Testing the Odds of Inherent versus Observed Over-dispersion in Neural Spike Counts Odds of Inherent versus Observed Over-dispersion
Extracted source text ยท arXiv 1611.04364
% !TEX TS-program = PdfLaTeX
% !TEX spellcheck = en-US
% !BIB TS-program = bibtex
% !TEX encoding = UTF-8 Unicode
% -*- coding: UTF-8; -*-
% vim: set fenc=utf-8
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\def\Draft{1}% draft=1 or no draft = 0
\def\Draft{0}% draft=1 or no draft = 0
\newcommand{\AuthorLuP}{Laurent U.~Perrinet}%
\newcommand{\AuthorWT}{Wahiba Taouali} %
\newcommand{\AuthorGB}{Giacomo Benvenuti}%
\newcommand{\AuthorPW}{Pascal Wallisch} %
\newcommand{\AuthorFC}{ Fr\'ed\'eric Chavane} %
\newcommand{\AddressA}{Institut de Neurosciences de la Timone, CNRS, Aix-Marseille Universit\'e, Marseille, France}% - Marseille, France
\newcommand{\AddressB}{Center for Neural Science, New York University, New York, USA}%
%\newcommand{\AddressB}{CNRS, Aix-Marseille University}
\newcommand{\LongAddressA}{
Institut de Neurosciences de la Timone (UMR 7289),
CNRS, Aix-Marseille Universit\'e \\
%Facult\'e de M\'edecine - B\^atiment Neurosciences,
27, Bd Jean Moulin,
13385 Marseille Cedex 05,
France}
\newcommand{\PhoneWT}{+33.662 866 453}%
\newcommand{\Website}{http://invibe.net/LaurentPerrinet}%
\newcommand{\EmailWT}{wahiba.taouali@univ-amu.fr}%
\newcommand{\Title}{
Testing the Odds of Inherent versus Observed Over-dispersion in Neural Spike Counts
%A doubly stochastic (model) decoder (application) accounting for over-dispersion (test) in spike counts
}%
\newcommand{\RunningTitle}{Odds of Inherent versus Observed Over-dispersion}
\newcommand{\Abstract}{
The repeated presentation of an identical visual stimulus in the receptive field of a neuron may evoke different spiking patterns at each trial. Probabilistic methods are essential to understand \xchanged{its}{the} functional role \xadded{of this variance} within the neural activity. In that case, a Poisson process is the most common model of trial-to-trial variability. \xchanged{However}{For a Poisson process}, the variance of the spike count is constrained to be equal to the mean, irrespective of \xchanged{measurement's duratio}{the duration of measurements}. Numerous studies have shown that this relationship does not generally hold. Specifically, a majority of electrophysiological recordings show an ``{\em \xchanged{overdispersion}{over-dispersion}}'' effect: Responses that exhibit more inter-trial variability than expected from a Poisson process alone. A model that is particularly well suited to quantify \xchanged{overdispersion}{over-dispersion} is the Negative-Binomial distribution model. This model is \xchanged{largely applied and studied}{well-studied and widely used} but has only recently been applied to neuroscience. In this paper, we address three main issues. First, we describe how the Negative-Binomial distribution provides a model apt to account for overdispersed spike counts. Second, we quantify the significance of this model for any neurophysiological data by proposing a statistical test, which quantifies the odds that \xchanged{overdispersion}{over-dispersion} could be due to the limited number of repetitions (trials). We apply this test to three neurophysiological tests along the visual pathway. Finally, we compare the performance of this model to the Poisson model on a population decoding task. \xchanged{This shows that more knowledge about the form of dispersion tuning is necessary to have a significant gain, uncovering a possible feature of the neural spiking code.}{We show that the decoding accuracy is improved when accounting for over-dispersion, especially under the hypothesis of tuned over-dispersion.}
}%
%%
\newcommand{\Keywords}{Spike counts, \xchanged{overdispersion}{over-dispersion}, Negative-Binomial distribution, Decoding, Tuning function.}%
\newcommand{\Acknowledgments}{%
WT, FC and LP were supported by EC FP7-269921, ``BrainScaleS'', GB by FACETS ITN project (EU funding, grant number 237955), a 'Marie-Curie Initial Training Network' and PW was supported by F32-EY019833 from the NIH. We wish to thank Philippe Foundation for a travel grant support to F.C. The authors are indebted to T. Movshon and his laboratory for giving the opportunity to F.C. to perform MT experiments at CNS/NYU. } %Code and supplementary material :\\ \url{\Website/Publications/Taouali15} %
%}%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\documentclass[a4paper, 10pt, twoside]{article}
%
%\documentclass[a4paper, 11pt]{article} % For LaTeX2e
%\usepackage{times}
%\usepackage[utf8]{inputenc}
%============ common ===================
\usepackage[english]{babel}%
%\usepackage{csquotes}%
\usepackage[autostyle]{csquotes}
%% Sans-serif Arial-like fonts
%\renewcommand{\rmdefault}{phv}
%\renewcommand{\sfdefault}{phv}
%
% \usepackage{textcomp}
% \usepackage{libertine}%[sb]
% \usepackage[varqu,varl]{inconsolata}% sans serif typewriter
% \usepackage[libertine,bigdelims,vvarbb]{newtxmath} % bb from STIX
% \usepackage[cal=boondoxo]{mathalfa} % mathcal
%% \useosf % osf for text, not math
% \usepackage[supstfm=libertinesups,%
% supscaled=1.2,%
% raised=-.13em]{superiors}
\usepackage{hyperref}
\if 1\Draft
\usepackage{setspace}
\usepackage{lineno}
\fi
\usepackage{url}
\usepackage{graphicx}
\usepackage{geometry}
\geometry{hscale=0.78,vscale=0.8,centering}
%\usepackage[utf8]{inputenc}
%\graphicspath{{figures1/}}
\usepackage{tikz}
%\newcommand*\rot{\rotatebox{90}}
\usetikzlibrary{positioning}
%\newcommand{\fix}{\marginpar{FIX}}
%\newcommand{\new}{\marginpar{NEW}}
\listfiles
%\usepackage[ngerman]{babel}
%\usepackage[doi=false,isbn=false,url=false,backend=biber,maxncitetames = 2,dashed=false,style=authoryear,maxbibnames=99,natbib=true]{biblatex}
\usepackage[natbib=true,
style=nature, %numeric-comp,
sortcites=true,
abbreviate=true,
maxcitenames=2,
maxnames = 5,
% firstinits=true,
% uniquename=init,
% sorting=none,
doi=true,
url=true,
isbn=false,
eprint=false,
texencoding=utf8,
bibencoding=latin1,
%autocite=superscript,
backend=bibtex,
%articletitle=false,
]{biblatex}%
\AtEveryBibitem{
\clearfield{month}
\clearfield{day}
\clearfield{note}
\clearfield{comment}
% \clearfield{edition}
% \clearfield{publisher}
}
\addbibresource{decoding.bib}%
%--%\AtEveryBibitem{\clearfield{labelday}\clearfield{labelmonth}}
%--%\usepackage[round]{natbib}
%--%\usepackage[noperiod]{jabbrv}
%--%\DeclareFieldFormat{titlecase}{\MakeSentenceCase{#1}}
\usepackage{fancyhdr}
\pagestyle{fancy}
\rhead{\RunningTitle}
\lhead{\AuthorWT\ {\it et al} (2015) \href{http://dx.doi.org/10.1152/jn.00194.2015}{doi:10.1152/jn.00194.2015}}
%--%\DeclareNameAlias{sortname}{last-first}
%--%\DeclareNameAlias{default}{last-first}
\title{\Title}
\date{1 January 2016}
\author{{\AuthorWT}$^{1}$, \AuthorGB$^{1}$, \AuthorPW$^2$,\\
\AuthorFC$^1$ and \AuthorLuP$^1$\\
}
%\usepackage{trackchanges}
\newcommand{\xchanged}[2]{#2}
\newcommand{\xremoved}[1]{}
\newcommand{\xadded}[1]{#1}
%\addeditor{Edition}
% \xadded[editor]{added text}
% \xremoved[editor]{removed text}
% \xchanged[editor]{removed text}{added text}
% \note[editor]{note text}
\begin{document}
\if\Draft1
\linenumbers
\fi
\maketitle
\paragraph{Institutions}~~\\
$^{1}$ \AddressA\\
$^{2}$ \AddressB\\
\paragraph{Abbreviated title for the running head}~~\\
\RunningTitle
\paragraph{Corresponding author contact}~~\\
Name : \AuthorWT \\
E-mail: \EmailWT\\
Phone number : \PhoneWT\\
Physical address: \LongAddressA \\
%----------------------------------------------------------------------------------------------%
\begin{abstract}
\Abstract
\end{abstract}
%\smallskip
%\indent\indent\indent
\textbf{Keywords} : \Keywords
%Fano factor
%--------------------------------------------------------------------------------------------------%
\paragraph{BibTex entry}~~\\
\begin{verbatim}
@article{Taouali15,
title = {Testing the Odds of Inherent versus Observed
Over-dispersion in Neural Spike Counts},
author = {Taouali, Wahiba and
Benvenuti, Giacomo and
Wallisch, Pascal and
Chavane, Frederic and
Perrinet, Laurent U.},
journal = {Journal of Neurophysiology},
year = {2015},
doi = {10.1152/jn.00194.2015},
issn = {1522-1598},
volume = {115},
number = {1},
pages = {434-444},
url = {http://dx.doi.org/10.1152/jn.00194.2015},
url = {http://invibe.net/LaurentPerrinet/Publications/Taouali15},
pmid = {26445864},
publisher = {American Physiological Society},
}
\end{verbatim}
\if\Draft1
\newpage
\doublespacing
\fi
%--------------------------------------------------------------------------------------------------%
\section*{Introduction}
%--------------------------------------------------------------------------------------------------%
One of the most common ways to study the central nervous system is to use a stimulus-response paradigm. A system, which could be a neuron or a population of neurons, is provided with an input (a stimulus) and the resulting state change (response) of the system is quantified in order to better characterize functional associations. A main problem with this approach is that the system's response to repeated presentations of an identical visual stimulus often exhibits high variability, particularly in the spike count statistics (see Figure~\ref{fig:obs}). As a consequence, statistical methods are essential to characterize the information carried by neural populations.
%The application of Fano's method to neuroscience (and indeed biology) happened very late. For cases of variability, I like to (also) cite the original papers by Fano, where he struggled with similar issues in physics. To my knowledge, this is the first time anyone did: http://journals.aps.org/pr/abstract/10.1103/PhysRev.72.26
The Fano factor~\citep{Fano47} is commonly used to quantify the variability of spike counts~\citep{Kara00, Alitto11, Zhuang13,Tolhurst83, Softky93, Shadlen98, Koch99}. It is defined by $FF(\Delta t) = \frac{\sigma^2(\Delta t)}{\mu(\Delta t)}$ where $\mu$ and $\sigma^2$ are the average and the variance over multiple trials of the number of spikes in some time window of length $\Delta t$, respectively. %: Poisson model
This variability is most often characterized by the Poisson spiking Model (PM). By the definition of this model, the spike count follows a homogeneous Poisson process with a rate parameter ${\lambda}$ (the expected number of spikes that occur per unit of time). If $k$ is the observed number of spikes fired by a neuron in response to a stimulus %defined by a variable $\theta$%
in a time window $\Delta t$, the probability of such an observation is given by a Poisson distribution of parameter $f=\lambda \Delta t$, which gives the expected number of spikes per sample in the considered time window:
\begin{equation}
P(k) = \frac{f ^{k}e^{-f}}{k!}
\end{equation}
It follows that $\mu=f$ and $\sigma^2=f$ and thus that the Fano factor under the Poisson model has a value of $1$, regardless of $\lambda$ or $\Delta t$.
%%%%
Such a property is in accordance with a wide range of observations of variability in cortical responses to visual stimuli~\citep{Shadlen98,Softky93,Britten93,Buracas98,McAdams99} showing variance-to-mean ratios of spike counts that equal or just exceed unity. However, for a large majority of reported evoked responses, the variability is higher than expected from a Poisson model, \xchanged{specially}especially in mammalian visual cortex \citep{Heggelund78, Dean81, Schiller76, Tolhurst83, Vogels89}, \xadded{whereas} Sub-Poisson variability has been observed only under specific conditions \citep{Kara00}.
In fact, the variability of neural responses depends on different parameters. Some studies reported that the Fano factor values can vary depending on the brain region~\citep{Kara00, Gur06, Kayser10}. Indeed, neural responses in higher cortical areas~\citep{Maimon09} or in other non-cortical sensory areas such as the retina~\citep{Berry97} and the thalamus~\citep{Kara00} show substantially lower variability. It has been also shown that some neural features such as the refractory period~\citep{Kara00, Berry97} lead to a decrease of variability. Moreover, the nature of the stimulus itself could contribute to the variability of neuronal responses. It has been suggested that increasing the complexity of the stimuli by using natural-like statistics increases the reliability~\citep{Baudot13,Haider10,Borst99}. In addition, it has been reported that the variability of evoked responses is not stationary even during a trial time-course and decreases after the presentation onset of most sensory stimuli \citep{Churchland10}. Finally, a recent study~\citep{Nienborg12} on behaving animals \xchanged{focused on}{explored} the role of decision and other cognitive processes in shaping variability.
%--------------------------------------------------%
%: -observations
% ------------------------------------------------- %
%-------------------------------------------------%
%: fig:obs
\begin{figure}[ht]
\includegraphics[width=\textwidth]{fig1.pdf}
\caption{{\bf Trial-to-trial variability of neural responses.}
(Left)~We show the spiking activity of a representative cell of the area MT of a macaque monkey (see Methods section) in response to oriented drifting gratings. Each subplot corresponds to a raster plot of the responses to 16 equally distributed motion directions. There is a high trial-to-trial variability in the spiking responses. %
(Right)~To quantify this variability, we plot \xchanged{for a population of 16 cells from the same area the directional tuning curves in response to the same 16 directions (mean spike count, black lines) along with the range of {$\pm$} the standard deviation (red lines)}{the directional tuning curves in response to the same 16 directions (mean spike count, black lines) along with the range of {$\pm$} the standard deviation (red lines) for a population of 16 cells from the same area}. The subplot with a star corresponds to the representative cell on the left. The area corresponding to the mean {$\pm$} the standard deviation expected under Poisson distribution hypothesis {($\sigma=\sqrt \mu$)} is plotted in blue. In this experiment, most of the cells show a higher variability than expected from the Poisson model.}
\label{fig:obs}
\end{figure}
%------------------------------------------------- %
%------------------------------------------------- %
In this paper, we mainly explore \xchanged{overdispersion}{over-dispersion}, that is, cases where there is greater variance in the spike count than might be expected by the Poisson model. Our hypothesis is that the Poisson process only accounts for the intrinsic noise induced by the spiking mechanism itself. Indeed, under the Poisson process hypothesis, the trial-to-trial variability only results from the stochastic properties of the neuron~\citep{Carandini04, Mainen95, Schneidman98}. Thus, the PM can not account for other significant sources of variability including hidden contextual variables such as attention, varying cortical states~\citep{Arieli96,Buracas98,Tsodyks99,Kenet03}, perceptual effects~\citep{Ress03} and overt behaviors (such as the precise eye position)~\citep{Gur97}. Thus, we consider an alternative compound distribution, called Negative-Binomial, that generalizes the Poisson distribution with a dispersion parameter that directly controls the ratio between the mean and \xremoved{variance }variance. Such a doubly-stochastic model was recently proposed to model the variability of neural dynamics showing more complex variability than what is expected from a Poisson model~\citep{Churchland11, Ponce13, Ponce10,Goris14}. Our aim is to further explore the potential suitability of this doubly-stochastic model to account for response variability in neurophysiological data.
%: Outline:
The paper is organized as follows. First, we describe the Negative-Binomial encoding model (NBM) and two possible implementations of \xchanged{overdispersion}{over-dispersion} sources. Second, we propose a statistical test for evaluating whether the observed \xchanged{overdispersion}{over-dispersion} is significantly better described by the Negative-Binomial model than the Poisson model. To assess the generality of this model, we then apply this test to three data sets of spiking responses in different animals, different states and different visual stimuli: in the lateral geniculate nucleus (LGN) neurons of anaesthetized mice presented with drifting gratings, in the primary visual cortex (V1) of awake macaque monkeys viewing oriented moving bars and in the middle temporal cortex (MT) of anaesthetized macaque monkeys presented with drifting gratings. Finally, we compared the efficiency of PM and NBM models on a population decoder using the most dispersed data set (MT).
% ------------------------------------------------- %% ------------------------------------------------- %%
%---------------------------------------------------------------------------------------------------------------- %
\section{ The Negative-Binomial model: encoding of inter-trial \xchanged{overdispersion}{over-dispersion}}
%--------------------------------------------------------------------------------------------------%% ------------------------------------------------- %
%We illustrate this approach with the negative binomial regression model, which is the best known and most widely available Poisson-based regression model that allows for \xchanged{overdispersion}{over-dispersion}.
Several models have been developed to account for over-dispersion such as generalized Poisson~\citep{Consul73}, zero-inflated Poisson~\citep{Lambert92, Miaou94, Shankar97} or quasi-Poisson~\citep{ Mccullagh89, Wedderburn74} models. However, \xadded{only a} few such as the Negative-Binomial model (NBM) are well known and widely available. \xremoved{Indeed, }The Negative-Binomial model has been \xremoved{largely} studied and tested in different fields such as epidemiology~\citep{Byers03}, accident statistics~\citep{Poch96}, biology~\citep{Anscombe50, Bliss53}, etc. However, it has only recently been applied to evoked neural responses~\xchanged{\protect \citep{Goris14}}{\protect \citep{Scott12,Goris14}} where the variability was traditionally supposed to be well modeled by a Poisson model (PM).
