[translation] pointprocesses done?

This commit is contained in:
Jan Grewe 2018-11-04 17:01:45 +01:00
parent e32d7f501c
commit 8db863514b

View File

@ -107,7 +107,7 @@ are described using the common measures.
\rangle)^2 \rangle}$\vspace{1ex}
\item \enterm{Coefficient of variation}: $CV_{ISI} =
\frac{\sigma_{ISI}}{\mu_{ISI}}$.
\item \enterm{Diffusion coefficient}): $D_{ISI} =
\item \enterm{Diffusion coefficient}: $D_{ISI} =
\frac{\sigma_{ISI}^2}{2\mu_{ISI}^3}$.
\end{itemize}
@ -294,9 +294,9 @@ The homogeneous Poisson process has the following properties:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Time-dependent firing rate}
So far we discussed stationary spiketrains. The statistical properties
So far we have discussed stationary spiketrains. The statistical properties
of these did not change within the observation time (stationary point
processes. Most commonly, however, this is not the case. A sensory
processes). Most commonly, however, this is not the case. A sensory
neuron, for example, might respond to a stimulus by modulating its
firing rate (non-stationary point process).
@ -304,215 +304,229 @@ How the firing rate $r(t)$ changes over time is the most important
measure, when analyzing non-stationary spike trains. The unit of the
firing rate is Hertz, i.e. the number of action potentials per
second. There are different ways to estimate the firing rate and three
of these methods will are illustrated in \figref{psthfig}. All of
of these methods are illustrated in \figref{psthfig}. All of
these have their own justifications and pros- and cons. In the
following we will discuss the methods shown in \figref{psthfig} more
following we will discuss these methods more
closely.
\begin{figure}[tp]
\includegraphics[width=\columnwidth]{firingrates}
\titlecaption{Bestimmung der zeitabh\"angigen
Feuerrate.}{\textbf{A)} Rasterplot eines Spiketrains. \textbf{B)}
Feurerrate aus der instantanen Feuerrate bestimmt. \textbf{C)}
klassisches PSTH mit der Binning Methode. \textbf{D)} Feuerrate
durch Faltung mit einem Gauss Kern bestimmt.}\label{psthfig}
\titlecaption{Estimating the time-dependent firing
rate.}{\textbf{A)} Rasterplot depicting the spiking activity of a
neuron. \textbf{B)} Firing rate calculated from the
\emph{instantaneous rate}. \textbf{C)} classical PSTH with the
\emph{binning} method. \textbf{D)} Firing rate estimated by
\emph{convolution} of the activity with a Gaussian
kernel.}\label{psthfig}
\end{figure}
\subsection{Instantane Feuerrate}
\subsection{Instantaneous firing rate}
\begin{figure}[tp]
\includegraphics[width=\columnwidth]{isimethod}
\titlecaption{Instantane Feuerrate.}{Skizze eines Spiketrains
(oben). Die Pfeile zwischen aufeinanderfolgenden
Aktionspotentialen mit den Zahlen in Millisekunden illustrieren
die Interspikeintervalle. Der Kehrwert des Interspikeintervalle
ergibt die instantane Feuerrate.}\label{instrate}
\titlecaption{Instantaneous firing rate.}{Sketch of the recorded
spiketrain (top). Arrows illustrate the interspike intervals and
number give the intervals in milliseconds. The inverse of the
interspike interval is the \emph{instantaneous firing
rate}.}\label{instratefig}
\end{figure}
Ein sehr einfacher Weg, die zeitabh\"angige Feuerrate zu bestimmen ist
die sogenannte \determ[Feuerrate!instantane]{instantane Feuerrate}
(\enterm[firing rate!instantaneous]{instantaneous firing rate}). Dabei
wird die Feuerrate aus dem Kehrwert der Interspikeintervalle, der Zeit
zwischen zwei aufeinander folgenden Aktionspotentialen
(\figref{instrate} A), bestimmt. Die abgesch\"atzte Feuerrate
(\figref{instrate} B) ist g\"ultig f\"ur das gesammte
Interspikeintervall. Diese Methode hat den Vorteil, dass sie sehr
einfach zu berechnen ist und keine Annahme \"uber eine relevante
Zeitskala (der Kodierung oder des Auslesemechanismus der
postsynaptischen Zelle) macht. $r(t)$ ist allerdings keine
kontinuierliche Funktion, die Spr\"unge in der Feuerrate k\"onnen
f\"ur manche Analysen nachteilig sein. Au{\ss}erdem wird die Feuerrate
nie gleich Null, auch wenn lange keine Aktionspotentiale generiert
wurden.
