 0021620806 and MSM
denoted as R1 and R2 can be described as random variables
6840770012 by the Ministry of Education, Youth and Sports,
with means μ1 and μ2 and standard deviation σ (we assume
by MPO FR-TI3/869 by the Ministry of Industry and Trade of
σ1 = σ2). Detection distance is defined as: d = (μ2 − μ1)/σ. When
the Czech Republic, by P103/11/0282 by the Grant Agency of
t → ∞, firing rates converge to their means, R1 →μ1 and R2 →μ2.
the Czech Republic, and by the Max Planck Society. Thanks
The computational precision performance of the circuit is depen-
to Philip Joris, Petr Lansky, David McAlpine and Raymond
dent upon speed of this convergence. Thus we measure the time
Meddis for discussions and to Deborah A. M. James, Martina
to reach the direction estimate TA. This variable is mainly depen-
Missikova and Linda Jayne Turner for copy-editing.
dent on the firing rate RITD and input sound frequency fin, as well
as other parameters.
REFERENCES
4.5. Simulations and parameters

The model circuit was simulated numerically with explora- Beckius, G.E., Batra, R., Oliver, D.L., 1999. Axons from anteroventral
cochlear nucleus that terminate in medial superior olive of cat:
tion of plausible parameter space. Each of the circuit modules
observations related to delay lines. J. Neurosci. 19 (8), 3146–3161.
has its own set of parameters which together define the cir- Brand, A., Behrend, O., Marquardt, T., McAlpine, D., Grothe, B.,
cuit parameters. A default value of jitter of TJ = 1 ms was 2002. Precise inhibition is essential for microsecond interaural
assigned since the experimentally observed values of jitter time difference coding. Nature 417 (6888), 543–547.
do not exceed 1 ms as shown for example in the experiment Carr, C.E., Konishi, M., 1988. Axonal delay lines for time measurement
involving the octopus type cells of the cochlear nucleus in a in the owl's brainstem. Proc. Natl. Acad. Sci. U S A 85 (21),
8311–8315.
domestic cat (Oertel et al., 2000). The width of the coincidence
Dean, I., Robinson, B.L., Harper, N.S., McAlpine, D., 2008. Rapid neural
window wCD has been assigned as 600 μs. This is in agreement
adaptation to sound level statistics. J. Neurosci. 28 (25), 6430–6438.
with the value used in a model part of Beckius et al. (1999). Drapal, M., Marsalek, P., 2010. Stochastic model shows how cochlear
Briefly, their model performed well as long as the coincidence implants process azimuth in real auditory space. Chin. J. Physiol.
detection window was shorter than the period of the 53 (6), 439–446.
 B RA I N R ES E A RC H 1 4 3 4 ( 2 01 2 ) 2 5 7 –26 5 265