A Negative-Binomial model depends on a distributional form equivalent to a compound stochastic process. \xadded{It has recently been represented as a Polya-Gamma mixture of normals \protect \citep{Polson12}}, but its standard representation is the Gamma-Poisson mixture. It corresponds to a doubly stochastic process where the Poisson process has a probability distribution $\mathcal{P}$ with a parameter $\lambda$ which is itself a random variable generated from a Gamma distribution $\mathcal{G}(\phi,\beta)$ with a shape parameter $\phi$ and a scale parameter $\beta$. When we consider a random variable $x$ generated by this compound process (i.e. $x \sim \mathcal{P}(\lambda)$ with $\lambda \sim \mathcal{G}(\phi,\beta)$), the resulting distribution of this variable corresponds to a Negative-Binomial distribution (NB) which is usually defined by its shape parameter $\phi$ and its scale parameter $p=\frac 1{\beta+1}$:
\begin{equation}
x \sim \mathcal{NB}(\phi, p=\frac 1{\beta+1})
\label{eq:NB}
\end{equation}
Using the Negative-Binomial model, the probability of getting $k$ spikes from a given cell with parameters $\phi$ and $p$ in response to a specific stimulus is given by:
\begin{equation}
P(k; \phi, p)=\frac{\Gamma(k+\phi) \cdot p^{\phi}\cdot(1-p)^{k}}{\Gamma(\phi)k!}\end{equation}
We deduce the mean ($\mu$), the variance ($\sigma$) and the Fano factor (${FF}$) as:
\begin{equation}
\mu= \phi \cdot \beta \mathrm{, ~} \sigma^2 = \mu + \frac{\mu^2}{\phi}\mathrm{~and~{FF}} = \frac{\sigma^2}{\mu}=1 + \frac{\mu}{\phi}
\end{equation}
This distribution approaches the Poisson probability of parameter $\mu$ for large values of $\phi$, but allows to reach larger-than-Poisson variances for small $\phi$ values. We end up with a multivariate distribution \xadded{that is} - in practice - \xremoved{is} driven by two parameters: the mean spike count ($\mu$) and the parameter ($\phi$) that we call the {\em ``inverse-dispersion parameter"} \footnote{ The inverse of the parameter $\phi$, usually denoted by $\alpha$, has been often called in the literature as the {\em ``dispersion parameter"}. It has been occasionally named the {\em ``\xchanged{overdispersion}{over-dispersion} parameter"} because it does not account for FF values smaller than one.}.
\xadded{Under the NBM hypothesis, the distribution of inter-spike intervals (ISIs) is given by a Loamax distribution (equivalent to an exponential distribution with a random parameter generated from a Gamma distribution). The mean and the variance of such a distribution are given by the following expressions: mean(ISIs) = $\frac {\beta} {(\phi-1)} $ defined for $\phi >1$, variance(ISIs) = $\frac {\beta^2 \phi} {(\phi-1)^2(\phi-2)} $ defined for $\phi >2$. The resulting squared coefficient of variation is given by :
\protect \begin{equation}
CV^2 = \frac {\phi} {(\phi-2)}=1+\frac {1} {\frac \phi 2-1}
\protect \end{equation}
defined for $\phi >2$. $CV^2$ is higher than 1, which implies than a Negative-Binomial process is more irregular than a Poisson process and approaches Poisson regularity for large values of $\phi$. However, unlike under the Gamma distribution hypothesis made in \protect~\citep{Nawrot08}, the Negative-binomial model do not allow us to systematically relate over-dispersion to the positive deviation from the null-hypothesis $FF=CV^2$.} Thus, to better qualify this \xchanged{overdispersion}{over-dispersion} parameter, we need to focus on how to generate the Gamma input drive that feeds the Poisson generator.
%------------------------------------------------- %
%: fig:model1
\begin{figure}[t]
\includegraphics[width=\textwidth]{fig2.pdf}
\caption{{\bf Over-dispersion may result from extrinsic sensory noise\xadded {: NBM as a mixture of Poisson and LogNormal distributions.}} We show here a simple linear / non-linear spiking model with three processing steps: (A)~The correlation between an noisy Gabor image (input) and a population of Gabor filters at different orientations models the linear processing in the receptive field (RF) of a visual neuron. The green, black and red lines correspond respectively to one trial (simulation), the average and the average $\pm$ the standard deviation (within 1000 trials). (B)~This input is transformed using a static, exponential non-linearity multiplied by a scaling factor corresponding to the maximum spike count. \xadded{The input drive resulting from a Gaussian input noise and the non-linearity follows a Log-normal distribution that is closely fitted by a Gamma distribution}. (C)~This drive is finally transformed into spike counts using a Poisson point process with \xadded{the} mean \xadded{as} the input drive. The resulting tuning curve (mean spike count) is plotted in black along with the range of {$\pm$} the standard deviation (red line). The area corresponding to the mean {$\pm$} the standard deviation expected given the Poisson model {($\sigma=\sqrt \mu$)} is plotted in blue. The resulting spike counts variance is higher than what is expected by a Poisson model, as can be observed in neurophysiological data (see Figure~\ref{fig:obs} - Right).
}
\label{fig:model1}
\end{figure}
%------------------------------------------------- %
%: Gamma as lognormal
\paragraph{Over-dispersion may result from extrinsic sensory noise: \xadded {NBM as a mixture of Poisson and LogNormal distributions.}} Many studies \xremoved{indistinctly} assume a Log-Normal or a Gamma distribution when analyzing spike count distributions as it most often produces similar results~\citep{ Mccullagh89,Atkinson82}. Indeed, a Log-Normal process might be a good approximative model for a Gamma process. The underlying spiking process could result from a simple encoding model: Let us assume that a noisy sensory input is transformed to a response by a linear correlation measure with a cell's receptive field (for instance, in the early visual system, selective to a given orientation) and then through a non-linearity into a stochastic spike count (Figure~\ref{fig:model1}). If we consider \xremoved{an} additive extrinsic Gaussian noise as the sensory noise and an exponential function as the non-linearity, the resulting input drive follows a Log-normal distribution that is closely fitted by a Gamma distribution. Using such a model, \xremoved{the}Figure~\ref{fig:model1} shows that \xchanged{overdispersion}{over-dispersion} could simply result from an additive normal sensory noise and the non-linearity of the spiking response.
However, even if log-Normal and Gamma distributions are often used to model the same phenomena, looking at their log functions - in general - shows clear differences in their respective skewness. Furthermore, there is no exact morphing between the two distributions that allows to link their respective parameters analytically. As such, this approximation does not allow a precise quantification of the relationship between the noise source and the \xchanged{overdispersion}{over-dispersion} parameter.
%: Gamma as sum of exponential distribution
\paragraph{Over-dispersion may result from redundancy within a neural population: \xadded {NBM as a mixture of Poisson and Exponential distributions.}}
The Gamma distribution could also result from a mixture of distributions. Indeed, it is well known that the sum of $n$ independent and identically distributed exponential random variables $\mathcal{E}(\frac 1\lambda)$ (the parameter of an exponential distribution is by definition the inverse of its mean, here called $\lambda$) follows a Gamma distribution $\mathcal G(n,\lambda)$, where $n$ and $\lambda$ are its shape and scale parameters respectively. As such, exponential distributions have previously been used as a simple model of variability in formal neural networks~\citep{Treves91, Levy96, Baddeley96}. They correspond to a model of the distribution of time between two Poisson events. In addition, such a distribution maximizes the entropy of spike counts (knowing the average count) and thus maximizes the carried information~\citep{Shannon48}.
%------------------------------------------------- %
%: fig:model22
\begin{figure}
\includegraphics[width=\textwidth]{fig3.pdf}
\caption{{\bf Over-dispersion may result from redundancy within a neural population.\xadded{NBM as mixture of Poisson and Exponential distributions.}} Let us consider, within a neural population, a subset of $\phi$ cells with a similar selectivity and thus driven by a similar correlation measure $\lambda$= $\mathcal C(\theta)$. This measures the correlation between the input Gabor and the Gabor representing the orientation $\theta$ characteristic of their receptive field (RF). Let us further assume that the filter cells exhibit an exponential distribution of spike counts $\mathcal E(\frac{1}{\lambda})$ with mean $\lambda$, thus maximizing the entropy of the input drive. Then, the pooling of the resulting spike counts results in an input drive $\mathcal G{(\phi,\lambda)}$ equal to the sum of exponential variables having the same mean equal to $\lambda$. It can be shown that the Poisson spiking process driven by this input drive is equivalently described by a Negative-Binomial probability distribution function with parameters $\lambda$ (mean) and $\phi$ (inverse-dispersion) (see equation~\ref{eq:NB}). Such a distribution describes \xchanged{well}{observed} trial-to-trial variability \xadded{well} and in this model, the parameter of dispersion is directly correlated with the redundancy, i.e., the convergence of similar input cells. }
%%%%
%The Gamma input to the Poisson generator with an inverse-dispersion parameter $\phi$ is driven by the sum of $\phi$ exponential variables having the same mean $\lambda$. These exponential variables could result from the correlation measure $\lambda(\theta)$ between the stimulus and different cells having the same filter under a multiplicative noise $-log(\mathcal U_{(0,1)})$, where $\mathcal U_{(0,1)}$ is a uniform distribution between 0 and 1. Thus, in this case, the parameter of dispersion is directly correlated with the redundancy (i.e. the convergence of similar input cells). }
\label{fig:model22}
\end{figure}
%------------------------------------------------- %
%
From a biological point of view, to drive a Gamma input - with an inverse-dispersion parameter $\phi$ equal to $n$ - to the output neuron considered as a Poisson generator, this neuron should receive as input the sum of $n$ exponential variables having the same mean $\lambda$ (see Figure~\ref{fig:model22}). Thus, $\phi$ reflects the amount of convergence to a neuron which is - for instance - consistent with a model of the convergent connectivity from V1 to area V5/MT~\citep{Simoncelli98, Rust06, Wang12}. In addition, it was observed that the simulation of networks based on biologically realistic parameters~\citep{Voges10neurocomp} could lead to complex dynamics, showing in particular an excess of variability~\citep{Voges12}. Thus, we could expect an \xchanged{overdispersion}{over-dispersion} in V5/MT resulting from dimensionality reduction (i.e. projecting inputs from a high dimensional space to a lower dimensional one~\citep{Haykin99}).
To summarize, we have focused on two main possible sources of \xchanged{overdispersion}{over-dispersion} in spike counts. The first, is an extrinsic source that accounts for sensory noise, \xchanged{for instance it could correspond to the thalamo-cortical pathway. The second is a source of noise that emerges from convergent inputs to a population, as one commonly sees in the cortical hierarchy, e.g. the visual processing cascade.}{ conveyed along the thalamo-cortical pathway. The second is a source of noise that emerges from convergent inputs to a population, including feedforward, lateral and feedback.} We have \xchanged{unveiled}{shown} that the convergence of inputs between two data spaces may both result in an \xchanged{overdispersion}{over-dispersion}. In the next section, we will focus on quantitatively testing evidence for this \xchanged{overdispersion}{over-dispersion}.
%But even if the Negative-Binomial model has been considered as the model of choice allowing to account for over-dispersion~\citep{Osgood00}, several studies were specifically interested on the special circumstances under which this claim is true~\citep{Berk07}.
%%------------------------------------------------- %%%
\section {The Fano Gamma test: How to quantify \xchanged{overdispersion}{over-dispersion} evidence?}
% ------------------------------------------------- %% ------------------------------------------------- %% ------------------------------------------------- %
%introduction%
\paragraph {The Fano factor test is misleading} Most studies concerned with the characterization of spike count variability using different methods and formal tests start with the Fano factor (FF). However, most of the methodological studies measuring trial-to-trial variability in firing rates~\citep{Teich97, Koch99, Dayan01} do not take the estimation uncertainty introduced by the limited number of trials into account. Each trial may be considered as a realization of the random process and when accumulating evidence from $N$ trials, the average and the variance converge with an expected error of $1/\sqrt{N}$. Instead, the computed mean and variance are often treated as real features of the distribution whereas they actually correspond to estimations given the (necessarily) limited set of measurements. Thus, considering only the FF value may induce errors in conclusions regarding the assumption of probability distribution. In particular, when the observed FF deviates from $1$, is this the result of the limited number of trials or due to an \xchanged{overdispersion}{over-dispersion}? Or when a value is measured close to $1$, can it be safely attributed to a Poisson model?
%: -previous methods
\paragraph {The chi-square test is problematic} To evaluate how likely the observed data was generated by a Poisson model, the most common approach is to use a chi-square goodness of fit test ($\chi^2$). This statistical tool can be used to estimate how closely an expected distribution matches an observed distribution of a discrete quantitative variable (having only finite possible values). However, it has been shown that small expected frequencies may result in a loss of power, that is, a tendency to not reject a false null hypothesis (or to reject a true null hypothesis)~\citep{Cochran54, Yates99, Campbell07}. This is a potential problem when we deal with small sample sizes and its outcome also depends on how data is pooled into categories. Indeed, it often makes sense to pool the small frequencies together. Moreover, \xadded{as} this test is one-tailed \xremoved{test}, \xremoved{so} it does not distinguish evidence for \xchanged{overdispersion}{over-dispersion} from evidence for \xchanged{underdispersion}{under-dispersion}.
%------------------------------------------------- %
%: Fano Gamma test
\paragraph {The Fano Gamma test} To better quantify the \xchanged{overdispersion}{over-dispersion} while accounting for sample size and without dealing with pooling problems, we propose a statistical test based on an empirical sampling distribution of the Fano factor under the Negative-Binomial assumption. Here, we opt for an analytic procedure that we call Fano Gamma test ($FG$). It allows for a simple computation (or tabulation) of probability bounds for specific distributions hypotheses. That is not the case for other existing procedures~\citep{Amarasingham06, Jackson07, Nawrot08}. This procedure characterizes the probability distribution of the estimated FF. Thus, it allows one to compute FF bounds with respect to p-value precision limits. Starting from previous analyses~\citep{Hoel43, Kathirgamatamby53, Eden10} about the distribution of the Fisher dispersion index, we propose here that, for a Negative-Binomial distribution, the FF asymptotically follows a Gamma distribution with a shape parameter $\frac{n-1}{2}$ and a scale parameter $\frac{2(\frac\mu\phi+1)}{(n-1)}$:
\begin{equation}
FF_{Negative-Binomial} \sim \Gamma(\frac{n-1}{2},\frac{2(\frac\mu\phi+1)}{(n-1)})
\label{eq:FF_NB}
\end{equation}where $n$ is the number of spike count observations. We tested the accuracy of this approximation through simulation by comparing it to the empirical estimates of the true distribution of the Fano factor for any ($\mu$, $\phi$, $n$) triplet. We found that the p-value upper bounds resulting from the approximative formula are slightly over-conservative. Thus, it is worthwhile mentioning that using it to reject the NBM hypothesis is based on a strong assumption (i.e. a rejected realization could be a good one because it's p-value is between the approximation and the empirical upper bound) but we do not detail this analysis here. For large values of $\phi$, the Negative-Binomial distribution approaches a Poisson distribution and the previous formula gives:
\begin{equation}
FF_{Poisson} \sim \Gamma(\frac{n-1}{2},\frac{2}{n-1})
\label{eq:FF}
\end{equation}
\citet{Eden10} studied the FF distribution under the Poisson distribution hypothesis, and have shown that the convergence to this asymptotic distribution is fast. Thus, this analysis provides a formal test for both hypotheses (PM and NBM) given the observed variability in the spiking.
%------------------------------------------------- %
%: fig:power
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]{fig4.pdf}
\end{center}
\caption{{\bf Evaluating the power of chi-square test ($\chi^2$) test versus Fano Gamma test ($FG$) when testing evidence against Poisson (PM) model (in blue) and Negative-Binomial (NBM) model (in red)}. (Top) Two examples of spike count histograms resulting from artificially generated data (surrogate data) using the NBM (A) and the PM (B). These examples were chosen such that, when using the $\chi^2$ test, the model used to synthesize the data is rejected. For each histogram, we plotted in blue the expected distribution from a PM and in red the expected distribution from a NBM. For these examples, the sample size is 50. (Bottom) $\chi^2$ test (dashed lines) and $FG$ test (solid lines) were applied to Negative-Binomial (C) and Poisson (D) surrogate data for different sample sizes with a significance level of 0.025. The mean is fixed to 10 spikes, the dispersion parameter is equal to 1 spike and the test was repeated 1000 times for each sample size. (E) and (F) show the results in (C) and (D), respectively, corresponding to small sample sizes. The rejection rate is higher for $\chi^2$ test when the null hypothesis is the correct distribution (the model generating the data) and lower for the incorrect distribution than for the proposed Fano Gamma test. The results thus show this test is more powerful than the classically used $\chi^2$ test. }
\label{fig:power}
\end{figure}
%------------------------------------------------- %
We first compared the power of this test to the chi-square test for the general NBM hypothesis and for the specific PM hypothesis on surrogate data of spike counts. Figure~\ref{fig:power} shows that the Fano Gamma test did better than a chi-square goodness of fit test at a significance level of 0.025. This latter test has a larger probability to make a type $II$ error for both distributions (dashed red lines in Figure~\ref{fig:power}.C and dashed blue lines in Figure~\ref{fig:power}.D).