A very simple method for estimating the time-dependent firing rate is
the \enterm[firing rate!instantaneous]{instantaneous firing rate}. The
firing rate can be directly estimated as the inverse of the time
between successive spikes, the interspike-interval
(\figref{instratefig}).
\begin{equation}
\label{instantaneousrateeqn}
r_i = \frac{1}{T_i} .
\end{equation}
The instantaneous rate $r_i$ is valid for the whole interspike
interval. The method has the advantage of being extremely easy to
compute and that it does not make any assumptions about the relevant
timescale (of the encoding in the neuron or the decoding of a
postsynaptic neuron). The resulting $r(t)$, however, is no continuous
function, the firing rate jumps from one level to the next. Since the
interspike interval between successive spikes is never infinitely
long, the firing rate never reaches zero despite that the neuron may
not fire an action potential for a long time.
\begin{exercise}{instantaneousRate.m}{}
Implementiere die Absch\"atzung der Feuerrate auf Basis der
instantanen Feuerrate. Plotte die Feuerrate als Funktion der Zeit.
Implement a function that computes the instantaneous firing
rate. Plot the firing rate as a function of time.
\note{TODO: example data!!!}
\end{exercise}
\subsection{Peri-Stimulus-Zeit-Histogramm}
W\"ahrend die Instantane Rate den Kehrwert der Zeit von einem bis zum
n\"achsten Aktionspotential misst, sch\"atzt das sogenannte
\determ{Peri-Stimulus-Zeit-Histogramm} (\enterm{peri stimulus time
histogram}, \determ[PSTH|see{Peri-Stimulus-Zeit-Histogramm}]{PSTH})
die Wahrscheinlichkeit ab, zu einem Zeitpunkt Aktionspotentiale
anzutreffen. Es wird versucht die mittlere Rate \eqnref{firingrate} im
Grenzwert kleiner Beobachtungszeiten abzusch\"atzen:
\subsection{Peri-stimulus-time-histogram}
While the \emph{instantaneous firing rate} uses the interspike
interval, the \enterm{peri stimulus time histogram} (PSTH) uses the
spike count within observation windows of the duration $W$. It
estimates the probability of observing a spike within that observation
time. It tries to estimat the average rate in the limit of small
obersvation times \eqnref{psthrate}:
\begin{equation}
\label{psthrate}
r(t) = \lim_{W \to 0} \frac{\langle n \rangle}{W} \; ,
\end{equation}
wobei die Anzahl $n$ der Aktionspotentiale, die im Zeitintervall
$(t,t+W)$ aufgetreten sind, \"uber trials gemittelt wird. Eine solche
Rate enspricht der zeitabh\"angigen Rate $\lambda(t)$ des inhomogenen
Poisson-Prozesses.
where $\langle n \rangle$ is the across trial average number of action
potentials observed within the interval $(t, t+W)$. Such description
matches the time-dependent firing rate $\lambda(t)$ of an
inhomogeneous Poisson process.
Das PSTH \eqnref{psthrate} kann entweder \"uber die Binning-Methode
oder durch Verfaltung mit einem Kern bestimmt werden. Beiden Methoden
gemeinsam ist die Notwendigkeit der Wahl einer zus\"atzlichen Zeitskala,
die der Beobachtungszeit $W$ in \eqnref{psthrate} entspricht.
The firing probability can be estimated using the \emph{binning
method} or by using \emph{kernel density estimations}. Both methods
make an assumption about the relevant observation time-scale ($W$ in
\eqnref{psthrate}).