Goldberg, J.M., Brown, P.B., 1969. Response of binaural neurons of Meddis, R., 2006. Auditory-nerve first-spike latency and auditory
dog superior olivary complex to dichotic tonal stimuli: some absolute threshold: a computer model. J. Acoust. Soc. Am. 119 (1),
physiological mechanisms of sound localization. J. Neurophysiol. 406–417.
32 (4), 613–636. Middlebrooks, J.C., Green, D.M., 1991. Sound localization by
Grothe, B., 2003. New roles for synaptic inhibition in sound human listeners. Annu. Rev. Psychol. 42 (1), 135–159.
localization. Nat. Rev. Neurosci. 4 (7), 540–550. Miles, R., Robert, D., Hoy, R., 1995. Mechanically coupled ears for
Jeffress, L.A., 1948. A place theory of sound localization. J. Comp. directional hearing in the parasitoid fly Ormia ochracea. J.
Physiol. Psychol. 41 (1), 35–39. Acoust. Soc. Am. 98 (6), 3059–3070.
Jennings, T.R., Colburn, H.S., 2010. Models of the superior olivary Mills, A.W., 1972. Auditory localization. In: Tobias, J.V. (Ed.),
complex. In: Meddis, R., Lopez-Poveda, E.A., Fay, R.R., Popper, Foundations of Modern Auditory Theory. Academic Press, New
A.N. (Eds.), Computational Models of the Auditory System. York, pp. 303–348.
Springer, New York, pp. 65–96. Moiseff, A., Konishi, M., 1981. Neuronal and behavioral sensitivity
Jercog, P.E., Svirskis, G., Kotak, V.C., Sanes, D.H., Rinzel, J., 2010. to binaural time differences in the owl. J. Neurosci. 1 (1), 40–48.
Asymmetric excitatory synaptic dynamics underlie interaural Oertel, D., Bal, R., Gardner, S.M., Smith, P.H., Joris, P.X., 2000.
time difference processing in the auditory system. PLoS Biol. etection of synchrony in the activity of auditory nerve fibers by
8 (6), e1000406 1–9. octopus cells of the mammalian cochlear nucleus. Proc. Natl.
Joris, P.X., Smith, P.H., Yin, T.C.T., 1998. Coincidence detection in the Acad. Sci. U S A 97 (22), 11773–11779.
auditory system: 50 years after Jeffress. Neuron 21 (6), 1235–1238. Papoulis, A., 1991. Probability, Random Variables, and Stochastic
Joris, P.X., Van de Sande, B., Louage, D.H., van der Heijden, M., Processes. McGraw-Hill, New York.
2006. Binaural and cochlear disparities. Proc. Natl. Acad. Sci. Pecka, M., Brand, A., Behrend, O., Grothe, B., 2008. Interaural time
U S A 103 (34), 12917–12922. difference processing in the mammalian medial superior olive:
Karino, S., Smith, P., Yin, T., Joris, P., 2011. Axonal branching the role of glycinergic inhibition. J. Neurosci. 28 (27), 6914–6925.
patterns as sources of delay in the mammalian auditory Reed, M.C., Blum, J.J., Mitchell, C.C., 2002. Precision of neural
brainstem: a re-examination. J. Neurosci. 31 (8), 3016–3031. timing: effects of convergence and time-windowing. J. Comput.
Marsalek, P., Lansky, P., 2005. Proposed mechanisms for Neurosci. 13 (1), 35–47.
coincidence detection in the auditory brainstem. Biol. Cybern. Shamma, S., Shen, N., Gopalaswamy, P., 1989. Stereausis: binaural
92 (6), 445–451. processing without neural delays. J. Acoust. Soc. Am. 86 (3),
Mathews, P., Jercog, P., Rinzel, J., Scott, L., Golding, N., 2010. Control 989–1006.
of submillisecond synaptic timing in binaural coincidence Tanner Jr., W.P., 1961. Physiological implications of psychophysical
detectors by K(v)1 channels. Nat. Neurosci. 13 (5), 601–609. data. Ann. N.Y. Acad. Sci. 89 (5), 752–765.
McAlpine, D., Jiang, D., Palmer, A., 1996. Interaural delay Wagner, H., Takahashi, T., Konishi, M., 1987. Representation of
sensitivity and the classification of low best-frequency interaural time difference in the central nucleus of the barn
binaural responses in the inferior colliculus of the guinea pig. owl's inferior colliculus. J. Neurosci. 7 (10), 3105–3116.
Hear. Res. 97 (1–2), 136–152. Wagner, H., Asadollahi, A., Bremen, P., Endler, F., Vonderschen, K.,
McAlpine, D., Jiang, D., Palmer, A.R., 2001. A neural code for von Campenhausen, M., 2007. Distribution of interaural time
low-frequency sound localization in mammals. Nat. Neurosci. difference in the barn owl's inferior colliculus in the low- and
4 (4), 396–401. high-frequency ranges. J. Neurosci. 27 (15), 4191–4200.
P. Šanda: Speeding Up the Algorithm for Finding Optimal Kernel Bandwidth in Spike Train Analysis, en 73-75 en73

Speeding Up the Algorithm for Finding Optimal Kernel Bandwidth

in Spike Train Analysis
P. Šanda1
1
Institute of Physiology, Academy of Sciences of the Czech Republic
Supervisor: Doc. RNDr. Petr Lánský, CSc.