\paragraph {Over-dispersion evidence from neurophysiological data} We applied these tests to evaluate evidence against\xadded{\footnote{Indeed, an hypothesis test is a statistical analogy to proof by contradiction. Thus, we used the ``evidence against'' term that is the result of the statistical test and that means the rejection plausibility. Importantly, the evidence against a claim $\mathcal{H}0$ does not imply the evidence (plausibility) of the opposite claim $\mathcal{H}1$, i.e., this is not the same as a likelihood based test that allow explicitly to quantify the weight of evidence for one hypothesis.}} the Poisson model on three data sets.
The first data set results from recording \xadded{signals from} \xremoved{in} mouse LGN \xchanged{cells}{neurons} (72 cells) in response to moving \xremoved{test}gratings. The stimulus was presented several times in 12 different directions. The second data set comes from extra-cellular recordings in area V1 (67 cells) of two awake macaque monkeys in response to an oriented moving bar. The task was repeated several times for 12 different directions (orthogonal to the bar's orientation). The third data set results from recording in area MT (40 cells) of three anaesthetized macaque monkeys in response to moving gratings. The stimulus was presented several times in 16 different directions. See the methods section for more details on the data collection.
We show the results in Figure~\ref{fig:test}. For the three data sets, we found that the Fano Gamma test gives lower evidence against the PM and the NBM compared to the FF and $\chi^2$ tests (as expected given our discussion above). The Fano Gamma test shows, as do the other tests, that the evidence of \xchanged{overdispersion}{over-dispersion} is only significant in MT (with a rate of $64\%$) among cell$/$condition pairs, compared to LGN and V1 (with rates of $19\%$ and $26\%$, respectively). Similarly, it shows that evidence against the NBM is significantly lower in the MT dataset ($15\%$) compared to LGN and V1 data sets ($52\%$ and $41\%$, respectively).
%------------------------------------------------- %
%: fig:test
\begin{figure}
\includegraphics[width=\textwidth]{fig5.pdf}
\caption{{\bf Evaluating evidence against Poisson and Negative-Binomial models on visual neurophysiological data} (A) The Fano Gamma test (FG) results : Evidence against the Poisson model. Each subplot shows the measures of variability for the different cells and conditions as points on a (mean, variance) plot. The straight line corresponds to FF=1 (Poisson), while dots correspond to observations. Blue dots correspond to observations with no evidence against Poisson, red dots correspond to no evidence against the Negative-Binomial model (NBM) and black dots correspond to strong evidence against the Poisson model (PM) on the top and against NBM on the bottom based on the empiric distribution of the Fano factor (FG test). The results show that there is more evidence of \xchanged{overdispersion}{over-dispersion} in MT and that the proportion of cells showing \xchanged{overdispersion}{over-dispersion} seems to increase along the visual hierarchy. (B) The percentage of observations with evidence against PM and NBM are given with respect to three criteria: the Fano factor test discards the observations with a Fano factor value higher or lower than its formal expected value (equal to one for the Poisson model), the $\chi^2$ test and the $FG$ test with a significance level of 0.025. The bar plots correspond to the percentage of observations with evidence against the NBM (left) and the PM (right). The tests were applied to 3 data sets (from the left to the right, respectively): LGN cells (mouse), V1 cells (awake macaque monkey) and MT cells (anaesthetized macaque monkeys). The results show that the Fano factor values are misleading when the number of trials is not taken into account. Looking only at the FF gives strong evidence against PM and NBM that is not confirmed by the other tests. The $\chi^2$ test shows slightly more evidence against the Poisson model than the FG test which could be explained by categorization issues. It shows also very strong evidence against the NBM model known to be suitable for over dispersed data. In summary, the FG test gives more coherent results: weak evidence of \xchanged{overdispersion}{over-dispersion} in LGN and V1 data sets and low evidence against the NBM in MT data set. }
\label{fig:test}
\end{figure}
%------------------------------------------------- %
%%%%%%%%%%%%%%%%%%%%%%%%%%
For these particular datasets where the number of trials is limited (mostly less than 20 trials per condition), we have shown that in the LGN and V1 data sets, the high values of Fano factor observed could be simply related to the sampling of the stochastic process itself. Indeed, looking only at the FF values induces a false belief that these datasets exhibit a significant degree of \xchanged{overdispersion}{over-dispersion}, while the FG test reliably rejects this hypothesis. The $\chi^2$ test shows slightly more evidence against the Poisson model than the FG test which could be explained by categorization issues. It shows also very strong evidence against the NBM model known to be suitable for over dispersed data. Finally, we have shown that the situation is different for the MT data set where we could indeed show evidence of a higher variability than expected from a Poisson model and less evidence against a Negative Binomial model. It is worth mentioning that our statistical test (as validated on surrogate data) allows us to make this conclusion, while other tests fail to do so. Since we mostly observe \xchanged{overdispersion}{over-dispersion} in the MT data set, we propose to compare the performance of an NBM-based to a PM-based direction decoder applied to this data set. %It also makes sense physiologically, given the Simoncelli and Heeger model~\citep{Simoncelli98}.
%%%%%%%%%%%%%%%%%%%%%%%%%%
%-------------------------------------------------------%%------------------------------- %
\section{NBM-based direction decoder: what if the \xchanged{overdispersion}{over-dispersion} is tuned ?}
% ------------------------- %% ------------------------------------ %% ------------------------------ %
%In this section, we would like to focus on a main problem Negative-Binomial specially for modeling inter-trial spike count variability when decoding.
%The estimation of dispersion parameter has been widely studied (Fisher, 1941, Anscombe, 1950, Shenton and Wallington, 1962, Ross and Preece, 1985, Dean 1994, Gourieroux and Visser, 1986, Cameron and Trivedi, 1986, 1990, Lawless, 1987, Clark and Perry, 1989, Piegorsch, 1990 and Saha and Paul, 2005, Hongsheng, 2013) in different fields () and several estimation methods were provided . Most of the studies have shown this estimation is very challenging and can be significally biased specially when dealing with small data sets (Clark and Perry 1989, Piegorsch 1990, Dean 1994, Lord 2006)
%%: describing potential biases:
%
%unstable variance and issues related to small sample sizes (Pieters et al., 1977; Willson et al., 1984, 1986; Davidian and Carroll, 1987; Clark and Perry, 1989; Piegorsch, 1990; Dean, 1994; Saha and Paul, 2005).
%
%Piegorsch, 1990, Dean (1994) and Cameron and Trivedi (1998) for additional information on different estimation techniques (Clark and Perry, 1989), (Ross and Preece, 1985))
%
%%bias: effect of small sample sizes
%The estimated dispersion parameter of NB models can be significantly affected when the data are characterized by small sample size and low sample mean values (Piegorsch 1990, Dean 1994, Lord 2006).
%When this occurs, many estimators provide a biased estimate of the dispersion parameter (the distribution becomes highly skewed) and has a high probability of being mis-estimated
%Clark and Perry (1989) provide a review of earlier work and investigate method of moment (MM) and maximum extended quasi-likelihood (EQL) estimates of alpha or different sample sizes and sample means
%Piegorsch (1990) discusses its maximum likelihood (ML) estimation.
%Dean (1994) evaluated the effects of small sample sizes on the estimation of the dispersion parameter for seven different estimators, including the MLE and MM.
%Lord (2006) reported that no matter which estimator, MM, MLE and Weighted Regression (Cameron and Trivedi, 1998) is used, the dispersion parameter is very likely to be mis-estimated when the data are characterized by extreme conditions.
%%%: correction
%Saha and Paul (2005) expanded on the work of Clark and Perry (1989) and Piegorsch (1990) and proposed a biased-corrected maximum likelihood (BCML) estimator for estimating the dispersion parameter of a NB regression model. and study its bias and efficiency properties in comparison with the ML, MM, and EQL estimates. They also construct a double-extended quasi-likelihood (DEQL) estimate of c using a more recent methodology by Lee and Nelder (2001). This estimator is also included in the bias and efficiency comparison.
% %Airoldi et al. (2006) proposed a modified parameterization of the NB distribution.
%Cox and Snell (1968) provide general results for first-order correction of biases of maximum likelihood estimators.
%\paragraph { Decoding : lack of knowledge about dispersion tuning.}
%Approach: Maximum likelihood direction decoding:
To evaluate the gain in decoding when accounting for the \xchanged{overdispersion}{over-dispersion} using NBM compared to PM, we extended the classical probabilistic decoding approach proposed by~\citet{Jazayeri06} and tested the population decoder on the MT data set (presenting the highest \xchanged{overdispersion}{over-dispersion} probability). Based on a Poisson model of trial-to-trial variability applied to single neurons and a model of tuning for a population coded variable (direction $\theta$). This decoding approach allows us to infer the stimulus (the hidden variable $\theta$) given a particular neuronal response (an observed spike raster Y).
The decoding algorithm consists of maximizing the posterior probability $P({\theta}|Y)$ function of the estimated direction ${\theta}$ given a distribution hypothesis. Bayes' rule gives: $P({\theta}|Y)=\frac{ P(Y|{\theta}) P(\theta) }{P(Y)}$, where the evidence term $P(Y)$ is a normalization term independent of ${\theta}$. So, if we assume that there is no prior knowledge on ${\theta}$ ($P(\theta)$ constant), maximizing the posterior under the PM hypothesis is equivalent to maximizing the following log-likelihood function $LP(\theta)$:
\begin{equation}
LP(\theta) = \sum_{i=1}^N{k_i\log[f_{i}(\theta)]}-\sum_{i=1}^N{f_{i}(\theta)}
- \sum_{i=1}^N \log[{k_i!}],\label{eq:lp}
\end{equation} where $f_{i}$ \xremoved{corresponds to the mean response of each cell $i$}is a function of the stimulus parameter $\theta$ \xadded{and corresponds to the mean response of each cell $i$}, $k_i$ is the number of trials and $N$ is the number of cells. Similarly, under the NBM hypothesis, we maximize the following log-likelihood function $LNB(\theta)$ in order to estimate ${\theta}$, where $p_i(\theta)=\frac{\phi_i(\theta)}{\phi_i(\theta)+f_i(\theta)}$ and $\phi_i$ corresponds to the inverse-dispersion parameter of each cell $i$ as a function of the stimulus parameter $\theta$: %(Equation~\ref{eq:lnb}).
\begin{equation}
LNB(\theta) = \sum_{i=1}^N{\log[\frac{\Gamma(k_i+\phi_i(\theta))}{\Gamma(\phi_i(\theta))k_i!}]} +\sum_{i=1}^N {\phi_i(\theta)\log[p_i(\theta)]}+\sum_{i=1}^N {k_i\log[1-p_i(\theta)]}, \label{eq:lnb}
\end{equation}
%\subsection*{Tuning curves}
Given that objective, let us first define the tuning functions of the cells. These correspond to the mean spike count (the mean parameter of the Negative-Binomial model) with respect to direction. Thus, we defined for the MT data set a generative tuning function as von Mises functions peaking on the preferred direction. The parameters corresponding to the scaling, maximum and minimum values ($\kappa_o$, $\kappa_d$, $R_0$, $R_0$+$R_m$) are free to vary. The main tuning function is given by:
%s for V1 and MT are respectively given by $f_{V1}$ and
%\begin{equation}
%f_{V1}(\theta) = R_0 + R_m \cdot \frac{e^{\kappa_o\cdot(\cos(2(\theta-\theta_0))-1)}}{e^{\kappa_o}-e^{-\kappa_o}} \cdot e^{\kappa_d\cdot(\cos(\theta-\theta_0)-1)}
%\end{equation}
\begin{equation}
f_{MT}(\theta) = R_0 + R_m \cdot \frac{e^{\kappa_d\cdot \cos(\theta-\theta_0)}-e^{-\kappa_d}}{e^{\kappa_d}-e^{-\kappa_d}}
\end{equation}
Figure~\ref{fig:tunings}.B shows the estimated tunings of $f$ for 16 different cells.
%%---------------------------------------------------%
%: fig:tuning
\begin{figure}
\includegraphics[width=\textwidth]{fig6.pdf}
\caption{{\bf Mean and dispersion tunings} (Left) The box plots represent the distribution of cells' mean spike count (A) and the distribution of their inverse-dispersion values (C) function of the different directions centered around the preferred direction of each cell. The ``whiskers'' for each direction indicate the 25th and the 75th percentiles. The colored dashes (blue for mean and red for inverse dispersion) represent the median values. The inverse-dispersion parameter (C) shows a tuning less regular than the mean (A) but that could, as a first approximation, be fitted by a bell-shaped function. (Right) Tuning curves of the mean spike count (B, in blue) and the inverse-dispersion (D, in red) for 16 cells resulting from, respectively, the fit of a von Mises function (equivalent to a circular Gaussian) to each cell estimated parameters from their repeated responses to the 16 directions used for training (gray points). Each subplot correspond\xadded{s} to a cell. The tuning functions of the inverse-dispersion are supposed to be centered on the cell's preferred direction deduced from the mean fit.}
\label{fig:tunings}
\end{figure}
%%---------------------------------------------------%
Then, we should consider the estimation of inverse-dispersion parameter $\phi$. We tested two hypotheses. First, having no prior on the tuning of $\phi$ as a function of the stimuli direction, we supposed that this parameter is stimulus independent (constant chosen as the average value of its estimates over the different conditions) and varying between cells. Second, we tested the hypothesis of a stimulus dependent inverse-dispersion parameter. We considered a von Mises tuning of $\phi$ centered on the preferred direction for each cell. Figure~\ref{fig:tunings}.D shows the estimated tunings of $\phi$ for 16 different cells. \xadded{Please note that our aim is not to have an optimal fit but to show that a simple tuning hypothesis of $\phi$ can improve the quality of decoding, even if the estimation of $\phi$ values is difficult (see discussion)}.
Finally, to perform population decoding, we used a leave-one-out (LOO) cross validation method. In this method, we randomly pick a single trial from the data set for each neuron and each condition. Then, we apply the decoding approach previously described: we infer an estimate of the encoded direction (stimulus direction) that maximizes the log-likelihood functions based on the Poisson model (eq.~\ref{eq:lp}) or the Negative-Binomial model (eq.~\ref{eq:lnb}) using the spike count of each cell given by the picked trial (as $k_i$) and the estimated values of mean and inverse-dispersion using the remaining trials(as $\phi_i$ and $f_i$). Then, we compare the estimated direction angle to the stimulus direction for the PM model and the NBM model under tuned and un-tuned dispersion hypotheses.
%%---------------------------------------------------%
%: fig:decoding
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fig7.pdf}
\end{center}
\caption{{\bf Population decoding : Poisson model versus Negative-Binomial model}. Variation of decoding error values using the Negative-Binomial model (NBM) function compared to that using the Poisson model (PM) both applied to real MT data (A and C), then to surrogate data generated using the parameters inferred from the real data (B and D). We tested two alternatives of the NBM : one with an \xchanged{overdispersion}{over-dispersion} parameter stimulus-independent (non-tuned dispersion) and the other fitting this parameter by a von Mises function of the direction centered on the preferred direction (tuned dispersion). Points correspond to the median error values and bars to 25\%-75\% confidence interval within 100 trials (horizontal for PM and vertical for NBM). The Poisson model and The Negative-\xadded{Binomial }model with non-tuned dispersion show similar results. However, the Negative-Binomial model with stimulus-dependent dispersion parameters shows better results compared to the two others. This \xremoved{shows}{suggests} that tuned \xchanged{overdispersion}{over-dispersion} could play a role in the neural code.}%
\label{fig:decoding}
\end{figure}
%%---------------------------------------------------%
The resulting decoding error values for both models applied to MT data show that the precision of decoding of the PM and the NBM under the non-tuned dispersion hypothesis are very similar (Figure~\ref{fig:decoding}.A). However, we found that there is a significant gain in decoding when using the NBM under the tuned dispersion hypothesis (Figure~\ref{fig:decoding}.C). To test if this is related to some data or estimation bias, we performed the same test on surrogate data. We generated Negative-Binomial data using the estimated parameter and applied the two decoding schemes. Similarly, the results of decoding on surrogate data show that the precision of decoding is not improved when using non-tuned dispersion (Figure~\ref{fig:decoding}.B). However, when considering additive knowledge about \xchanged{overdispersion}{over-dispersion} tuning, there is a clear gain in accuracy and precision (mean and variance Figure~\ref{fig:decoding}.D) which seems to follow decoding performance in real data.
\xadded{Moreover, comparing the goodness of fit of these models to MT data favors strongly the tuned NBM. Indeed, we used the Bayesian information criterion (BIC) that penalizes models with additional parameters \protect \citep{Box94} to determine which model best explained our data. With the BIC measure, 37 (respectively 28) over the 40 cells were best fit with the tuned NBM (respectively non-tuned NBM) and the remaining were best fit with the Poisson model.}
In summary, we have shown two main results in this section. First, the Poisson model performs as well as the Negative-Binomial model in the decoding task most of the time but not for good reasons as the data set is not well described by a Poisson variability model. Second, the Negative-Binomial model with tuned \xchanged{overdispersion}{over-dispersion} performs better and could be a generic model of neural computation.