\subsubsection{Binning-Methode}
\subsubsection{Binning-method}
\begin{figure}[tp]
\includegraphics[width=\columnwidth]{binmethod}
\titlecaption{Bestimmung des PSTH mit der Binning Methode.}{Der
gleiche Spiketrain wie in \figref{instrate}. Die grauen Linien
markieren die Grenzen der Bins und die Zahlen geben die Anzahl der Spikes
in jedem Bin an (oben). Die Feuerrate ergibt sich aus dem
mit der Binbreite normierten Zeithistogramm (unten).}\label{binpsth}
\titlecaption{Estimating the PSTH using the binning method.}{ The
same spiketrain as shown in \figref{instratefig} is used
here. Vertical gray lines indicate the borders between adjacent
bins in which the number of action potentials is counted (red
numbers). The firing rate is then the histogram normalized to the
binwidth.}\label{binpsthfig}
\end{figure}
Bei der Binning-Methode wird die Zeitachse in gleichm\"aßige
Abschnitte (Bins) eingeteilt und die Anzahl Aktionspotentiale, die in
die jeweiligen Bins fallen, gez\"ahlt (\figref{binpsth} A). Um diese
Z\"ahlungen in die Feuerrate umzurechnen muss noch mit der Binweite
normiert werden. Das ist \"aquivalent zur Absch\"atzung einer
Wahrscheinlichkeitsdichte. Es kann auch die \code{hist()} Funktion zur
Bestimmung des PSTHs verwendet werden. \sindex[term]{Feuerrate!Binningmethode}
Die bestimmte Feuerrate gilt f\"ur das gesamte Bin (\figref{binpsth}
B). Das so berechnete PSTH hat wiederum eine stufige Form, die von der
Wahl der Binweite anh\"angt. $r(t)$ ist also keine stetige
Funktion. Die Binweite bestimmt die zeitliche Aufl\"osung der
Absch\"atzung. \"Anderungen in der Feuerrate, die innerhalb eines Bins
vorkommen k\"onnen nicht aufgl\"ost werden. Mit der Wahl der Binweite
wird somit eine Annahme \"uber die relevante Zeitskala des Spiketrains
gemacht.
The binning method separates the time axis into regular bins of the
bin width $W$ and counts for each bin the number of observed action
potentials (\figref{binpsthfig} top). The resulting histogram is then
normalized with the bin width $W$ to yield the firing rate shown in
the bottom trace of figure \ref{binpsthfig}. The above sketched
process is equivalent to estimating the probability density. It is
possible to estimate the PSTH using the \code{hist()} method
\sindex[term]{Feuerrate!Binningmethode}
The estimated firing rate is valid for the total duration of each
bin. This leads to the step-like plot shown in
\figref{binpsthfig}. $r(t)$ is thus not a contiunous function in
time. The binwidth defines the temporal resolution of the firing rate
estimation Changes that happen within a bin cannot be resolved. Thus
chosing a bin width implies an assumption about the relevant
time-scale.
\pagebreak[4]
\begin{exercise}{binnedRate.m}{}
Implementiere die Absch\"atzung der Feuerrate mit der ``binning''
Methode. Plotte das PSTH.
Implement a function that estimates the firing rate using the
``binning method''. The method should take the spike-times as an
input argument and returns the firing rate. Plot the PSTH.
\end{exercise}
\subsubsection{Faltungsmethode}
\subsubsection{Convolution method --- Kernel density estimation}
\begin{figure}[tp]
\includegraphics[width=\columnwidth]{convmethod}
\titlecaption{Bestimmung des PSTH mit der Faltungsmethode.}{Der
gleiche Spiketrain wie in \figref{instrate}. Bei der Verfaltung
des Spiketrains mit einem Faltungskern wird jeder Spike durch den
Faltungskern ersetzt (oben). Bei korrekter Normierung des
Kerns ergibt sich die Feuerrate direkt aus der \"Uberlagerung der
Kerne.}\label{convrate}
\titlecaption{Estimating the firing rate using the convolution
method.}{The same spiketrain as in \figref{instratefig}. The
convolution of the spiketrain with a convolution kernel basically
replaces each spike event with the kernel (top). A Gaussian kernel
is used here, but other kernels are also possible. If the kernel
is properly normalized the firing rate results directly form the
superposition of the kernels.}\label{convratefig}
\end{figure}
Bei der Faltungsmethode werden die harten Kanten der Bins der
Binning-Methode vermieden. Der Spiketrain wird mit einem Kern
verfaltet, d.h. jedes Aktionspotential wird durch den Kern ersetzt.