Summary data differs from paper to paper and spike train with a specific kernel, thus
One of the important tasks in the spike train various methods were proposed [2]. obtaining a smooth estimate of l, for
analysis is to estimate the underlying firing example see Fig. 1. In this case we use
rate function. The aim of this article is to Here we consider the method based on the fixed Gaussian kernel
improve the time performance of an convolution of a spike train with a fixed
algorithm which can be used for the (Gaussian) kernel, which in result leads to
estimation. a smooth estimate of firing rate and has
As there is no unique way how to infer the been widely used in the past decades [3],
firing rate function, several different [4], [5], [6], [7], [8]. The most difficult part of and the problem is reduced to the question
methods have been proposed. A popular this method is the selection of the kernel how to select the optimal bandwidth w, so
Ů
method how to estimate this function is the bandwidth, because it affects substantially that the difference between l and l is
convolution of the spike train with the “quality” of the estimate, while there is minimal. The method itself is beyond the
Gaussian kernel with appropriate kernel no obvious clue how the optimal bandwidth scope of this article, for details see [1].
bandwidth. The definition of what should be chosen. In [1] authors propose a What is important here is that the core of
“appropriate” means remains a matter of kernel density estimator based on the computation can be summarized in the
discussion and a recent paper [1] proposes mean integrated squared error principle following statement: find w*, such that the
a method how to exactly compute optimal (MISE) and formulate a precise algorithm formula (1) is minimal:
bandwidth under certain conditions. For how to infer optimal (fixed) kernel
large sets of spike train data the bandwidth.
elementary version of the algorithm is
unfortunately too inefficient in terms of For larger sets of recorded spike trains the
computational time complexity. time complexity of the straightforward (1),
We present a refined version of the version of the algorithm increases, so that where ti is the time of i-th spike, n is the
algorithm which in turn allows us to use the it becomes unusable for online queries number of trials,
original method even for large data sets. when studying the features of the method.
The achieved performance improvement Here we provide a solution, which ,
is demonstrated on a particular results and improves the time complexity of the
shows usability of proposed method. implementation. That at the end allows us
to work with experimental data in defines the time range of the record and kw
Keywords: action potential, spike train, a reasonable manner. It is also worth to
is as the kernel used. Since we will study
neural coding, firing rate, convolution, note that the proposed solution does not
the Gaussian kernel the equation (1) can
Gaussian kernel, kernel bandwidth, interfere with the actual result of the
be rewritten as
Brent's minimization, parallel computing, original method - for the properties and
MPI comparison with other methods look in the
original paper [1].
1. Introduction
Many neurophysiological studies are 2. Methods (2),
based on the assumption that the majority 2.1 Original method
of information flow between neurons is The firing rate is a non-negative function l where N is the number of spikes. Note that
provided by spikes. Spike trains are for which the integral ňab l(t)dt gives the the term 2Öpn2 is constant and has no
believed to form a neuronal code and many expected number of spikes during the time effect on w*. Let us denote the inner term of
coding models successfully predict interval [a,b). In the experimental recording the sum in (2) as E(w, t,i t)j . We will denote
experimental stimuli features when only we get only one or more trials of spike train W the set of possible values of w, in which
the resulting spike train is given. It has data. The problem is how to assess the
Ů we are searching. We denote the size
been shown that important aspects of the firing rate l , which will be as close as
stimuli are coded by the neuron's firing W = |W | and assume that its points are
possible to the original l, which is believed equidistant.
rate, however, the exact procedure how to to stand in the background of spike
obtain such a rate from the experimental discharges. The method is to convolve the

EJBI – Volume 6 (2010), Issue 1 © 2010 EuroMISE s.r.o.

 en74 P. Šanda: Speeding Up the Algorithm for Finding Optimal Kernel Bandwidth in Spike Train Analysis, en 73-75