% ------------------------------------------------- %
% ------------------------------------------------- %
% -------------------------------------------------
%
\section{Discussion}
%: summary%
% ------------------------------------------------- %% ------------------------------------------------- %% %------------------------------------------------- %
In this paper, we explored \xchanged{overdispersion}{over-dispersion} in neural spike counts using a doubly stochastic model, namely the Negative-Binomial model (NBM), that could disentangle different sources of \xchanged{overdispersion}{over-dispersion}. We have first explained that some commonly used procedures to quantify \xchanged{overdispersion}{over-dispersion} (the Fano factor and chi-square goodness of fit tests) could be problematic. Then, we have shown that the alternative analytical procedure that we propose (the Fano Gamma test) better characterizes \xchanged{overdispersion}{over-dispersion} in three different data sets (different animals, tasks, conditions and areas). Using the Fano Gamma test we show that there is weak evidence of \xchanged{overdispersion}{over-dispersion} in LGN and V1 data sets and low evidence against the NBM in MT data set, \xadded{especially for high mean spiking values,} a trend expected from~\citep{Goris14}.
In the last section, we compared the performances of the traditional Poisson model to the more recent Negative-Binomial model on a population decoding algorithm (direction decoding from spike recording in MT cells of an anesthetized macaque monkey). Under the Poisson hypothesis, a linear decoding model is optimal\xadded{,} as proposed by~\citet{Jazayeri06}. We show here that this decoder is still efficient (mean decoding error $<$ 10 degrees) even though there is strong evidence against the Poisson hypothesis. This is partly due to the tuning profile of the likelihood functions: the direction corresponding to the maximum likelihood will converge toward the stimulus' direction even if the noise model is not optimal.
%In the last section, we compared the performances of the traditional Poisson model to the more recent Negative-Binomial model on a population decoding task (direction decoding from spike recording in MT cells of an anaesthetized macaque monkey). Under the Poisson hypothesis, a linear decoding model is optimal as proposed by~\citep{Jazayeri06}. In that paper, they proposed a feed forward model implementing the computation of a decoder. This model starts from a sensory representation, use a pooling of spike counts weighted by the mean firing rates that should be merged as being the individual likelihood values. Surprisingly, we show here that this decoder is still efficient (mean decoding error $<10$ degrees). This is partly due to the tuning profile of the likelihood function, the direction corresponding to the maximum likelihood converges toward the stimulus' direction even if the noise model is not optimal.
Intuitively, one could expect that using the compound model that accounts for the observed \xchanged{overdispersion}{over-dispersion}, one should obtain better decoding results. We first considered the hypothesis of non-tuned cells' dispersion (i.e. each cell dispersion parameter is constant). The Poisson model and The Negative-Binomial model with non-tuned dispersion show similar direction decoding results, not only on real data, but also on surrogate data. However, note that, similarly to what was reported before~\citep{Goris14}, we found that the likelihood values of the estimated directions are always higher for the Negative-Binomial based decoding compared to the Poisson based decoding (not shown), but with similar bandwidths. This leads to similar decoding accuracies, despite the fact that the NBM fits better to the data. Indeed, the spike counts corresponding to the response of the cell to its preferred direction are more likely to be generated by an NBM with higher inverse-dispersion values, i.e. Poisson-like. This could explain why PM performance could be comparable to NBM performance in decoding.
%NB non tuned is as good as Poisson
%Intuitively, one could expect that using the compound model that accounts for the over-dispersion widely observed (i.e. the Poisson process is not adapted), one should observe better decoding results. We first considered the hypothesis of non-tuned cells' dispersion (i.e. each cell dispersion parameter is constant). Indeed, that would underlie a generic neural computation which a relative invariance with stimulus directions of the information represented by the neural populations. However, the Poisson model and The Negative-Binomial model with non-tuned dispersion show similar direction decoding results not only on real data, but also on surrogate data. Similarly to what was reported before~\citep{Goris14}, we found that the likelihood values of the estimated directions are always higher for the Negative-Binomial based decoding compared to the Poisson based decoding (not shown in this paper), but with similar bandwidths. This leads to similar decoding accuracies, despite the fact that the NBM fits better to the data. Indeed, the spike counts corresponding to the response of the cell to its preferred direction are more likely to be generated by an NBM with higher inverse-dispersion values, so more Poisson-like distributions, which could explain why PM performance could be comparable to NBM performance in decoding.
% NB with tuned is better
In addition, the estimation of the inverse-dispersion parameter is known to be difficult. Indeed, the estimation of this parameter has been widely studied in different fields~\citep{Gourieroux84, Lawless87, Mccullagh89, Dai13} and several estimation methods were provided such as \xadded{the} maximum likelihood method~\citep{Piegorsch90},\xadded{ or the} method of moment and maximum extended quasi-likelihood method~\citep{Clark89}. A wide range of studies have shown that the estimation of this parameter is very challenging and can be significantly biased, especially when dealing with small data sets. It is worthwhile to point out this problem here: Since neurophysiological experiments are costly and time consuming, we often deal with small data sets. Thus, having more knowledge and control about the dispersion tuning is crucial.
%%%%
Therefore, we tested a simple possible form of dispersion tuning, that is, the tuning of the inverse-dispersion having similar preferred direction as the tuning of the spike count (using a von Mises function). We found that under the hypothesis of a stimulus-dependent inverse-dispersion, the NBM performs qualitatively better.
%------------------------------------------------- %
Moreover, such a tuning function could explain complex Fano factor tuning profiles that were recently reported~\citep{Ponce13}. Our results lead us to propose that the reported trial-by-trial variability among MT neurons - showing a directional tuning that is not trivially explained by firing rate variations alone - may simply be explained by a bell-shaped tuning of the \xchanged{overdispersion}{over-dispersion} (see Figure~\ref{fig:tunings2}). \xadded{However, contrary to what they expect, we did not observe a decrease of over-dispersion with respect to the spontaneous level. Indeed, applying the Fano-Gamma test to MT spontaneous spiking data (results not shown) yields lower evidence against a PM (47.5$\%$) compared to that for evoked spiking (64.1$\%$, Figure 5.D) but similar evidence against NBM ($15\%$) compared to evoked spiking (16.3$\%$, Figure 5.D). This is more in line with an increase of under-dispersion, but which still in accordance with the known fact that FF is decreased when a stimulus is applied.}
%%---------------------------------------------------%
\begin{figure}
\begin{center}
\includegraphics[width=0.9\textwidth]{fig8.pdf}\\
\end{center}
\caption{{\bf Complex tuning of Fano factor (FF) values for different motion directions.} \xchanged{The plot on the right shows the expected FF profile resulting from the mean spike count ($f$) profile (on the right) and the inverse-dispersion ($\phi$) profile (in the middle) using the relationship equation ($FF=1+\frac{f}{\phi}$) provided by the Negative-Binomial model.}{(A) An example of a mean response tuning curve. (B) An example of an inverse-dispersion tuning curve. (A) and (B) were analytically generated by Von Mises (circular Gaussian) functions centered on a given preferred direction (PD). (C) Let us suppose that the spiking count of a cell is driven from a Negative-Binomial model characterized by mean values given in (A) and inverse dispersion values given in (B). The expected resulting Fano factor values are given by the relationship equation ($FF=1+\frac{f}{\phi}$). For example, the mean value (7.14) and the inverse dispersion value (21.47) corresponding to the PD result in an FF value of 1.33. The resulting FF profile could not be explained only by the mean response tuning as in a Poisson model.}}
\label{fig:tunings2}
\end{figure}
%%---------------------------------------------------%
%
Where could such a tuning of the \xchanged{overdispersion}{over-dispersion} come from? One first hypothesis could be that \xchanged{overdispersion}{over-dispersion} results from the sensory noise. Then, as suggested by~\citet{Goris14}, \xchanged{overdispersion}{over-dispersion} could also reflect an excitability state of the cell. However, both of these hypotheses cannot account for \xremoved{a}stimulus dependent variability. Alternatively, we \xremoved{can}suppose that this property is better apprehended at the population scale. Going back to the encoding paradigm described in the first section (Figure~\ref{fig:model22}), an alternative hypothesis would be that the \xchanged{overdispersion}{over-dispersion} arise from a stimulus-dependent degree of network convergence (within MT or from V1). Such \xadded{a} scheme could result from a center-surround connectivity profile in the direction domain: more excitatory input for a stimulus close to the preferred direction of the recipient cell.
%%%source of over-dispersion
%Where could such a tuning of the over-dispersion come from? From the encoding paradigms described in the first section, we can suppose that this property is better apprehended at the population scale. First, considering the first hypothesis of an over-dispersion resulting from the sensory noise do not explain a stimulus dependent variability. Second, in the recent work of~\citet{Goris14}, they suggested that over-dispersion could reflect an excitability state of the cell. However, a tuned dispersion hypothesis implies that this over-dispersion should not be considered as a stationary property of the cell. Thus, the most reliable hypothesis is that the over-dispersion could be explained by a stimulus dependent network convergence (within MT from V1), i.e, a larger degree of neuronal convergence for a stimulus close to the preferred direction. In fact, a center-surround connectivity profile combined in the direction domain (with a topographic mapping of cells based on their selectivities) may result in a bell-shaped tuning of the dispersion, i.e, more excitatory inputs to the cell encoding the preferred direction.
%As a consequence, this study calls for an extension over a wider range of experimental observations. First, we would like to extend this work to more datasets with more trials to explore if our predictions are still valid, both for teasing out sources of variability but also to evaluate population decoding. It would be mostly interesting to compare this extra variability along the sensory pathways to associative areas to confirm the tendency that we observed with a gradual increase of over-dispersion from the lower to associative areas. A second important perspective is to consider the role of experimental conditions. Indeed, it is likely that results would greatly vary depending on the arousal sate of the animal, the nature of the stimulation or the behavioral task. In particular, behaviorally more relevant stimulations (such as the model-based synthesis of textures, Sanz- Leon et al., 2012) may lead to a more precise response (Baudot et al., 2013), that would greatly benefit from the tools that we developed herein.
%: perspectives
This study calls for an extension over a wider range of experimental observations. First, we would like to extend this work to more datasets with more trials to explore if our predictions are still valid, both for teasing out sources of variability but also to evaluate population decoding. It would be mostly interesting to compare this extra variability along the sensory pathways to associative areas to confirm the tendency that we observed with a gradual increase of \xchanged{overdispersion}{over-dispersion} from the lower to associative areas. A second important perspective is to consider the role of experimental conditions. Indeed, it is likely that results would greatly vary depending on the arousal \xchanged{sate}{state} of the animal, the nature of the stimulation or the behavioral task. In particular, behaviorally more relevant stimulations (such as the model-based synthesis of textures~\citep{Leon12} may lead to a more precise response~\citep{Baudot13}.%and that the tools that we developed herein could be a marker for elaborating such experiments.
%A wide range of the studies have shown that the estimation of this parameter is very challenging and can be significally biased specially when dealing with small data sets.%\citep{Willson86, Clark89, Piegorsch90, Dean94, Lord06}.
%We predict also from the FG test results on the three provided data sets that the over-dispersion is more likely to increase from low to high level areas that could be explained by more noise sources and dimensionality reduction.
%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Materials and Methods}
In our analyses, we used three different data sets from neurophysiological recordings.
\paragraph{LGN data set}%%%%%%%%%%%%%%%%%%%%%%%%%%
This data set~\citep{Scholl13data} contains the spiking responses of 72 LGN neurons in the anesthetized mouse to drifting grating (square wave (n=24) and sinusoidal (n=48)). Many of these neurons are already contained in the data set published in~\citep{Scholl13}. Contrarily to the two other data sets, \xremoved{this one is not from us, }it is an openly available data set.
\paragraph{V1 data set}%%%%%%%%%%%%%%%%%%%%%%%%%%
Experiments were conducted at the INT on two adult male rhesus macaque monkeys (Macaca mulatta). Experimental protocols have been approved by the Marseille Ethical Committee in Neuroscience (approval $\#A10/01/13$, official national registration $\#71-$French Ministry of Research). All procedures complied with the French and European regulations for animal research, as well as the guidelines from the Society for Neuroscience. \\
\textit{Surgical preparation : }
Monkeys were chronically implanted with a head-holder and a recording chamber located above the V1 and V2 cortical areas. In a second surgery, a search coil was inserted below the ocular sclera to record eye movements~\citep{Robinson63}.\\
%%%%%%%%%%%%%%%%%%%%%%%%%%
\textit{Behavioral task and training : }
Monkeys were trained to fixate - within a window of 1 to 2 degrees of diameter - on a red target presented at the center of the screen during \xremoved{all}the \xadded{entire} duration of the trial (1.5 $s$). The eye position was monitored by the scleral search coil technique~\citep{Collewijn75,Robinson63} and the animal's behavior controlled using the REX package~\citep{Hays82}. \\
%%%%%%%%%%%%%%%%%%%%%%%%%%
\textit{Unit Recording :}
The two monkeys were chronically implanted with a head-holder and a recording chamber located above \xremoved{the}cortical area V1. A computer controlled microdrive (MT-EPS, Alpha Omega, Israel) was used to trans-durally insert a micro-electrode (FHC, 0.5-1.2 M$\Omega$ at 1 KHz) in the right hemisphere. Spikes were sorted online using a template matching algorithm (MSD, Alpha Omega).\\
%%%%%%%%%%%%%%%%%%%%%%%%%%
\textit{Visual Stimulation : }
Visual stimulation protocols have been produced using in-house software (developed by G\'erard Sadoc, Acquis1-Elphy, Biologic CNRS-UNIC/ANVAR). Stimuli were back-projected on a translucent screen covering $80^\circ$$\times$$80^\circ$ of the monkey's visual field, at a distance of 1 $m$, using a retro-projector (resolution: 1280$\times$1024 pixels at 60 $Hz$). The mean luminance of the motion stimulus was 22.2 $cd/m^2$ and the background was kept constant to about 2.24 $cd/m^2$. The display was gamma calibrated by means of a lookup table.
Once a cell is isolated, a sparse noise (SN) stimulus was first used to quantitatively map out the RF (a grid of 10x10 squares with a side length of 0.6 degrees, 10.9 $cd/m^2$ darker or brighter than the background of 11.1 $cd/m^2$, $\sim$15 trials). Once the RF was properly located and estimated, a direction tuning paradigm was launched. In this paradigm, a moving bar (0.5$\times$4$^\circ$) translated across the receptive field (the displacement covers $3^\circ$ on either side of the RFc) in 12 possible directions (spaced by 30 degrees of polar angle, at a speed of $6.6^\circ/s$).
\paragraph{MT data set }%%%%%%%%%%%%%%%%%%%%%%%%%%
Experiments were conducted at NYU-CNS on 3 anesthetized, paralyzed, adult macaque monkeys (Macaca nemestrina) of either sex. All procedures were conducted in compliance with the National Institute of Health Guide for the Care and Use of Laboratory Animals, and with the approval of the New York University Animal Welfare Committee.\\
%%%%%%%%%%%%%%%%%%%%%%%%%%
\textit{Surgical preparation : }The standard procedures for the surgical preparation of animals and single-unit recordings have been reported in detail previously~\citep{Cavanaugh02}. Briefly, experiments typically lasted 5 to 6 days, during which we maintained anesthesia with infusion of sufentanil citrate (6-30 $\mu g$ $kg^{-1}$ $h^{-1}$) and paralysis with infusion of vecuronium bromide (Norcuron; 0.1 $mg$ $kg^{-1}$ $h^{-1}$) in isotonic dextrose-normosol solution. We monitored vital signs (heart rate, lung pressure, end-tidal pCO2, EEG, body temperature, urine flow, and osmolarity) and maintained them within appropriate physiological ranges. Pupils were dilated with topical atropine. The eyes were protected with gas-permeable contact lenses, and refracted with supplementary lenses chosen through direct ophthalmoscopy. At the conclusion of data collection, the animal was sacrificed with an overdose of sodium pentobarbital.\\
%%%%%%%%%%%%%%%%%%%%%%%%%%
\textit{Unit Recording : }Extracellular recordings were made with quartz-platinum-tungsten microelectrodes (Thomas Recording), advanced mechanically into the brain through a craniotomy and small durotomy. To record from MT we passed microelectrodes through a small durotomy centered roughly 16 $mm$ lateral to the midline and 3 $mm$ posterior to the lip of the lunate sulcus at an angle of 20 degrees from horizontal in a ventro-anterior direction. Area MT was identified \xchanged{from}{by} the brisk direction-selective responses of isolated neurons. We made recordings from every single unit with a spike waveform that rose sufficiently above noise to be isolated. Stimuli were presented in random order.\\
%%%%%%%%%%%%%%%%%%%%%%%%%%
\textit{Visual Stimulation : }We presented visual stimuli on a gamma-corrected CRT monitor (Eizo T966; mean luminance: 33 $cd/m^2$) at a resolution of 1.280$\times$960 with a refresh rate of 120 $Hz$. Stimuli were presented using Expo software (\url{ http://corevision.cns.nyu.edu}) on an Apple Macintosh computer. For each isolated unit, we first determined its ocular dominance and occluded the non-preferred eye. We presented circularly windowed sinusoidal grating stimuli (16 directions around the clock, 22.5 degrees apart) to map each cell's receptive field, determined its preferred size and speed, and then measured selectivity for orientation or spatial frequency.
The time window considered to calculate the spike counts was the duration of stimulation for the grating stimuli (LGN and MT data sets) and 400 $ms$ covering approximatively the time when the bar stimulus \xchanged{crosses}{crossed} the cell's receptive field for V1 data set. All data were analyzed using Python and Matlab (The Mat\xadded{h}Works Inc., Natick, MA, USA).
%%--------------------------------------------------------------------------------------------------%
\section*{Acknowledgments}
\Acknowledgments
%--------------------------------------------------------------------------------------------------%
%--%\printbibliography
%\bibliographystyle{jneurophysiol}
%\bibliography{Taouali_etal2015}
\printbibliography
%-------------------------------------------------
%\section*{Figures}
\end{document}
% $ biblatex auxiliary file $
% $ biblatex version 2.5 $
% Do not modify the above lines!