Zur Berechnung wird die Aktionspotentialfolge zun\"achst
``bin\"ar'' dargestellt. Dabei wird ein Spiketrain als
(Zeit-)Vektor dargestellt, in welchem die Zeitpunkte der
Aktionspotentiale als 1 notiert werden. Alle anderen Elemente des
Vektors sind 0. Anschlie{\ss}end wir dieser bin\"are Spiketrain mit
einem Gau{\ss}-Kern bestimmter Breite verfaltet:
\[r(t) = \int_{-\infty}^{\infty} \omega(\tau) \, \rho(t-\tau) \, {\rm d}\tau \; , \]
wobei $\omega(\tau)$ der Filterkern und $\rho(t)$ die bin\"are Antwort
ist. Bildlich geprochen wird jede 1 in $\rho(t)$ durch den Filterkern
ersetzt (Abbildung \ref{convrate} A). Wenn der Kern richtig normiert
wurde (Integral gleich Eins), ergibt sich die Feuerrate direkt aus der
\"Uberlagerung der Kerne (Abb. \ref{convrate} B). \sindex[term]{Feuerrate!Faltungsmethode}
Die Faltungsmethode f\"uhrt, anders als die anderen Methoden, zu einer
stetigen Funktion was insbesondere f\"ur spektrale Analysen von
Vorteil sein kann. Die Wahl der Kernbreite bestimmt, \"ahnlich zur
Binweite, die zeitliche Aufl\"osung von $r(t)$. Die Breite des Kerns
macht also auch wieder eine Annahme \"uber die relevante Zeitskala des
Spiketrains.
With the convolution method we avoid the sharp edges of the binning
method. The spiketrain is convolved with a \enterm{convolution
kernel}. Technically speaking we need to first create a binary
representation of the spike train. This binary representation is a
series of zeros and ones in which the ones denote the spike. Then this binary vector is convolved with a kernel of a certain width:
\[r(t) = \int_{-\infty}^{\infty} \omega(\tau) \, \rho(t-\tau) \, {\rm d}\tau \; , \]
where $\omega(\tau)$ represents the kernel and $\rho(t)$ the binary
representation of the response. The process of convolution can be
imagined as replacing each event of the spiketrain with the kernel
(figure \ref{convratefig} top). The superposition of the replaced
kernels is then the firing rate (if the kerel is correctly normalized
to an integral of one, figure \ref{convreffig}
bottom). \sindex[term]{Feuerrate!Faltungsmethode}
In contrast to the other methods the convolution methods leads to a
continuous function which is often desirable (in particular when
applying methods in the frequency domain). The choice of the kernel
width defines, similar to the bin width of the binning method, the
temporal resolution of the method and thus makes assumptions about the
relevate time-scale.
\pagebreak[4]
\begin{exercise}{convolutionRate.m}{}
Verwende die Faltungsmethode um die Feuerrate zu bestimmen. Plotte
das Ergebnis.
Implement the function that estimates the firing rate using the
convolution method. The method takes the spiketrain, the temporal
resolution of the recording (as the stepsize $dt$, in seconds) and
the width of the kernel (the standard deviation $\sigma$ of the
Gaussian kernel, in seconds) as input arguments. It returns the
firing rate. Plot the result.
\end{exercise}
\section{Spike-triggered Average}
Die graphischer Darstellung der Feuerrate allein reicht nicht aus um
den Zusammenhang zwischen neuronaler Antwort und einem Stimulus zu
analysieren. Eine Methode um mehr \"uber diesen Zusammenhang zu
erfahren, ist der \enterm{spike-triggered average}
(\enterm[STA|see{spike-triggered average}]{STA}). Der STA
The graphical representation of the neuronal activity alone is not
sufficient tot investigate the relation between the neuronal response
and a stimulus. One method to do this is the \enterm[STA|see
spike-triggered average]{spike-triggered average}. The STA
\begin{equation}
STA(\tau) = \langle s(t - \tau) \rangle = \frac{1}{N} \sum_{i=1}^{N} s(t_i - \tau)
\end{equation}
der $N$ Aktionspotentiale zu den Zeiten $t_i$ in Anwort auf den
Stimulus $s(t)$ ist der mittlere Stimulus, der zu einem
Aktionspotential in der neuronalen Antwort f\"uhrt.
Der STA l\"a{\ss}t sich relativ einfach berechnen, indem aus dem
Stimulus f\"ur jeden beobachteten Spike ein entsprechender Abschnitt
ausgeschnitten wird und diese dann gemittelt werde (\figref{stafig}).
of $N$ action potentials observed at the times $t_i$ in response to
the stimulus $s(t)$ is the average stimulus that led to a spike in the
neuron. The STA can be easily extracted by cutting snippets out of
$s(t)$ that surround the times of the spikes. The resulting stimulus
snippets are then averaged (\figref{stafig}).