The straightforward implementation will computational cost or (3) split (l2) into
find the minimum value via evaluating this many small subtasks which are
term in three nested loops: successively distributed to CPUs
according to their load. In real-life
(11) for wÎW implementation we have chosen (3)
(12) for jÎ[1,...N] because (1) tends to produce high
(13) for iÎ[1,...,j - 1] overhead of the parallelization engine and
Tmp=E(w, t,i t)j (2) assumes that the underlying CPUs are
if (min>tmp)min=tmp,w*=w equivalent in performance and accessi-
bility (that breaks in many distributed
environments).
thus obtaining the time complexity O(N2W).
The selection of W is dependent on the
3. Results
interval [a, b) and required precision of the Fig. 2. Typical shape of the function for
experimental spike train data. 3.1 Tuning parameters
optimal value. In a typical case w* << b - a
The algorithm was implemented based on
we can select the upper bound of W to
the sections above, allowing all the
log(b - a), in the case of bisection (see Splitting the task for the parallel execution strategies - exhaustive search or bisecting
below) the upper bound is not so vital. at the loop (l1) level will not allow us to use both in sequential and parallelized
parallelization in case of the bisection run, versions. The language used was C++, for
2.2 Bisecting thus we will split the task on p parts at the parallelization openMPI implementation
Now we will use that in a typical case where level (l2). That will give us the final estimate [9] of MPI standard was chosen, for
C forms an unimodal function, see Fig 2., 2
for the time complexity of O(N log(W)/p). bisection we used Brent minimization
though this cannot be asserted in general algorithm [10]. In order to find the proper
(such a problem can, for example, occur Let us stick with the implementation details splitting of the subtasks we analyzed
when searching for bandwidths is smaller now. measured time demands for a different
than the sampling resolution of input data). fragmentation of the tasks, see Fig. 3.
Having the unimodal function and a 2.3.1 Splitting
sensible estimate of lower and upper Since the upper bound in the loop (l3) is not We ran the optimality search for two sets of
bounds we can use any of the extremum constant, trivial splitting of (l2) will produce 1000 and 18000 spikes. In the case of the
search algorithms based on sectioning the p subtasks [1,...,N/p),[N/p,...,2N/p),..., larger set we could see that taking any
domain. This will reduce loop (l1) time [(p-1)N/p,...,N] with increasing time value below ¦ = 500 gives approximately
complexity from a linear to a logarithmic complexity of subtasks. At the end this the same time demands. On the very
factor and as a result we obtain the would produce a situation where the first beginning there is a visible peak caused by
complexity of O(N2log(W)). As hinted subtasks are completed having the the growth of the load by the parallelization
above while this method helps a lot certain relevant CPUs idling while the last maintenance (i.e. the cost of distributing
attention needs to be paid before its use. subtasks would still be in computation. subtasks starts to be larger than
computation of subtasks themselves).
2.3 Parallelization There are more ways how to solve it - (1) From¦ = 500 we could see gradual growth
Because the evaluation of E(w, t,i t)j is move the splitting of task into (l3), (2) caused by the insufficient fragmentation
independent of the previous computations, splitting p tasks in (l2) in a proportional (i.e. some CPUs are needlessly idle and
it is a natural target for parallelization. way, so that each subtask has the same waiting for other unfinished subtasks).

Fig.1. Illustration of the problem. The thick line Fig. 3. Figure shows how splitting affects time performance of the computation. The left panel
is the original l, the top line shows presents the case, where the input data were 1000 spikes, while the input for the right panel
experimentally measured spike train generated Ů
was 18000 spikes. Both sets were taken from the real experimental data, bisection was used
from this function, the thin line is firing rate l , in this case. f is the size of one (fixed) subtask, that is the number of (l3) iterations.
which we try to optimize.

EJBI – Volume 6 (2010), Issue 1 © 2010 EuroMISE s.r.o.