%
% This is an auxiliary file used by the 'biblatex' package.
% This file may safely be deleted. It will be recreated as
% required.
%
\begingroup
\makeatletter
\@ifundefined{ver@biblatex.sty}
{\@latex@error
{Missing 'biblatex' package}
{The bibliography requires the 'biblatex' package.}
\aftergroup\endinput}
{}
\endgroup
\entry{Fano47}{article}{}
\name{author}{1}{}{%
{{}%
{Fano}{F.}%
{U.}{U.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{APS}%
}
\strng{namehash}{FU1}
\strng{fullhash}{FU1}
\field{number}{1}
\field{pages}{26}
\field{title}{Ionization yield of radiations. II. The fluctuations of the
number of ions}
\field{volume}{72}
\field{journaltitle}{Physical Review}
\field{year}{1947}
\endentry
\entry{Kara00}{article}{}
\name{author}{3}{}{%
{{}%
{Kara}{K.}%
{P.}{P.}%
{}{}%
{}{}}%
{{}%
{Reinagel}{R.}%
{P.}{P.}%
{}{}%
{}{}}%
{{}%
{Reid}{R.}%
{R.~C.}{R.~C.}%
{}{}%
{}{}}%
}
\keyw{cat, lgn, neurophysiology, response\_variability, retina, spike
statistics, taouali15, v1, variability}
\strng{namehash}{KPRPRRC1}
\strng{fullhash}{KPRPRRC1}
\field{abstract}{%
The response of a cortical cell to a repeated stimulus can be highly variable
from one trial to the next. Much lower variability has been reported of
retinal cells. We recorded visual responses simultaneously from three
successive stages of the cat visual system: retinal ganglion cells ({RGCs}),
thalamic ({LGN}) relay cells, and simple cells in layer 4 of primary visual
cortex. Spike count variability was lower than that of a Poisson process at
all three stages but increased at each stage. Absolute and relative
refractory periods largely accounted for the reliability at all three stages.
Our results show that cortical responses can be more reliable than previously
thought. The differences in reliability in retina, {LGN}, and cortex can be
explained by (1) decreasing firing rates and (2) decreasing absolute and
relative refractory periods.%
}
\field{number}{3}
\field{pages}{635\bibrangedash 646}
\field{title}{Low response variability in simultaneously recorded retinal,
thalamic, and cortical neurons.}
\verb{url}
\verb http://view.ncbi.nlm.nih.gov/pubmed/11055444
\endverb
\field{volume}{27}
\list{location}{1}{%
{Department of Neurobiology, Harvard Medical School, Boston, Massachusetts
02115, USA.}%
}
\field{journaltitle}{Neuron}
\field{month}{09}
\field{year}{2000}
\endentry
\entry{Alitto11}{article}{}
\name{author}{4}{}{%
{{}%
{Alitto}{A.}%
{Henry~J.}{H.~J.}%
{}{}%
{}{}}%
{{}%
{Moore}{M.}%
{Bartlett~D.}{B.~D.}%
{}{}%
{}{}}%
{{}%
{Rathbun}{R.}%
{Daniel~L.}{D.~L.}%
{}{}%
{}{}}%
{{}%
{Usrey}{U.}%
{Martin~M.}{M.~M.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Wiley-Blackwell}%
}
\keyw{taouali15}
\strng{namehash}{AHJMBDRDLUMM1}
\strng{fullhash}{AHJMBDRDLUMM1}
\field{abstract}{%
Despite the increasing use of alert animals for studies aimed at
understanding visual processing in the cerebral cortex, relatively little
attention has been focused on quantifying the response properties of neurons
that provide input to the cortex. Here, we examine the response properties of
neurons in the lateral geniculate nucleus ({LGN}) of the thalamus in the
alert macaque monkey and compare these responses to those in the
anaesthetized animal. Compared to the anaesthetized animal, we show that
magnocellular and parvocellular neurons in the alert animal respond to visual
stimuli with significantly higher firing rates. This increase in
responsiveness is not accompanied by a change in the shape of neuronal
contrast response functions or the strength of centre--surround antagonism;
however, it is accompanied by an increased ability of neurons to follow
stimuli drifting at higher spatial and temporal frequencies.%
}
\verb{doi}
\verb 10.1113/jphysiol.2010.190538
\endverb
\field{number}{Pt 1}
\field{pages}{87\bibrangedash 99}
\field{title}{A comparison of visual responses in the lateral geniculate
nucleus of alert and anaesthetized macaque monkeys.}
\verb{url}
\verb http://dx.doi.org/10.1113/jphysiol.2010.190538
\endverb
\field{volume}{589}
\field{journaltitle}{The Journal of physiology}
\field{month}{01}
\field{year}{2011}
\endentry
\entry{Zhuang13}{article}{}
\name{author}{7}{}{%
{{}%
{Zhuang}{Z.}%
{Jun}{J.}%
{}{}%
{}{}}%
{{}%
{Stoelzel}{S.}%
{Carl~R.}{C.~R.}%
{}{}%
{}{}}%
{{}%
{Bereshpolova}{B.}%
{Yulia}{Y.}%
{}{}%
{}{}}%
{{}%
{Huff}{H.}%
{Joseph~M.}{J.~M.}%
{}{}%
{}{}}%
{{}%
{Hei}{H.}%
{Xiaojuan}{X.}%
{}{}%
{}{}}%
{{}%
{Alonso}{A.}%
{Jose-Manuel~M.}{J.-M.~M.}%
{}{}%
{}{}}%
{{}%
{Swadlow}{S.}%
{Harvey~A.}{H.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Society for Neuroscience}%
}
\keyw{taouali15}
\strng{namehash}{ZJ+1}
\strng{fullhash}{ZJSCRBYHJMHXAJMMSHA1}
\field{abstract}{%
Extracellular recordings were obtained from two cell classes in layer 4 of
the awake rabbit primary visual cortex (V1): putative inhibitory interneurons
[suspected inhibitory interneurons ({SINs})] and putative excitatory cells
with simple receptive fields. {SINs} were identified solely by their
characteristic response to electrical stimulation of the lateral geniculate
nucleus ({LGN}, 3+ spikes at >600 Hz), and simple cells were identified
solely by receptive field structure, requiring spatially separate {ON} and/or
{OFF} subfields. Notably, no cells met both criteria, and we studied 62
simple cells and 33 {SINs}. Fourteen cells met neither criterion. These layer
4 populations were markedly distinct. Thus, {SINs} were far less linear
({F1/F0} < 1), more broadly tuned to stimulus orientation, direction, spatial
and temporal frequency, more sensitive to contrast, had much higher
spontaneous and stimulus-driven activity, and always had spatially
overlapping {ON}/{OFF} receptive subfields. {SINs} responded to drifting
gratings with increased firing rates (F0) for all orientations and
directions. However, some {SINs} showed a weaker modulated (F1) response
sharply tuned to orientation and/or direction. {SINs} responded at shorter
latencies than simple cells to stationary stimuli, and the responses of both
populations could be sustained or transient. Transient simple cells were more
sensitive to contrast than sustained simple cells and their visual responses
were more frequently suppressed by high contrasts. Finally, cross-correlation
between {LGN} and {SIN} spike trains confirmed a fast and precisely timed
monosynaptic connectivity, supporting the notion that {SINs} are well suited
to provide a fast feedforward inhibition onto targeted cortical populations.%
}
\verb{doi}
\verb 10.1523/jneurosci.0863-13.2013
\endverb
\field{number}{28}
\field{pages}{11372\bibrangedash 11389}
\field{title}{Layer 4 in primary visual cortex of the awake rabbit:
contrasting properties of simple cells and putative feedforward inhibitory
interneurons.}
\verb{url}
\verb http://dx.doi.org/10.1523/jneurosci.0863-13.2013
\endverb
\field{volume}{33}
\field{journaltitle}{The Journal of Neuroscience}
\field{month}{07}
\field{year}{2013}
\endentry
\entry{Tolhurst83}{article}{}
\name{author}{3}{}{%
{{}%
{Tolhurst}{T.}%
{D.~J.}{D.~J.}%
{}{}%
{}{}}%
{{}%
{Movshon}{M.}%
{J.~Anthony}{J.~A.}%
{}{}%
{}{}}%
{{}%
{Dean}{D.}%
{A.~F.}{A.~F.}%
{}{}%
{}{}}%
}
\keyw{cat, monkey, noise, spike statistics, taouali15, v1, variability}
\strng{namehash}{TDJMJADAF1}
\strng{fullhash}{TDJMJADAF1}
\field{abstract}{%
The variability of the discharge of visual cortical neurons in cats and
macaque monkeys limits the reliability with which such neurons can relay
signals about weak visual stimuli. In general, the variance of a neuron's
firing rate is directly proportional to its mean firing rate. The probability
that a neuron will fire a criterion number of impulses on a stimulus trial
grows monotonically with the contrast of a sinusoidal grating stimulus.
Neural probability functions prepared either by computing the probability of
criterion response or by integrating receiver operating characteristics to
yield the probability of correct choice in a two-alternative forced-choice
situation resemble psychometric functions obtained in psychophysical and
behavioral experiments on humans and animals, but are shallower in slope. The
slopes of neuronal probability functions are slightly higher when they are
estimated over short time periods, but even so do not equal the slopes
measured psychophysically in human and monkey observers. This discrepancy in
slope could be explained if the whole observer responded only when about four
neurons were active together.%
}
\field{number}{8}
\field{pages}{775\bibrangedash 785}
\field{title}{The statistical reliability of signals in single neurons in cat
and monkey visual cortex.}
\verb{url}
\verb http://view.ncbi.nlm.nih.gov/pubmed/6623937
\endverb
\field{volume}{23}
\field{journaltitle}{Vision {R}esearch}
\field{year}{1983}
\endentry
\entry{Softky93}{article}{}
\name{author}{2}{}{%
{{}%
{Softky}{S.}%
{W.~R.}{W.~R.}%
{}{}%
{}{}}%
{{}%
{Koch}{K.}%
{Christof}{C.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Society for Neuroscience}%
}
\keyw{taouali15}
\strng{namehash}{SWRKC1}
\strng{fullhash}{SWRKC1}
\field{abstract}{%
How random is the discharge pattern of cortical neurons? We examined
recordings from primary visual cortex (V1; Knierim and Van Essen, 1992) and
extrastriate cortex ({MT}; Newsome et al., 1989a) of awake, behaving macaque
monkey and compared them to analytical predictions. For nonbursting cells
firing at sustained rates up to 300 Hz, we evaluated two indices of firing
variability: the ratio of the variance to the mean for the number of action
potentials evoked by a constant stimulus, and the rate-normalized coefficient
of variation (Cv) of the interspike interval distribution. Firing in
virtually all V1 and {MT} neurons was nearly consistent with a completely
random process (e.g., Cv approximately 1). We tried to model this high
variability by small, independent, and random {EPSPs} converging onto a leaky
integrate-and- fire neuron (Knight, 1972). Both this and related models
predicted very low firing variability (Cv << 1) for realistic {EPSP}
depolarizations and membrane time constants. We also simulated a
biophysically very detailed compartmental model of an anatomically
reconstructed and physiologically characterized layer V cat pyramidal cell
(Douglas et al., 1991) with passive dendrites and active soma. If
independent, excitatory synaptic input fired the model cell at the high rates
observed in monkey, the Cv and the variability in the number of spikes were
both very low, in agreement with the integrate-and-fire models but in strong
disagreement with the majority of our monkey data. The simulated cell only
produced highly variable firing when {Hodgkin-Huxley}- like currents ({INa}
and very strong {IDR}) were placed on distal dendrites. Now the simulated
neuron acted more as a millisecond- resolution detector of dendritic spike
coincidences than as a temporal integrator. We argue that neurons that act as
temporal integrators over many synaptic inputs must fire very regularly. Only
in the presence of either fast and strong dendritic nonlinearities or strong
synchronization among individual synaptic events will the degree of predicted
variability approach that of real cortical neurons.%
}
\field{number}{1}
\field{pages}{334\bibrangedash 350}
\field{title}{The highly irregular firing of cortical cells is inconsistent
with temporal integration of random {EPSPs}}
\verb{url}
\verb http://www.jneurosci.org/content/13/1/334.abstract
\endverb
\field{volume}{13}
\list{location}{2}{%
{Division of Physics, Mathematics,}%
{Astronomy, California Institute of Technology, Pasadena 91125.}%
}
\field{journaltitle}{The Journal of Neuroscience}
\field{month}{01}
\field{year}{1993}
\endentry
\entry{Shadlen98}{article}{}
\name{author}{2}{}{%
{{}%
{Shadlen}{S.}%
{Michael~N.}{M.~N.}%
{}{}%
{}{}}%
{{}%
{Newsome}{N.}%
{William~T.}{W.~T.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Society for Neuroscience}%
}
\keyw{taouali15}
\strng{namehash}{SMNNWT1}
\strng{fullhash}{SMNNWT1}
\field{abstract}{%
Cortical neurons exhibit tremendous variability in the number and temporal
distribution of spikes in their discharge patterns. Furthermore, this
variability appears to be conserved over large regions of the cerebral
cortex, suggesting that it is neither reduced nor expanded from stage to
stage within a processing pathway. To investigate the principles underlying
such statistical homogeneity, we have analyzed a model of synaptic
integration incorporating a highly simplified integrate and fire mechanism
with decay. We analyzed a '' high-input regime'' in which neurons receive
hundreds of excitatory synaptic inputs during each interspike interval. To
produce a graded response in this regime, the neuron must balance excitation
with inhibition. We find that a simple integrate and fire mechanism with
balanced excitation and inhibition produces a highly variable interspike
interval, consistent with experimental data. Detailed information about the
temporal pattern of synaptic inputs cannot be recovered from the pattern of
output spikes, and we infer that cortical neurons are unlikely to transmit
information in the temporal pattern of spike discharge. Rather, we suggest
that quantities are represented as rate codes in ensembles of 50--100
neurons. These column-like ensembles tolerate large fractions of common
synaptic input and yet covary only weakly in their spike discharge. We find
that an ensemble of 100 neurons provides a reliable estimate of rate in just
one interspike interval (10--50 msec). Finally, we derived an expression for
the variance of the neural spike count that leads to a stable propagation of
signal and noise in networks of neurons---that is, conditions that do not
impose an accumulation or diminution of noise. The solution implies that
single neurons perform simple algebra resembling averaging, and that more
sophisticated computations arise by virtue of the anatomical convergence of
novel combinations of inputs to the cortical column from external sources.%
}
\field{number}{10}
\field{pages}{3870\bibrangedash 3896}
\field{title}{The Variable Discharge of Cortical Neurons: Implications for
Connectivity, Computation, and Information Coding}
\verb{url}
\verb http://www.jneurosci.org/content/18/10/3870.abstract
\endverb
\field{volume}{18}
\list{location}{3}{%
{Department of Physiology}%
{Biophysics}%
{Regional Primate Research Center, University of Washington, Seattle,
Washington 98195-7290, USA.}%
}
\field{journaltitle}{The Journal of Neuroscience}
\field{month}{05}
\field{year}{1998}
\endentry
\entry{Koch99}{article}{}
\name{author}{2}{}{%
{{}%
{Koch}{K.}%
{Christof}{C.}%
{}{}%
{}{}}%
{{}%
{Poggio}{P.}%
{Tomaso}{T.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{[London: Macmillan Journals], 1869-}%
}
\keyw{bicv-motion, khoei14thesis, motion\_anticipation, taouali15}
\strng{namehash}{KCPT1}
\strng{fullhash}{KCPT1}
\verb{doi}
\verb 10.1038/4511
\endverb
\field{number}{1}
\field{pages}{9\bibrangedash 10}
\field{title}{Predicting the visual world: silence is golden}
\verb{url}
\verb http://dx.doi.org/10.1038/4511
\endverb
\field{volume}{2}
\field{journaltitle}{Nature {N}euroscience}
\field{month}{01}
\field{year}{1999}
\endentry
\entry{Britten93}{article}{}
\name{author}{4}{}{%
{{}%
{Britten}{B.}%
{K~H.}{K.~H.}%
{}{}%
{}{}}%
{{}%
{Shadlen}{S.}%
{M~N.}{M.~N.}%
{}{}%
{}{}}%
{{}%
{Newsome}{N.}%
{W~T.}{W.~T.}%
{}{}%
{}{}}%
{{}%
{Movshon}{M.}%
{J~A.}{J.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Cambridge Univ Press}%
}
\strng{namehash}{BKHSMNNWTMJA1}
\strng{fullhash}{BKHSMNNWTMJA1}
\field{number}{06}
\field{pages}{1157\bibrangedash 1169}
\field{title}{Responses of neurons in macaque MT to stochastic motion
signals}
\field{volume}{10}
\field{journaltitle}{Visual {N}euroscience}
\field{year}{1993}
\endentry
\entry{Buracas98}{article}{}
\name{author}{4}{}{%
{{}%
{Buracฬas}{B.}%
{Giedrius~T.}{G.~T.}%
{}{}%
{}{}}%
{{}%
{Zador}{Z.}%
{Anthony~M.}{A.~M.}%
{}{}%
{}{}}%
{{}%
{DeWeese}{D.}%
{Michael~R.}{M.~R.}%
{}{}%
{}{}}%
{{}%
{Albright}{A.}%
{Thomas~D.}{T.~D.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{BGTZAMDMRATD1}
\strng{fullhash}{BGTZAMDMRATD1}
\field{abstract}{%
Although motion-sensitive neurons in macaque middle temporal ({MT}) area are
conventionally characterized using stimuli whose velocity remains constant
for 1--3 s, many ecologically relevant stimuli change on a shorter time scale
(30--300 ms). We compared neuronal responses to conventional
(constant-velocity) and time-varying stimuli in alert primates. The responses
to both stimulus ensembles were well described as rate-modulated Poisson
processes but with very high precision (3 ms) modulation functions underlying
the time-varying responses. Information-theoretic analysis revealed that the
responses encoded only 1 bit/s about constant-velocity stimuli but up to 29
bits/s about the time-varying stimuli. Analysis of local field potentials
revealed that part of the residual response variability arose from '' noise''
sources extrinsic to the neuron. Our results demonstrate that extrastriate
neurons in alert primates can encode the fine temporal structure of visual
stimuli.%
}
\verb{doi}
\verb 10.1016/s0896-6273(00)80477-8
\endverb
\field{number}{5}
\field{pages}{959\bibrangedash 969}
\field{title}{Efficient Discrimination of Temporal Patterns by
{Motion-Sensitive} Neurons in Primate Visual Cortex}
\verb{url}
\verb http://dx.doi.org/10.1016/s0896-6273(00)80477-8
\endverb
\field{volume}{20}
\field{journaltitle}{Neuron}
\field{month}{05}
\field{year}{1998}
\endentry
\entry{McAdams99}{article}{}
\name{author}{2}{}{%
{{}%
{McAdams}{M.}%
{Carrie~J.}{C.~J.}%
{}{}%
{}{}}%
{{}%
{Maunsell}{M.}%
{John H.~R.}{J.~H.~R.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{MCJMJHR1}
\strng{fullhash}{MCJMJHR1}
\field{abstract}{%