\begin{figure}[t]
\includegraphics[width=\columnwidth]{sta}
\titlecaption{Spike-triggered Average eines P-Typ Elektrorezeptors
und Stimulusrekonstruktion.}{Der STA (links): der Rezeptor
wurde mit einem ``white-noise'' Stimulus getrieben. Zeitpunkt 0
ist der Zeitpunkt des beobachteten Aktionspotentials. Die Kurve
ergibt sich aus dem gemittelten Stimulus je 50\,ms vor und nach
einem Aktionspotential. Stimulusrekonstruktion mittels
STA (rechts). Die Zellantwort wird mit dem STA gefaltet um eine
Rekonstruktion des Stimulus zu erhalten.}\label{stafig}
\titlecaption{Spike-triggered average of a P-type electroreceptors
and the stimulus reconstruction.}{The neuron was driven by a
``white-noise'' stimulus (blue curve, right panel). The STA (left)
is the average stimulus that surrounds the times of the recorded
action potentials (40\,ms before and 20\,ms after the
spike). Using the STA as a convolution kernel for convolving the
spiketrain we can reconstruct the stimulus from the neuronal
response. In this way we can get an impression of the stimulus
features that are encoded in the neuronal response (right
panel).}\label{stafig}
\end{figure}
Aus dem STA k\"onnen verschiedene Informationen \"uber den
Zusammenhang zwischen Stimulus und neuronaler Antwort gewonnen
werden. Die Breite des STA repr\"asentiert die zeitliche Pr\"azision,
mit der Stimulus und Antwort zusammenh\"angen und wie weit das Neuron
zeitlich integriert. Die Amplitude des STA (gegeben in der gleichen
Einheit wie der Stimulus) deutet auf die Empfindlichkeit des Neurons
bez\"uglich des Stimulus hin. Eine hohe Amplitude besagt, dass es
einen starken Stimulus ben\"otigt, um ein Aktionspotential
hervorzurufen. Aus dem zeitlichen Versatz des STA kann die Zeit
abgelesen werden, die das System braucht, um auf den Stimulus zu
antworten.
Der STA kann auch dazu benutzt werden, aus den Antworten der Zelle den
Stimulus zu rekonstruieren (\figref{stafig} B). Bei der
\determ[invertierte Rekonstruktion]{invertierten Rekonstruktion} wird
die Zellantwort mit dem STA verfaltet.
From the STA we can extract several pieces of information about the
relation of stimulus and response. The width of the STA represents the
temporal precision with which the neuron encodes the stimulus waveform
and in which temporal window the neuron integrates the (sensory)
input. The amplitude of the STA tells something about the sensitivity
of the neuron. The STA is given in the same units as the stimulus and
a small amplitude indicates that the neuron needs only a small
stimulus amplitude to create a spike, a large amplitude on the
contrary suggests the opposite. The temporal delay between the STA and
the time of the spike is a consequence of the time the system (neuron)
needs to process the stimulus.
We can further use the STA to do a \enterm{reverse reconstruction} and
estimate the stimulus from the neuronal response (\figref{stafig},
right). For this, the spiketrain is convolved with the STA as a
kernel.
\begin{exercise}{spikeTriggeredAverage.m}{}
Implementiere eine Funktion, die den STA ermittelt. Verwende dazu
den Datensatz \file{sta\_data.mat}. Die Funktion sollte folgende
R\"uckgabewerte haben:
Implement a function that calculates the STA. Use the dataset
\file{sta\_data.mat}. The function expects the spike train, the
stimulus and the temporal resolution of the recording as input
arguments and should return the following information:
\vspace{-1ex}
\begin{itemize}
\setlength{\itemsep}{0ex}
\item den Spike-Triggered-Average.
\item die Standardabweichung der individuellen STAs.
\item die Anzahl Aktionspotentiale, die zur Berechnung des STA verwendet wurden.
\vspace{-2ex}
\item the spike-triggered average.
\item the standard deviation of the STA across the individual snippets.
\item The number of action potentials used to estimate the STA.
\end{itemize}
\end{exercise}
\begin{exercise}{reconstructStimulus.m}{}
Rekonstruiere den Stimulus mithilfe des STA und der Spike
Zeiten. Die Funktion soll Vektor als R\"uckgabewert haben, der
genauso gro{\ss} ist wie der Originalstimulus aus der Datei
Do the reverse reconstruction using the STA and the spike times. The
function should return the estimated stimulus in a vector that has
the same size as the original stimulus contained in file
\file{sta\_data.mat}.
\end{exercise}