P. Šanda: Speeding Up the Algorithm for Finding Optimal Kernel Bandwidth in Spike Train Analysis, en 73-75 en75

4. Discussion [3] Sanderson A.C.: Adaptive Filtering of

Tab. 1. Comparison of time demands. Neuronal Spike Train Data. IEEE
We have proposed and implemented a
Transactions on Biomedical Engineering
parallel algorithm for optimal kernel 1980; 27 (5): 271 274.
Method Time (min:sec) bandwidth search which has better time [4] Richmond B.J., Optican L.M., Podell M.,
performance than its “straightforward” Spitzer H.: Temporal encoding of two-
Sequential search 58:41 version. Moreover when the function (1) is dimensional patterns by single units in
unimodal on the given range, we can use primate inferior temporal cortex. I.
the bisecting version, which reduces the Response characteristics. J Neurophysiol
Parallel search 4:14 1987; 57 (1): 132146.
time even more drastically. To check the [5] Richmond B.J., Optican L.M., Spitzer H.:.
reasonable ranges, one can do the first trial Temporal encoding of twodimensional
Sequential bisection 2:43 run which uses only few sampling points patterns by single units in primate primary
(and in fact cannot be omitted even in visual cortex. I. Stimulus-response
Parallel bisection 0:09 normal case). relations. J Neurophysiol 1990; 64 (2):
351369.
[6] Paulin M.G.: Digital filters for firing rate
This performance boost does not play an
estimation. Biological cybernetics 1992; 66
important role in the case of small input (6): 525531.
As the number of spikes in the input set will sets of spikes, however, it is significant in [7] Paulin M.G., Hoffman L.F.: Optimal firing
decrease, this value will also decrease, as case of large sets. The whole work was rate estimation. Neural Networks 2001; 14
the results for the set of 1000 spikes show. motivated by real demands, when (6-7): 877 881.
Here we could see similar properties as far experimental sets of ~20000 spikes were [8] Nawrot M., Aertsen A., Rotter S.: Single-trial
as the shape is concerned, but the total evaluated and, moreover, their subsets estimation of neuronal firing rates: From
time needed for computation is now single-neuron spike trains to population
also needed to be evaluated the activity. Journal of Neuroscience Methods
negligible. approximate knowledge of the function (1) 1999; 94 (1): 81 92.
shape reduced the need for a slow version [9] Gabriel E., Fagg G.E., Bosilca G., Angskun
This leads to the final choice of ¦ = 100, of the algorithm. T., Dongarra J.J., Squyres J.M. et al:. Open
which will be always sufficient for any MPI: Goals, Concept, and Design of a Next
larger input sets. As it can be seen in the At the end we proposed tuning parameters Generation MPI Implementation. In:
Proceedings, 11th European PVM/MPI
left panel of Fig. 3, it is a reasonable value for an example cluster configuration and Users' Group Meeting 2004; 97104.
even for small sets, but that is not so provided actual results of the performance [10] Brent R.P.: Algorithms for minimization
important due to small total time demands. improvement. without derivatives. Dover Pubns; 2002.
This value is, of course, dependent on the
particular computational setting - in our Acknowledgments
case all tests have been done on a small The work was supported by the grant SVV- Contact
cluster with 20 CPU cores. 2010-265 513. Mgr. Pavel Šanda
Institute of Physiology,
3.2 Real time demands References Academy of Sciences of the Czech
For the comparison of real-time improve- [1] Shimazaki H., Shinomoto S:. Kernel Republic
ments we offer the table below. The input bandwidth optimization in spike rate Vídeňská 1083
data and parameters were the same for all estimation. Journal of Computational 142 20 Prague 4
Neuroscience 2010; 29: 171182. Czech Republic
the tasks: 18000 spikes, [1;400] ms range
[2] Cunningham J.P., Gilja V., Ryu S.I.,
for bandwidth, precision of 1 ms Shenoy, K.V.: Methods for estimating neural
e-mail: [email protected]
(W=400/1). firing rates, and their application to brain-
machine interfaces. Neural Networks 2009;
22 (9): 12351246.

EJBI – Volume 6 (2010), Issue 1 © 2010 EuroMISE s.r.o.