To determine the physiological mechanisms underlying the enhancement of
performance by attention, we examined how attention affects the ability of
isolated neurons to discriminate orientation by investigating the reliability
of responses with and without attention. Recording from 262 neurons in
cortical area V4 while two rhesus macaques did a delayed match-to-sample task
with oriented stimuli, we found that attention did not produce detectable
changes in the variability of neuronal responses but did improve the
orientation discriminability of the neurons. We also found that attention did
not change the relationship between burst rate and response rate. Our results
are consistent with the idea that attention selects groups of neurons for a
multiplicative enhancement in response strength.%
}
\verb{doi}
\verb 10.1016/s0896-6273(01)80034-9
\endverb
\field{number}{4}
\field{pages}{765\bibrangedash 773}
\field{title}{Effects of Attention on the Reliability of Individual Neurons
in Monkey Visual Cortex}
\verb{url}
\verb http://dx.doi.org/10.1016/s0896-6273(01)80034-9
\endverb
\field{volume}{23}
\field{journaltitle}{Neuron}
\field{month}{08}
\field{year}{1999}
\endentry
\entry{Heggelund78}{article}{}
\name{author}{2}{}{%
{{}%
{Heggelund}{H.}%
{P.}{P.}%
{}{}%
{}{}}%
{{}%
{Albus}{A.}%
{K.}{K.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Springer-Verlag}%
}
\keyw{taouali15}
\strng{namehash}{HPAK1}
\strng{fullhash}{HPAK1}
\field{abstract}{%
The response of single cells in the striate cortex of cat to a moving light
bar of variable orientation was measured by a method providing data on the
mean response as well as the standard deviation ({SD}) at the different
stimulus orientations. At the optimal stimulus orientation the {SD} was about
1/3 of the mean response. Marked differences in this respect were found
between simple and complex cells, i.e., the {SD} for the simple cells was
about 1/2 of the mean response and about 1/4 for the complex cells. The
variation coefficient (Vc = {SD}/mean) was minimal at the optimal orientation
and increased relatively in the same manner for simple and complex cells as
the stimulus orientation was varied away from optimal orientation. The Vc
varied with the mean response at optimal orientation in a nonlinear manner. A
function is proposed which fits this relationship and which is equally
applicable for both simple and complex cells. The mean orientation
discrimination ({MOD}) was defined as that change in orientation angle away
from the optimal which produced a response statistically different --- on the
1 \% level --- from the response to the optimal orientation. There were
differences in {MOD} between the two sides of the orientation tuning curve:
the mean of the smaller of the two values was 13.5 deg and of the larger 19.7
deg. No significant difference in {MOD} was found between simple and complex
cells despite the fact that the halfwidth of the tuning curves for the two
cell types was 19.5 deg and 31.6 deg, respectively. The preciseness in
localization of the most sensitive part within the receptive field of single
cells was calculated from the variability in time of occurrence of the
smallest interspike interval. The degree of preciseness was found to be of
the order of 1/4 of the receptive field diameter in both simple and complex
cells. When nonoptimal stimulus orientations were presented, the preciseness
significantly decreased in complex cells whereas it remained unchanged in
simple cells. It is suggested that the same type of intracortical wiring
produces orientation selectivity in simple and complex cells, and that the
differences in tuning width are mainly due to a larger extension of
inhibitory fields in the simple cells. Considering the cortical visual cells
as elementary units in a network built for orientation detection and
discrimination, the tuning width seems of minor importance for that
function.%
}
\field{booktitle}{Experimental Brain Research}
\verb{doi}
\verb 10.1007/bf00239727
\endverb
\field{number}{2}
\field{pages}{197\bibrangedash 211}
\field{title}{Response variability and orientation discrimination of single
cells in striate cortex of cat}
\verb{url}
\verb http://dx.doi.org/10.1007/bf00239727
\endverb
\field{volume}{32}
\field{journaltitle}{Experimental Brain Research}
\field{year}{1978}
\endentry
\entry{Dean81}{article}{}
\name{author}{1}{}{%
{{}%
{Dean}{D.}%
{A~F.}{A.~F.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Springer}%
}
\strng{namehash}{DAF1}
\strng{fullhash}{DAF1}
\field{number}{4}
\field{pages}{437\bibrangedash 440}
\field{title}{The variability of discharge of simple cells in the cat striate
cortex}
\field{volume}{44}
\field{journaltitle}{Experimental Brain Research}
\field{year}{1981}
\endentry
\entry{Schiller76}{article}{}
\name{author}{3}{}{%
{{}%
{Schiller}{S.}%
{P.~H.}{P.~H.}%
{}{}%
{}{}}%
{{}%
{Finlay}{F.}%
{B.~L.}{B.~L.}%
{}{}%
{}{}}%
{{}%
{Volman}{V.}%
{S.~F.}{S.~F.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{SPHFBLVSF1}
\strng{fullhash}{SPHFBLVSF1}
\field{number}{2}
\field{pages}{347\bibrangedash 349}
\field{title}{Short-term response variability of monkey striate neurons.}
\verb{url}
\verb http://view.ncbi.nlm.nih.gov/pubmed/816424
\endverb
\field{volume}{105}
\field{journaltitle}{Brain research}
\field{month}{03}
\field{year}{1976}
\endentry
\entry{Vogels89}{article}{}
\name{author}{3}{}{%
{{}%
{Vogels}{V.}%
{R.}{R.}%
{}{}%
{}{}}%
{{}%
{Spileers}{S.}%
{W.}{W.}%
{}{}%
{}{}}%
{{}%
{Orban}{O.}%
{G.~A.}{G.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Springer-Verlag}%
}
\keyw{taouali15}
\strng{namehash}{VRSWOGA1}
\strng{fullhash}{VRSWOGA1}
\field{abstract}{%
In order to relate single cell performance to behavioral discrimination one
needs measurements of the response variance of the units. We recorded from
183 single units of area V1 of monkeys performing an orientation
discrimination task. The response variance was found to increase with
increasing response strength. This relationship between response variance and
response strength was well described by a power function with a power close
to one. The response variance was on average 1.9 times the response strength.
Despite important differences in preparation, the behaving monkey data are in
good agreement with those previously obtained in paralysed and anesthetised
animals.%
}
\field{booktitle}{Experimental Brain Research}
\verb{doi}
\verb 10.1007/bf00275002
\endverb
\field{number}{2}
\field{pages}{432\bibrangedash 436}
\field{title}{The response variability of striate cortical neurons in the
behaving monkey}
\verb{url}
\verb http://dx.doi.org/10.1007/bf00275002
\endverb
\field{volume}{77}
\field{journaltitle}{Experimental Brain Research}
\field{year}{1989}
\endentry
\entry{Gur06}{article}{}
\name{author}{2}{}{%
{{}%
{Gur}{G.}%
{Moshe}{M.}%
{}{}%
{}{}}%
{{}%
{Snodderly}{S.}%
{D.~Max}{D.~M.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Oxford University Press}%
}
\keyw{taouali15}
\strng{namehash}{GMSDM1}
\strng{fullhash}{GMSDM1}
\field{abstract}{%
The reliability of neuronal responses determines the resources needed to
represent the external world and constrains the nature of the neural code.
Studies of anesthetized animals have indicated that neuronal responses become
progressively more variable as information travels from the retina to the
cortex. These results have been interpreted to indicate that perception must
be based on pooling across relatively large numbers of cells. However, we
find that in alert monkeys, responses in primary visual cortex (V1) are as
reliable as the inputs from the retina and the thalamus. Moreover, when the
effects of fixational eye movements were minimized, response variability
(variance/mean --- Fano factor, {FF}) in all V1 layers was low. When
presenting optimal stimuli, the median {FF} was 0.3. High variability, {FF}
โผ 1, was found only near threshold. Our results suggest that in natural
vision, suprathreshold perception can be based on small numbers of optimally
stimulated cells.%
}
\verb{doi}
\verb 10.1093/cercor/bhj032
\endverb
\field{number}{6}
\field{pages}{888\bibrangedash 895}
\field{title}{High Response Reliability of Neurons in Primary Visual Cortex
(V1) of Alert, Trained Monkeys}
\verb{url}
\verb http://dx.doi.org/10.1093/cercor/bhj032
\endverb
\field{volume}{16}
\field{journaltitle}{Cerebral Cortex}
\field{month}{06}
\field{year}{2006}
\endentry
\entry{Kayser10}{article}{}
\name{author}{3}{}{%
{{}%
{Kayser}{K.}%
{C.}{C.}%
{}{}%
{}{}}%
{{}%
{Logothetis}{L.}%
{N~K.}{N.~K.}%
{}{}%
{}{}}%
{{}%
{Panzeri}{P.}%
{S.}{S.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{KCLNKPS1}
\strng{fullhash}{KCLNKPS1}
\field{number}{1}
\field{pages}{19\bibrangedash 24}
\field{title}{Visual enhancement of the information representation in
auditory cortex}
\field{volume}{20}
\field{journaltitle}{Current Biology}
\field{year}{2010}
\endentry
\entry{Maimon09}{article}{}
\name{author}{2}{}{%
{{}%
{Maimon}{M.}%
{Gaby}{G.}%
{}{}%
{}{}}%
{{}%
{Assad}{A.}%
{John~A.}{J.~A.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{MGAJA1}
\strng{fullhash}{MGAJA1}
\field{abstract}{%
{SummaryCortical} areas differ in their patterns of connectivity, cellular
composition, and functional architecture. Spike trains, on the other hand,
are commonly assumed to follow similarly irregular dynamics across neocortex.
We examined spike-time statistics in four parietal areas using a method that
accounts for nonstationarities in firing rate. We found that, whereas neurons
in visual areas fire irregularly, many cells in association and motor-like
parietal regions show increasingly regular spike trains by comparison.
Regularity was evident both in the shape of interspike interval distributions
and in spike-count variability across trials. Thus, Poisson-like randomness
is not a universal feature of neocortex. Rather, many parietal cells have
reduced trial-to-trial variability in spike counts that could provide for
more reliable firing-rate signals. These results suggest that spiking
dynamics may play different roles in different cortical areas and should not
be assumed to arise from fundamentally irreducible noise sources.
{IntroductionIt} is widely held that neocortical neurons in vivo fire action
potentials with irregular interspike intervals. Extensive effort has been
aimed toward understanding the functional significance of this observation.
In one view, irregularity reflects noise: cells fire at certain rates, but
the exact timing of spikes is random (Shadlen and Newsome, 1998) and is often
modeled as a Poisson process with a brief refractory period between action
potentials ([Gerstein and Mandelbrot, 1964] and [Kara et al., 2000]).
Consistent with a Poisson model, variance-to-mean ratios of spike counts in
visual cortex often equal or just exceed unity ([Britten et al., 1993],
[Buracas et al., 1998], [{McAdams} and Maunsell, 1999] and [Tolhurst et al.,
1983]). It has been posited that Poisson-like spike trains are the
fundamental unit of cortical communication ([Geisler et al., 2005], [Ma et
al., 2006], [Mazurek and Shadlen, 2002] and [Shadlen and Newsome,
{1998]).The} Poisson interpretation of irregularity is inconsistent with some
recent observations. For example, in rat auditory cortex, many cells fire a
single spike in response to pure tone stimuli, and the probability of
observing a spike is better modeled as a binomial rather than a Poisson
process ({DeWeese} et al., 2003). Other studies report that neurons in visual
cortex can exhibit {sub-Poisson} variability in anesthetized, paralyzed
preparations (Kara et al., 2000) when responding to natural images
(Amarasingham et al., 2006) or after accounting for the effects of small eye
movements (Gur and Snodderly, 2006). These data argue that the Poisson model
may provide an overestimate of noise levels in the system and that spike-time
statistics may vary across the neocortical {mantle.In} this light, we decided
to systematically examine spike-time statistics in parietal cortex. An
important axis upon which spiking neurons vary is the regularity of
interspike intervals. At one end of such a spectrum are irregular-spiking
cells: neurons that show Poisson-like, or {super-Poisson}, variability
(Figure 1, top left). At the other end are highly regular, clock-like
neurons, such as those found in the brainstem or cerebellum, in which there
is very little difference between neighboring interspike intervals (Figure 1,
bottom right). In this study, we wished to determine where cortical cells lie
along this continuum. In particular, we wanted to answer two questions.
First, do parietal neurons show characteristic, first-order, interspike
interval statistics independent of firing rate or task? Individual cortical
cells are known to have distinctive anatomy, connectivity, and intrinsic
electrophysiological properties; perhaps single parietal neurons also show
idiosyncratic interspike-interval statistics in vivo. Second, if individual
cells do show distinctive regularity or irregularity in their interspike
intervals, do the spiking statistics of single cells differ across parietal
regions? Whereas regular spiking is common in the peripheral nervous system
(Werner and Mountcastle, 1965) and subcortically ([Calvin and Stevens, 1968]
and [Pfeiffer and Kiang, 1965]), neocortical neurons are thought to fire in a
Poisson-like fashion in awake animals. This view, however, arises mainly from
recordings in the visual system. Fewer studies have examined spike timing
outside of sensory areas ([Averbeck and Lee, 2003], [Churchland et al.,
2006], [Compte et al., 2003], [Joelving et al., 2007] and [Lee et al.,
1998]). Working with awake, behaving monkeys, we recorded from single neurons
in four parietal cortical areas (Figure {2A}): the middle temporal ({MT}) and
medial superior temporal ({MST}) visual areas, the lateral intraparietal
({LIP}) association area, and area 5, a motor-like subdivision of the
parietal lobe in which cells activate with contralateral arm movements. We
found that cells in {MT} and {MST} typically show irregular interspike
intervals and their spike-count variance-to-mean ratios are near 1. In
contrast, many neurons in {LIP}, and an even larger fraction of cells in area
5, fire with more regular interspike intervals, with variance-to-mean ratios
clearly below 1. We show that cells with increased regularity from Poisson
tend to have large, broad extracellular waveforms, suggesting that at least
some of these neurons are pyramidal cells ([Bartho et al., 2004],
[{McCormick} et al., 1985] and [Nowak et al., 2003]) that may function to
send a lower-noise spike-rate signal subcortically or to other cortical
sites. {ResultsWe} recorded neural activity from two monkeys as they
performed the following task. The animals viewed a dot approaching a bar on a
computer screen and were trained to make visually guided arm movements in one
of two modes (Figure {2B}) (Experimental Procedures and Supplemental
Experimental Procedures). In the proactive mode, the monkeys pressed a lever
to cause the dot to reverse its motion direction before it collided with the
bar. In the reactive mode, the dot reversed its motion direction under
computer control, and the monkeys pressed a lever in response to this abrupt
cue. This task allowed us to characterize brain processes underlying the
initiation of proactive versus reactive arm movements, described previously
([Maimon and Assad, 2006a] and [Maimon and Assad, {2006b]).In} addition, this
experiment was advantageous for assessing spiking statistics of cortical
neurons. First, we recorded from parietal regions across the visual-to-motor
spectrum, using the same methodology in each region. Second, the monkeys
performed a large number of identical trials allowing us to robustly estimate
statistical distributions. Third, as the task was demanding, spike-count
variability on correct trials was unlikely to arise from trial-to-trial
differences in overall states of arousal/vigilance. {Single-Neuron}
{ExamplesFigure} {2C} (top) shows activity from two visually responsive
cells, aligned to the time of the dot's motion reversal. The stimulus was
oriented such that these direction-selective neurons were stimulated in their
preferred direction after the motion reversal. Whereas the two cells fired at
similar rates, they exhibited very different patterns of spiking. The {MST}
cell fired irregularly, whereas the {LIP} cell had more evenly spaced spike
times. To the right, we show interspike interval ({ISI}) distributions from a
400 ms time window, starting 100 ms after the turn. The {MST} neuron had
intervals that approximate an exponential distribution, a signature of a
Poisson process. The {LIP} neuron had a more rounded, compact {ISI}
distribution, showing that for a similar mean interval length the {LIP} cell
had fewer long and short intervals about that mean, a phenomenon we call ''
pulsed firing.'' Data from two motor-like cells in area 5 (Figure {2C},
bottom), aligned to the time of the hand movement, make the same point. Some
cortical neurons fire with Poisson-like irregularity, but others fire in a
more regular fashion than Poisson. Definition of Pulsed Firing and Pulsing
{NeuronsAt} high firing rates, where the mean interspike interval approaches
the timescale of the refractory period, even a Poisson-like process appears
more regular (Gabbiani and Koch, 1998). This is because refractoriness
precludes very short intervals and a high firing rate precludes very long
intervals, which together give rise to {ISI} distributions that are
necessarily more compact and non-exponential-like. A related issue is that
some cortical neurons fire action potentials in bursts of two or more spikes
in rapid succession, during which firing rates can be extremely high.
Bursting can thereby introduce periodic structure to spike trains (Bair et
al., 1994). Neurons we recorded in parietal cortex, however, showed regular
spiking at both low rates and high rates, and they did not burst (e.g.,
Figure {2C}, cells in red). We therefore define a '' pulsing neuron'' as a
cell whose interspike intervals are regular independent of firing rate, i.e.,
not due to simple refractoriness, and whose regularity does not result from a
tendency to burst. We use the term '' pulsing neuron'' by analogy to the
pulse of the heart, which changes rate over time, but maintains a fundamental
regularity. As will become clear, pulsing neurons have interspike intervals
that are more regularly distributed than those derived from a Poisson-like
process; however, such cells in cortex do not approach the level of
oscillatory precision observed in many subcortical neurons, let alone
man-made clocks. To emphasize this latter point, we avoid the label ''
periodic neuron'' or '' oscillatory neuron.'' {Rate-Parsed} Interspike
Interval {HistogramsAs} a measure of regularity, we quantitatively
characterized {ISI} distributions. One complication with analyzing {ISI}
plots was the fact that each cell's spike rate varied considerably over time.
This nonstationarity meant that we could not simply plot a single {ISI}
histogram for each cell, as this histogram would have undoubtedly resembled a
sum of more basic distributions, one for each rate at which the neuron
{fired.To} address this issue, we adopted a method for plotting multiple
{ISI} histograms per neuron, in which each {ISI} histogram consists of
interspike intervals associated with a similar mean firing rate (Softky and
Koch, 1993) (Experimental Procedures). In this analysis, we first calculated
a firing-rate histogram across all trials (peri-event time histogram, {PETH};
50 ms bins) aligned on a trial-specific event, such as the time of the hand
movement or the dot's motion reversal on the screen (Figure {3A}, gray
curve). Then, for each trial we uniformly multiplied this overall {PETH} by a
scalar, defined as the number of spikes on the currently analyzed trial
divided by the average number of spikes across all trials (Softky and Koch,
1993). For example, if we observed 50 spikes on a given trial, whereas the
mean number of spikes was 30, we multiplied each bin of the overall {PETH} by
1.67 (Figure {3A}, red curve), so as to account for the increased response
gain on that specific trial. If we observed 17 spikes, then we multiplied
each bin by 0.57 (Figure {3A}, black curve), so as to account for the lower
gain. Next, we associated each interspike interval with the trial-specific
firing rate at the central time point of that interval (e.g., dotted red line
in Figure {3A} associates an interspike interval of 43 ms with a firing rate
of 61 spikes/s). Finally, we plotted an {ISI} histogram for intervals
associated with 0--5 Hz firing, a separate histogram for 5--10 Hz firing, and
so on, in 5 Hz steps, up to the maximum rate at which the cell fired. Having
generated a set of {ISI} histograms for each cell, we fit these histograms to
simple functions so as to describe their shape with a small number of
parameters. Visual inspection of the data, as well as past work ([Kuffler et
al., 1957] and [Seal et al., 1983]), argued that {ISI} plots would be well
modeled as gamma distributions. Gamma distributions are defined by a scale
parameter (ฮธ) and a shape parameter (k), with k providing a useful measure
of {ISI} regularity (Figures 1 and S1). That is, when k = 1, gamma
distributions are mathematically equivalent to exponential distributions,
suggestive of a Poisson process. When k > 1, the peaks of the gamma
distributions are offset from zero, suggestive of increased regularity from
Poisson. When k < 1, gamma distributions have an excess of very short and
very long intervals, suggestive of increased irregularity from Poisson. Thus,
the shape parameter provides a convenient axis for separating pulsed spike
trains, with relatively high k, from more irregular spike trains, with low
{k.Before} applying this {ISI} analysis to our complete dataset, there were
some important issues to consider. First, we only fit {ISI} histograms with
>150 interspike intervals, as histograms with fewer data points were
undersampled (e.g., lower-left histogram in Figure {3B}). Second, since we
wanted to detect neurons whose {ISI} regularity was not a direct result of
refractoriness when firing at a high rate, we focused on the same relatively
low spike rates, 10--40 Hz, for all neurons. More than 90\% of cells provided
sufficient interspike intervals in the 10--40 Hz range (see below), allowing
us to effectively compare the shape of {ISI} histograms across neurons after
normalizing for firing rate, without excluding many cells from analysis.
Third, since we were interested in examining neurons whose {ISI} regularity
was not due to simple bursting or refractoriness, we deliberately ignored
very short interspike intervals (โค8 ms) when fitting gamma distributions to
{ISI} plots (Experimental Procedures; we return to these short intervals
{later).ISI} histograms, across the board, were very well described as gamma
distributions (Experimental Procedures). In Figure {3B}, we show rate-parsed
{ISI} histograms from the four example cells in Figure {2C}. For each cell,
we found that the shape parameter typically increased slightly with firing
rate but otherwise remained consistent across the fitted histograms (e.g.,
Figure {3B} k1 -- k6, quantified in more detail below). The area 5 neuron in
the bottom row of Figure {3B} shows one of the largest variations in shape
parameter observed in our dataset, \\\\$\\\\\\\\pm\\\\$35\% about the mean,
yet even this cell's histograms did not change character from a higher-order
gamma to a more exponential-like appearance across different rates. For a
given neuron, we averaged the shape parameter, k, across each of the fitted
histograms, generating a single measure of {ISI} regularity for each cell.
The {MST} neuron had a mean k of 0.9, consistent with irregular interspike
intervals. The {LIP} neuron had a mean k of 5.1, evidence for pulsed firing.
The irregular-firing neuron from area 5 had a mean k of 0.6, well below 1,
suggestive of {super-Poisson} irregularity. The other area 5 neuron had a
mean k of 7.6, evidence for pulsed {firing.To} verify that our gamma model
was not generating a biased picture, we also calculated the coefficient of
variation ({CV}) of each {ISI} distribution, which provided a parameter-free
method of describing the {ISI} plots. {CV} is a measure of dispersion,
defined as the standard deviation divided by the mean. Exponential
distributions have a {CV} of 1, whereas distributions derived from more
regular interspike intervals have {CV} values below 1. As with the gamma
fits, we only calculated {CV} for well-sampled histograms (>150 interspike
intervals), and we averaged {CV} across histograms. In Figure {3B}, the two
irregular cells had mean {CV} values of 1.1 and 1.2, whereas the two pulsing
cells had {CV} values of 0.52 and 0.39. What Is the Function of Pulsed
{Spiking?One} important consideration is that pulsed spike trains are
expected to have lower spike-count variance than irregular spike trains; in
the extreme, a grandfather clock will always tick ten times in 10 s, whereas
a Geiger counter measuring radioactive decays at a rate of 1 per s will show
variability about this value. Thus, to the extent that cortical neurons
communicate with spike-rate signals, pulsing neurons should provide a more
reliable form of this signal. For each bin of the {PETH}, we calculated the
mean and variance of the number of spikes recorded across trials and plotted
a single point on a variance-versus-mean log-log plot (Figure {3B}, right).
The two irregular-firing cells had points that fell along the diagonal, or
slightly above it, indicating that spike-count means and variances were
roughly equal, as predicted from a Poisson process (Figure {3B}, first and
third rows). In contrast, the two pulsing cells had points that fell below
the diagonal, indicating lower variance than predicted from a Poisson process
(Figure {3B}, second and fourth rows). To quantify these data, we calculated
the mean residual (vertical distance) of points from the diagonal. The two
irregular cells had mean residuals that were slightly positive (0.13, 0.25),
whereas the two pulsing cells had mean residuals that were substantially
negative (--0.75, --1.2). In summary, for each neuron, we calculated three
metrics to describe its spike-train statistics. Two metrics, k and {CV},
captured the shape of {ISI} distributions, whereas the third metric, mean
residual, captured spike-count variability across trials. Note that with mean
residual we did not parse the intervals in any way, and we examined all
firing rates for each cell, not just 10--40 Hz. Randomness and Regularity
across Parietal {CortexOur} complete dataset consisted of 539 parietal
neurons: 102 in {MT}/{MST}, 203 in {LIP}, and 234 in area 5. We analyzed
94\\\\\\\\ of {MT}/{MST} cells (96 of 102), 95\% of {LIP} cells (193 of 203),
and 94\\\\\\\\ of area 5 cells (219 of 234); other neurons did not fire
sufficiently between 10 and 40 Hz to provide at least one {ISI} histogram
with >150 interspike intervals. Overall, spiking in {MT}/{MST} was irregular
(median k = 1.23), with increasing regularity in {LIP} (median k = 1.95) and
area 5 (median k = 2.70) (Figure {4A}, left column). All three
shape-parameter distributions were significantly different from one another
(rank sum test; p < 10โ7), demonstrating that first-order spike-time
statistics vary across parietal regions. Cells with high k we call '' pulsing
neurons'' and those with low k '' irregular neurons,'' but cells fell along a
continuum with no obvious boundary separating the two extremes. Nevertheless,
if one defines an arbitrary threshold at k > 3, one finds that 41\% of area 5
neurons exceeded this threshold and had spike trains that appeared
definitively regular to the naked eye (Supplemental Data). 21\% of {LIP}
cells, but only 4\% of {MT}/{MST} cells (4 neurons), had k > 3. Results were
comparable with our alternative measure of {ISI} regularity, mean {CV}
(Figure {4A}, middle column). As {CV} is a ratio ({SD}/mean), we plotted {CV}
on a logarithmic scale in which values near zero indicated Poisson-like
spiking and values below zero indicated pulsed spiking. We observed a median
{log(CV}) of 0.01 for {MT}/{MST}, --0.04 for {LIP}, and --0.13 for area 5.
All three distributions were significantly different from one another (rank
sum test; p < 10โ4). For reference, we also show {CV} distributions on a
linear scale (Figure S2). {CV} and k were highly correlated (Figure S3)
arguing that each provided a similar, reliable measure of {ISI}
{regularity.For} each cell, we generated a variance-versus-mean log plot and
calculated the mean residual from the equivalence line, averaged across all
bins. Neurons with low {log(CV}) tended to have negative mean residuals
(Figure {4A}, right column). Correlation coefficients were 0.90 for
{MT}/{MST}, 0.84 for {LIP}, and 0.85 for area 5, all highly significant (F
test, p < 10โ15). In other words, pulsing cells have lower spike-count
variability than irregular cells, demonstrating that pulsing neurons carry a
more reliable spike-rate signal than previously thought to exist in the
neocortex. This finding was not critically dependent on the bin width used
for {PETHs} or on how we aligned and processed the data (Figure S4). Firing
{RatesOur} analysis assumes that spike-train irregularity or regularity is a
defining feature of a neuron and not something that changes dramatically with
firing rate. Stated formally, our working model is that, for a given neuron,
interspike intervals are drawn from a gamma distribution with a consistent
shape parameter (k), but a scale parameter (ฮธ) that changes with spike rate
(see Figure S1). If this model is valid, k measured for a given neuron should
be the same independent of firing rate, which was largely the case (Figure
{4B}). Pulsing neurons did not turn irregular at certain spike rates or vice
{versa.There} was, however, a weak tendency for k to increase with higher
rates, and for {log(CV}) to decrease (Figure {4B}), likely because of the
refractory period. Thus, if area 5 neurons fired faster than {MT}/{MST}
neurons this could have contributed to the increased regularity in area 5. To
control for firing-rate differences across areas, we only analyzed {ISI}
plots in the 10--40 Hz range. Nevertheless, {MT}/{MST} cells might have fired
closer to 10 Hz and area 5 cells closer to 40 Hz, for {example.To} address
this issue, we examined spike rates across areas (Figure 5). We plotted a
vertical line from the minimum to the maximum firing rate exhibited by each
cell and positioned this line along the x axis based on the cell's {log(CV})
measure (Figure 5, left). There was substantial overlap in the range of
firing rates among individual cells across all cortical areas. We also
plotted the maximum, minimum, mean, and standard deviation of spike rates as
a function of {log(CV}) (Figure 5, right). Mean spike rates in {MT}/{MST}
(mean = 16.8 Hz) and area 5 (mean = 16.6 Hz) did not differ significantly
(rank sum test, p = 0.93) and thus cannot explain the observed differences in
{ISI} regularity. Standard deviations of firing rates in {MT}/{MST} (mean =
12.3 Hz) were significantly higher than those in {LIP} (mean = 10.3 Hz) and
area 5 (mean = 7.6) (rank sum test, p < 10โ4). This larger range of spike
rates for {MT}/{MST} neurons was likely because we optimally positioned the
visual stimulus along their preferred-null directional axis (Experimental
Procedures). In contrast, the animals always performed the same lever press,
even though many area 5 cells are selective for the direction of arm
movements (Kalaska et al., 1983). Only 1 of 12 plots relating firing rate
parameters to {log(CV}) (Figure 5 right) showed a significant correlation (p
< 0.05, Bonferroni correction for 12 comparisons); area 5 neurons with low
{log(CV}) tended to have slightly elevated minimum spike rates (r = --0.29; p
= 1.8 ร 10โ5). Overall, these data emphasize the extensive overlap in
firing rates between irregular and pulsing cells and across cortical regions.
Differences in {ISI} regularity are thus unlikely to be due to systematic
differences in firing rate. {BurstingAnother} concern is that increased {ISI}
regularity could result from bursting (Bair et al., 1994). For example, we
observed {ISI} histograms that had a peak at short interspike intervals,
reflective of bursting, followed by a trough and then a second, lower peak
(Figure {6A}). Since we excluded the first 8 ms in our gamma fits, these
cells would have had elevated k parameters simply because of the postburst
trough, which likely reflects a postburst refractory period (Bair et al.,
1994). To examine the role of bursting in pulsed firing, we measured the
proportion of spikes in the 0--8 ms bins of each {ISI} histogram and compared
this value with the proportion of spikes predicted based on the gamma fit,
which excluded these bins. For example, the histogram in Figure {6A} shows
data from a cell with a large difference between the actual and predicted
number of spikes in the 0--8 ms bins. In contrast to this example, however,
we generally observed a good match between predicted (red symbols) and
observed (black symbols) values, with perhaps slightly more intervals
observed than predicted for rates close to 40 Hz (Figure {6B}). The most
bursting was in {MT}/{MST}, less in {LIP}, least in area 5. That is, 57\% of
{ISI} histograms in {MT}/{MST} had more 0--8 ms intervals than predicted by a
Poisson process, i.e., an exponential distribution (Figure {6B}, gray lines),
whereas only 27\% of histograms in {LIP} and 17\% in area 5 showed this
trend. Bursting did not contribute to the pulsed spike trains characterized
here. In particular, area 5 cells showed the least bursting and most
spike-time regularity. Controls for the Visual Stimulus and
{MicrosaccadesFor} two reasons, it is possible we overestimated spike-time
irregularity and spike-count variability in {MT}/{MST}. First, we typically
aligned {MT}/{MST} spike trains to the time of the dot's motion reversal,
which had
3\\\\$\\\\\\\\,^{\\\\\\\\circ}\\\\$--5\\\\$\\\\\\\\,^{\\\\\\\\circ}\\\\$ of
trial-to-trial jitter (Experimental Procedures). Since receptive fields in
{MT}/{MST} are not homogeneous, this trial-to-trial difference in the visual
stimulus could have increased the variance of spike counts across trials.
Second, microsaccades (small, rapid eye movements) within the fixation window
would have produced irregular movements of the visual stimulus on the retina
that could have led to irregular changes in the firing rate of {MT}/{MST}
cells, particularly since our stimuli had sharp contrast edges. Area 5 cells
have very weak, if any, visual responses and would not have been pushed
toward irregularity in the same {manner.To} address trial-to-trial jitter in
the location of the dot's motion reversal, we analyzed spike trains up to,
but not including, the dot's earliest motion reversal among all trials
(Supplemental Experimental Procedures). In this analysis, the visual stimulus
on the monitor was identical across trials, yet we observed the same trends
of irregularity and variability across the three brain regions (Figure {7A}).
To address the role of small eye movements, we extracted the time of
microsaccades from our dataset and found, serendipitously, that one of our
monkeys performed microsaccades at twice the rate of the other animal (Figure
S5). Despite this difference in saccade rate, the two animals showed very
similar spiking statistics in {MT}/{MST} (as well as in {LIP} and area 5),
suggesting that microsaccades did not have a strong influence on our results.