Original Article en51

Jitter Effect on the Performance of the Sound Localization

Model of Medial Superior Olive Neural Circuit
Pavel Šanda1,2
1
Institute of Physiology, Academy of Sciences of the Czech Republic
2
3rd Medical Department, First Faculty of Medicine, Charles University in Prague, Czech Republic

Abstract
Objectives: Additional properties of the stochastic neural
circuit model suggested in [1] were studied.
Methods: The performance of the whole circuit when the
system employs a different jitter was studied by extensive
simulations. By performance we mean the time needed to
obtain a reliable estimate of ITD.
Results: It was found that the relation between jitter
and performance is nonlinear and we estimated a plausible Mgr. Pavel Šanda
range of jitter values for the model.
Conclusion: To conclude, there exists an upper bound
of the timing jitter since the number of neurons needed
to compensate the injected noise grows exponentially and Keywords
above certain jitter values becomes unrealistically high.
Medial superior olive (MSO), stochastic model, timing jit-
ter, interaural time difference (ITD)
Correspondence to:
EJBI 2011; 7(1):51–54
Mgr. Pavel Šanda recieved: September 10, 2011
Institute of Physiology, Academy of Sciences of the Czech Republic accepted: October 31, 2011
Address: Videnska 1082, 142 20 Prague 4, Czech Republic published: November 20, 2011
E–mail: [email protected]

1 Introduction rate is driven by coincidence detection of the action po-

tentials coming from time locked ipsi- and contralateral
The way mammalian brain localizes sound azimuth re- inputs shifted by ITD and additional jitter added to the
mains a matter of discussion. The current textbook view system. Under certain conditions each ITD value corre-
is based on the theory of delay lines proposed a long time sponds to a unique value of the firing rate, thus the imag-
ago by [2]. inary observer monitoring output of such a neuron is able
to estimate ITD only by interpolation from its firing rate.
Although there is a strong experimental evidence that
delay lines implemented by the branching pattern of neu- The role of noise in this model is ambiguous. On the
ronal fibers are present in the Nucleus Laminaris in birds one hand it allows a finer distribution of recognized ITD
[3], experimental evidence for such branching pattern in values, on the other hand higher values deteriorate the
the Medial Superior Olive (MSO - counterpart of bird’s estimation performance of the circuit.
NL) in mammals remain weak [4] and alternative theories
have been proposed [5]. This performance decline was indicated in [1] for two
circuits with different jitter. The aim of this report is to
In a specific variant of the slope-encoding model [6] extend the previous result and show quantitatively how
proposed in [1] the interaural time difference (ITD) is en- jitter affects performance of the whole range of circuits
coded by the firing rate of the first binaural neuron. This defined by different jitter values.

c 2011 EuroMISE s.r.o. EJBI – Volume 7 (2011), Issue 1

en52 Šanda – Jitter Effect on the Performance of the Sound Localization Model

2 Methods information can, in principle, be obtained from the

rate coded presented by a single binaural neuron.
The circuit operates at an abstract level of description
without explicit membrane potential regarding spikes as Details of the stages above are identical to those in [1]
single time point events and consists of several consecutive except for one important feature. Fixed parameters of the
processing stages (see Fig. 1): circuit define the ITD interpolation curve as seen in Fig.
2. In our previous study this curve was carefully fitted to
a fixed sinusoidal function and the inverse of this function
was used to interpolate ITD from estimated firing rate.
In Fig. 2 we can see how jitter J dramatically changes
this curve. Since we will use the whole range of different
jitter values we cannot rely on the fitted function any-
more and we shall use directly this interpolation curve.
Conceptually, this is not adding anything new, however,
it leads to additional computational difficulties - for each
jitter value a circuit ITD curve must be recomputed anew
and an inverse mapping from firing rate to ITD must use
a more elaborate interpolation mechanism since the curve
is not locally strictly monotonous.

Figure 1: Scheme of successive processing stages of the circuit.

• A generator of action potentials simulates time-

locked inputs, impulses from the contralateral side
are shifted in time by ITD value. The frequency of
generation is set to 140 Hz as in the previous study.
• A jitter generator which represents noise occurring
in the circuit during the signal transmission along
the auditory pathway. It is parametrized by a sin-
gle value. It should be noted that each different
parameter value defines a different circuit since it Figure 2: ITD interpolation curves for circuits with a differ-
changes the characteristic ITD interpolation curve ent jitter value. Each firing rate value corresponds to an ITD
used for interpolation. Together with the spike value and is uniquely determined in case the function is strictly
generator they can be considered as a very sim- increasing in the ITD values under scrutiny. We see that in-
creasing jitter leads to smaller slopes of the interpolation curve
plified counterpart to the auditory pathway up to
and we expect a deteriorated circuit performance for higher jit-
the MSO (where the signal from the left and right
ter values.
ear used for sound localization based on low sound
frequencies converges). In this stage each spike is
shifted in time by small random jitter ∆ which is
parametrized by jitter magnitude J, more precisely 3 Results
∆ = J(B(2, 4) − 0.5), where B(a, b) is a random
variable from the beta distribution with parameters Each jitter value defines a new circuit and after com-
a, b. puting its interpolation curve we let the circuit estimate
a single ITD value while observing how the estimate de-
• A coincidence detector representing the first binau-
velops in time. This way we obtain asymptotic behaviour
ral neuron. It generates a new spike only in case two
for each circuit, see Fig. 3.
input spikes occur within a short time window and
in a specific order when contralateral spike precedes From psychophysical experiments we know that the
the ipsilateral one. precision of azimuth estimation in a human is appro-
ximately 4◦ in the head-on direction [7]. We define that
• An observer which collects output of the previous the time needed for reliable estimation of ITD is identical
processing stages and estimates the ITD value com- with the last-passage-time (LPT) of the 4◦ precision re-
puted by the circuit. It can be seen as a counterpart gion, see the area delineated by horizontal dotted lines in
of higher processing stages which measure how much Fig. 3.

EJBI – Volume 7 (2011), Issue 1 c 2011 EuroMISE s.r.o.

Šanda – Jitter Effect on the Performance of the Sound Localization Model en53

Figure 3: Asymptotic behaviour of ITD estimation produced by observer for selected values of jitter J. The original azimuth
was selected as IT D = 0. Horizontal lines delineate the region when desired precision of ITD estimate was achieved (±2◦ ).
For each line we can define the last passage time (LPT) when the function enters the region and remains inside of it. We see
that increasing jitter leads to the increase of LPT value. Each line is an averaged function from 1000 simulation runs.

Figure 4: Dependence of last-passage-times on different jitter values. On the right hand side the same plot in logarithmic scale.
We fit ff it (x) in such a way to be as close as possible in the interval of 0.1 - 10 s of LPT. This will be subsequently used for
relating plausible jitter ranges, see text.

In this way we obtain unique LPT for each circuit with This allows us at least to connect specific jitter value
specific jitter value, as plotted in Fig. 4. As we can see, J with the required number of neurons n in order to ob-
the functional dependence is nonlinear and can be appro- tain tA (let us fix tA = 0.2 s). By employing the er-
ximately fitted by ff it (x) = e1.9(x−1.25) − 0.2. godic hypothesis we get tA = LP Tn (J) and from fitting
Obtained circuit’s LPT time tA corresponds to the pro- ff it (J)LP T (J), hence
cessing time of a single binaural neuron in MSO needed to
estimate ITD. Because the auditory pathway consists of ff it (J) e1.9(J−1.25) − 0.2
n= = . (1)
many parallel fibers and processing of the signal is simul- t A 0.2
taneous, we used the ergodic hypothesis in our previous
study. To sum up, we obtain that the physiologically plausible
range of simultaneously working neurons n ∈ [1; 100] cor-
In short, we assume that when a single neuron of this responds to jitter range J ∈ [0.7; 2.8], which also implies
type requires the time tA , n neurons working in parallel plausible jitter values for the canonical set of parameters
need the time tA /n to produce equivalent information sub- of this model.
sequently used in higher stages of the pathway (repre-
sented by the concept of the observer).
The number of binaural neurons working in parallel 4 Discussion
is difficult to estimate but does not exceed hundreds of
units. Next, we know from psychophysical experiments Irregularities in spike timings observed in physiologi-
that the time tA needed for azimuth estimation ranges cal recordings were originally thought to be the result of
around 150 - 300 ms in human subjects [8]. neuronal cells unreliability and it was assumed that the

c 2011 EuroMISE s.r.o. EJBI – Volume 7 (2011), Issue 1

en54 Šanda – Jitter Effect on the Performance of the Sound Localization Model

firing-rate neural coding scheme is used because of its ro- Acknowledgments

bustness against the noise present in neuronal activity.
Later decades have shown that what was often considered The work was supported by the grant SVV-2011-262
as erratic behaviour was rather a misunderstanding of the 514 of Charles University in Prague.
transmitted code [9] and it turned out that neurons are
capable of reliable and precise spike timing [10] needed
for so-called temporal coding. Coincidence detection of
References
precisely timed input spikes is an important concept in [1] Sanda P, Marsalek P. Stochastic InterpolationModel of the Me-
theories of binaural hearing and we suggested one variant dial Superior Olive Neural Circuit. Brain Res. 2011;in press.
of such a model in a stochastic neural circuit in [1].
[2] Jeffress LA. A place theory of sound localization. J Comp
Physiol Psychol. 1948;41(1):3539.
This time we focused specifically on the role of jitter.
In the previous study the jitter parameter was fixed to [3] Carr CE, Konishi M. Axonal delay lines for time mea-
surement in the owls brainstem. Proc Natl Acad Sci USA.
J = 1 ms which is in a good agreement with experimental 1988;85(21):83118315.
findings [11]. Here we took a further step and estimated a
range of possible values based on circuit performance. We [4] Grothe B. New roles for synaptic inhibition in sound localiza-
tion. Nat Rev Neurosci. 2003;4(7):540-50.
should, however, note that this analysis is bound to the
canonical set of basic circuit parameters. For example, [5] Jennings TR, Colburn HS. Models of the Superior Olivary
the spike generator frequency also has an impact on the Complex. In: Meddis R, Lopez-Poveda EA, Fay RR, Popper
AN, editors. Computational Models of the Auditory System.
overall performance of the circuit; in the previous study Springer, New York; 2010. p. 6596.
we employed a more detailed model of the auditory peri-
phery [12] and we could observe a decrease of overall per- [6] McAlpine D, Jiang D, Palmer AR. A neural code for low-
frequency sound localization in mammals. Nat Neurosci.
formance of the circuit. This result cannot be, however, so 2001;4(4):396401.
easily incorporated since one processing stage (bushy cells
layer) is missing. There are indications that this layer is [7] Mills AW. Auditory Localization. In: Tobias JV, editor. Foun-
dations of Modern Auditory Theory. New York: Academic
able to provide better time locking and consequently im-
Press; 1972. p. 303348.
prove coincidence detection in binaural neurons — that
can be another example of a somewhat unexpected ob- [8] Middlebrooks JC, Green DM. Sound Localization by Human
servation that higher processing stages of neural circuity Listeners. Annu Rev of Psychol. 1991;42(1):135159.
increase the accuracy of phase locking [13]. [9] Barlow HB. Single units and sensation: a neuron doctrine for
perceptual psychology. Perception. 1972;1(4):371394.
Another problematic point is that the number of pa- [10] Mainen ZF, Sejnowski TJ. Reliability of spike timing in neo-
rallel circuits employed in ITD estimation is not exper- cortical neurons. Science. 1995;268(5216):15031506.
imentally known. This parallelism would have a strong
[11] Oertel D, Bal R, Gardner SM, Smith PH, Joris PX. Detection
impact on the overall performance as well, and we have of synchrony in the activity of auditory nerve fibers by octopus
at least shown the correspondence between jitter and the cells of the mammalian cochlear nucleus. Proc Natl Acad Sci
required number of neurons (or vice versa). By employ- USA. 2000;97(22):1177311779.
ing the ergodic hypothesis we can conclude that due to [12] Meddis R. Auditory-nerve first-spike latency and auditory
(1) the number of neurons needed to compensate the in- absolute threshold: A computer model. J Acoust Soc Am.
jected noise grows exponentially and above certain jitter 2006;119(1):406417.
values becomes unrealistically high. This gives us an ap-
[13] Carr CE, Heiligenberg W, Rose GJ. A time-comparison cir-
proximate upper bound of jitter allowed for this type of cuit in the electric fish midbrain. I. Behavior and physiology.
circuit. J Neurosci. 1986;6(1):107.

EJBI – Volume 7 (2011), Issue 1 c 2011 EuroMISE s.r.o.