We also analyzed data from the baseline fixation period (Supplemental
Experimental Procedures), during which there was no visual stimulus on the
screen and thus microsaccades are less likely to have introduced visually
driven irregularity. In addition, we recorded from the same population of
cells as the animals performed a memory-saccade task, and in the memory-delay
epoch of this task there was no visual stimulus in the receptive field
(Supplemental Experimental Procedures). In both the fixation period of the
main task as well as the memory period of the memory-saccade task, we
observed the expected pattern of irregularity and variability across parietal
areas (Figures {7B} and {7C}). Note that we recorded fewer trials in the
memory-saccade task and fewer cells provided sufficient interspike intervals
for analysis. Overall, {MT}/{MST} neurons exhibited similarly irregular
firing across a broad range of conditions, arguing that spike-time
irregularity is a general feature of {MT}/{MST} and not an artifact of data
alignment or eye movements. Indeed, in all areas, spike-time regularity or
irregularity remained a consistent feature of neurons across different tasks
and time epochs within each task. Pulsed Firing Is Not Poisson-like Firing
with an Extended Refractory {PeriodTo} minimize the influence of the
classical refractory period on our results, we focused on relatively long
interspike intervals, associated with 10--40 Hz spiking, in our {ISI}
analyses. However, the relative refractory period could still play a role at
these rates. One possibility is that pulsing neurons were in fact
Poisson-like but had very long relative refractory periods, lasting tens of
milliseconds, which would result in fewer short intervals and thus higher k.
If so, one would expect the tails of {ISI} histograms, after the effects of
refractoriness have completely subsided, to resemble exponential
distributions, not gamma distributions, but this was not the case (Figure
S6). Rather, interspike intervals from pulsing neurons seemed genuinely well
captured as draws from gamma distributions with k > 1. One could still
consider pulsing cells to have extended refractory periods, in that these
cells have reduced probability of very short intervals. However, these
neurons also have reduced probability of very long intervals, meaning that an
extended refractory period is not the only difference between pulsed and
irregular spiking. Spike Shapes and {AmplitudesDo} irregular and pulsing
neurons correspond to distinct cellular subtypes? Extracellular spike
waveforms can provide some insight into this question. The width of
action-potential waveforms in cortex fall into two distinct categories,
narrow and broad ([Bartho et al., 2004], [{McCormick} et al., 1985],
[Mitchell et al., 2007], [Mountcastle et al., 1969] and [Rao et al., 1999]),
and some have argued that narrow spikes correspond mainly to inhibitory
interneurons, whereas broad spikes correspond to a wider set of neurons,
including pyramidal cells ([{McCormick} et al., 1985], [Mitchell et al.,
2007] and [Nowak et al., 2003]). We measured the time interval between trough
and peak of our spike waveforms (Experimental Procedures). 5\%--12\% of
neurons in each area had waveforms in which the peak occurred before the
trough; these cells were not analyzed further. For the remaining cells, we
found a bimodal distribution of spike widths in {MT}/{MST}, {LIP}, and area 5
(Figure {8A}), much like reported previously in area V4 (Mitchell et al.,
2007). We observed a weak, but significant correlation between spike width
and {log(CV}) both in area 5 (r = โ0.33, p = 1.6 ร 10โ6) and {LIP} (r =
โ0.29, p = 1.1 ร 10โ4), but not in {MT}/{MST} (r = โ0.21, p = 0.06)
(Figure {8B}). As a measure of spike amplitude, we also calculated a ''
signal-to-noise'' ratio for each spike waveform (Figure {8C}) (Experimental
Procedures). Mean signal-to-noise ratios were 3.3 for {MT}/{MST}, 4.2 for
{LIP}, and 4.8 for area 5 (Figure {8C}). All distributions were significantly
different from one another (rank sum test, p < 10โ3). Cells with large
signal-to-noise ratios tended toward pulsed firing. That is, there was a weak
but significant correlation between signal-to-noise and {log(CV}) in area 5
(r = --0.22, p = 0.0012), {LIP} (r = --0.21, p = 0.004), and {MT}/{MST} (r =
--0.45, p = 5.2 ร 10โ6) (Figure {8D}). These correlations remain
significant if we exclude one outlier in {MT}/{MST} with a signal-to-noise
ratio of 12 and one outlier in area 5 with a signal-to-noise ratio of 27
(data point not shown). Big, broad waveforms suggest that many pulsing cells
are large pyramidal neurons rather than smaller interneurons. {DiscussionAn}
influential view holds that spike timing in the neocortex is fundamentally
irregular. In contrast to this model, we have identified parietal neurons
that fire with increased {ISI} regularity. These '' pulsing neurons'' fire in
a notably regular fashion across different spike rates (Figure 4), different
time periods within a task (Figure 7), and different tasks (Figure 7),
suggesting that regularity is a defining feature of these cells and not a
property that changes with task demands. Pulsed spiking serves at least one
important function: increasing the reliability of spike-rate signals (Figure
{4A}, right column). Pulsing neurons are most common in association and
motor-like regions of the parietal lobe, less common in visual areas (Figure
{4A}, left and center column). Although past reports have not emphasized
regular spiking in neocortex of awake primates, {ISI} histograms in rat
archicortex (hippocampus) also appear better described by gamma rather than
Poisson models, after accounting for the extra parameter in the gamma model
(Barbieri et al., 2001). Additionally, a common method for estimating
neuronal latencies relies on modeling {ISI} histograms as gamma distributions
(Seal and Commenges, 1985), a method first developed using spike trains from
area 5 (Seal et al., {1983).Since} pulsed spiking reduces variance in
spike-rate signals, why do so few {MT}/{MST} cells fire in this fashion? The
answer is not clear; however, if {MT}/{MST} cells must respond to visual
inputs with millisecond precision, pulsed firing is problematic, as there
would be interspike periods, sometimes lasting many milliseconds, with
extremely low probability of spiking. Motor computations, in comparison, may
operate on a longer timescale---as these signals ultimately get filtered by
muscles---and thus may be more suitable for pulsing. Consistent with this
notion, spike trains in the spinal cord are extremely regular (Calvin and
Stevens, 1968). Our results argue that irregular spiking in visual areas does
not reflect irreducible noise inherent to all cortical computation (Shadlen
and Newsome, 1998), since this noise is reduced in other cortical areas.
Rather, widespread irregularity in the visual system is more likely to
reflect noise specific to visual circuits or, alternatively, the need for
rapid temporal {signaling.Aside} from increasing the fidelity of spike-count
signals, what functional role is served by pulsed spiking? We examined
whether pulsing cells were more likely to participate in the proactive timing
of arm movements---the cognitive process assayed by our behavioral task---but
found no supporting evidence (Figure {S7A}). Additionally, there was no
detectable relationship between the width of spike waveforms and cognitive
firing-rate modulations in our task (Figure {S7B}). To constrain functional
interpretations, it will be important to determine the anatomical inputs and
outputs of pulsing neurons. {LIP} and area 5 send and receive signals from
subcortical motor circuits, which commonly show regular spiking ([Clower et
al., 2001], [Lynch et al., 1985] and [Sasaki et al., 1976]). Furthermore, it
has been shown that motor and cognitive regions of the frontal lobe exhibit
closed-loop connectivity with the basal ganglia and cerebellum, regions with
extensive rhythmicity (Kelly and Strick, 2003). Perhaps {LIP} and area 5 are
more influenced by this closed-loop architecture (Clower et al., 2005) than
{MT}/{MST}.A general concern with extracellular recording techniques is the
likely bias toward recording from large cells with high spike rates
(Olshausen and Field, 2005). In this study, we made an effort to record from
well-isolated units, large or small, that had even a minimal response to the
task. Nevertheless, it remains a possibility that spiking statistics were
identical across parietal areas, but pulsing cells in {MT}/{MST}, for
example, were too small for us to consistently isolate. Arguing against this
idea, however, pulsing cells in {LIP} and area 5 tended to have
large-amplitude waveforms (Figure 8) and high minimum firing rates (Figure
5). It is unlikely that we missed such neurons in our {MT}/{MST} recordings.
More probable is a true areal difference in the statistics of {firing.Past}
studies ([Bair and Koch, 1996], [Buracas et al., 1998] and [Masse and Cook,
2008]) have reported that {MT} neurons can generate spike trains with high
temporal precision across trials, which seems at odds with the notion that
spike trains in area {MT} are fundamentally irregular. In these studies,
however, temporal precision in {MT} responses was a direct result of dynamic
visual stimuli, which drove the cells rapidly from a low to high firing rate
at multiple time points in a trial. Spike trains were consistent across
trials precisely because the stimulus-induced rate changes occurred at
consistent moments across trials. However, the underlying spike-generating
process in {MT} neurons was still reported to be well modeled, to
first-order, as a fundamentally irregular Poisson-like process ([Bair and
Koch, 1996] and [Buracas et al., 1998]). For example, variance-to-mean ratios
of spike counts in one study were near 1 (Buracas et al., 1998). Indeed,
irregular neurons may be more capable than pulsing cells of phase locking to
dynamic inputs because irregular cells do not experience a prolonged
refractory state between spikes. More generally, one should carefully
distinguish stimulus-induced regularity, as found in {MT}, from a regular
underlying mechanism of firing, which appears to be a stimulus-independent
feature of cells in {LIP} and area 5 (Figure {7).Along} similar lines, a
notable set of papers have examined the responses of somatosensory cortical
neurons to vibrotactile stimuli ([Mountcastle et al., 1969], [Recanzone et
al., 1992], [Romo et al., 1998] and [Salinas et al., 2000]). In these
studies, it was shown that many neurons in the primary somatosensory cortex
(S1) respond with an oscillating firing rate to an oscillating tactile input
in the range of 5 to 100 Hz (Mountcastle et al., 1969). In the secondary
somatosensory cortex (S2), however, many fewer neurons respond periodically
to the oscillating input (Salinas et al., 2000). A progression from more
periodic firing in S1 to less periodic firing in S2 might seem at odds with
our finding of increased pulsed firing in higher compared to lower parietal
regions. Again, however, the periodicity observed in these studies was
directly due to periodicity present within the incoming stimulus. The results
are to be expected if S1 neurons have intrinsically irregular interspike
intervals, with short refractory periods, which allow them to better phase
lock to periodic inputs compared to S2 neurons. Indeed, a Poisson model for
the underlying spike-generating process provided a good fit for
information-transfer rates observed in S1 cells (Salinas et al., 2000). An
intriguing, albeit speculative, hypothesis is that higher somatosensory
regions may show more pulsed firing, and less periodic following of inputs,
than lower somatosensory {areas.At} this point, one might interpret our
results to mean that spike-time regularity increases consistently from
primary sensory to higher cortical regions. Whereas this trend appears to
hold in the parietal areas studied, it is unlikely to be uniformly true. For
example, it has been observed that many neurons in dorsolateral prefrontal
cortex fire quite irregularly, especially during memory periods of delayed
response tasks (Compte et al., 2003). Perhaps a more reasonable
interpretation of our data is that neurons involved in oculomotor and
especially skeletomotor computations show disproportionate levels of pulsed
firing. This interpretation raises the testable prediction of more pulsed
firing in primary motor cortex and frontal eye fields compared to {MT} or
{MST}. Indeed, if information is communicated in Morse-code like patterns of
interspike intervals in neocortex, a debated hypothesis, irregular spiking
would be associated with higher entropy levels and thus allow cells to
transmit more information per unit time. In this view, one might imagine
greater irregularity in higher centers of cognitive management, like the
dorsolateral prefrontal cortex, which could possibly require a higher
capacity for information transfer compared to regions involved in motor
{control.Increased} {ISI} regularity could result from membrane properties in
single neurons, circuit properties of networks of neurons, or a combination
of both. A clearer understanding of the mechanisms for pulsed spiking may
require work in more reduced preparations. Specifically, to disambiguate
cellular versus network mechanisms, one could test for regular firing in
dissociated neurons (e.g., Raman and Bean, 1997), where synaptic input does
not contribute to the {activity.A} related question is whether pulsing cells
show higher-order structure in their spike trains or whether, as our
first-order model suggests, interspike intervals are best captured as
independent draws from gamma distributions. Future work will be needed to
definitively sort this issue out. Nonetheless, it should be mentioned that
global electrical oscillations in the brain provide a potential source for
long-range order. The relationship between such oscillations and spike timing
is an important area of active research (Buzsaki, 2006). We could not
directly address how pulsed spiking relates to global brain rhythms because
we did not record multiple neurons simultaneously, nor did we measure
local-field or scalp potentials. Nevertheless, one way to approach this
question in our dataset is to examine single-cell autocorrelograms, which
plot the probability of observing a spike at a given time delay after another
spike. Unlike {ISI} distributions, autocorrelograms examine all intervals,
not just nearest-neighbor intervals. If pulsing cells were locked to a
circuit-level oscillation, one might expect distinct oscillatory peaks in
autocorrelograms, whereas if interspike intervals were generated
cell-autonomously then one would predict autocorrelograms to appear more
flat, especially at long time delays (Figure S8). Autocorrelograms of highly
regular area 5 spike trains did not exhibit strong oscillatory peaks,
suggesting that these interspike intervals may have been divorced from global
rhythms. This conclusion is tentative, however, given that very few parietal
cells fired with sufficient oscillatory precision so as to expect detectable
peaks in autocorrelograms, even if these cells were locked to network
oscillations (Figure S8). Additionally, nonstationary firing rates corrupt
periodic structure in autocorrelograms, and it is not clear how to normalize
autocorrelograms for firing rate given that these plots require an extended
time window for {analysis.No} single analysis provides a complete
characterization of neuronal spiking. We therefore also show unprocessed
rasters from our entire dataset (Supplemental Data). Across cells there was
an excellent match between the regularity of spike trains as assessed by eye
and as assessed by k or {CV} of rate-parsed {ISI} histograms. Examining the
entirety of our spike-train data makes clear the pronounced differences in
spike-time statistics across parietal neurons and places emphasis on
deciphering the origin of this impressive variability. Experimental
Procedures Electrophysiological {TechniqueTwo} male rhesus monkeys (Macaca
mulatta, 10 kg) were surgically implanted with a head post, scleral search
coil, and recording chamber. All surgical and experimental procedures
followed Harvard Medical School and {NIH} guidelines. The chambers were
dorsally positioned at stereotactic coordinates P3, L10 in the hemisphere
contralateral to the hand used for pressing a lever. Extracellular recordings
were made from single neurons using tungsten microelectrodes and a
guide-tube/grid system (see below). We recorded from four distinct cortical
regions: the middle temporal ({MT}) area, the medial superior temporal
({MST}) area, the lateral intraparietal ({LIP}) area, and area 5. Recording
sites were localized based on physiological responses and structural {MRI}
images, as described previously (Maimon and Assad, {2006a).We} recorded spike
times with 1 ms resolution using a dual time-and-amplitude window
discriminator (Bak Electronics, Mount Airy, {MD}). In parallel, the complete
voltage trace was streamed to disk and used offline for investigating spike
waveforms (see '' Extracellular Waveforms'' below). We recorded neurons one
by one, examining action potentials on an oscilloscope as they appeared. Only
unambiguous single units, as determined by eye, were analyzed. With 1 ms
spike-time resolution, even a perfectly isolated neuron could fire in
adjacent time bins; thus, we did not employ a quantitative requirement for a
refractory period. However, for reference, we show extracellular waveforms
from our entire population of cells (Supplemental Data). To minimize bias, we
recorded from all neurons, large or small in amplitude, that responded even
minimally to our behavioral {task.We} recorded a total of 539 cells: 309 in
monkey B and 230 in monkey R. From monkey B, we recorded 54 cells in
{MT}/{MST}, 127 cells in {LIP}, and 128 cells in area 5. From monkey R, we
recorded 48 cells in {MT}/{MST}, 76 cells in {LIP}, and 106 cells in area 5.
In early recordings, we used tungsten microelectrodes with a 250 ฮผm shaft
diameter ({FHC}, Bowdoin, {ME}). In later experiments, we used 75 ฮผm
electrodes ({FHC}, Bowdoin, {ME}), as these yielded more stable isolations.
With monkey B, 54/54 {MT}/{MST} cells, 5/127 {LIP} cells, and 25/128 area 5
cells were recorded with 250 ฮผm electrodes; the rest were recorded with 75
ฮผm electrodes. With monkey R, all cells were recorded with 75 ฮผm
electrodes. Our results in {MT}/{MST} were very similar between monkeys
(e.g., Figure S5), even though electrodes of different widths were used in
each animal. Extracellular {WaveformsThe} electrode signal was amplified,
band-pass filtered at 100--6000 Hz (3 db/down) with a notch filter at 60 Hz
(Bak Electronics, Mount Airy, {MD}) and streamed to disk at a 20 {kHz}
sampling rate ({CED}, Cambridge, England). Offline, we extracted the 80
voltage samples (4 ms) surrounding each spike time and averaged together the
first 100 spike waveforms to generate one mean waveform for each {cell.Spike}
times were recorded with 1 ms of uncertainty by the window discriminator;
thus, we needed a method of aligning waveforms before averaging. Instead of
aligning to the time of peak or trough, which could vary by up to 50 ฮผs (one
sample period), we used an algorithm that took into account all waveform
sample points. For each cell, we slid the first waveform past the second
waveform and found the time shift that maximized the correlation between the
two curves. The pair of waveforms were aligned to this time shift and
averaged together. We then slid the third waveform past the mean waveform
from the first two spikes and, again, found the time point that maximized the
correlation. The third waveform and mean spike were aligned to this time
shift and a weighted average was taken, in which the mean waveform was
weighted by 2/3 (as it included two spikes) and the third waveform weighted
by 1/3. We continued this process for 100 spikes, thus generating the final
mean waveform for each {cell.We} interpolated the mean waveform with a cubic
spline to achieve 10 ฮผs temporal resolution. We then extracted two metrics
for each cell, spike width and signal-to-noise. Spike width was defined as
the time difference between the maximum and minimum voltages of the mean
waveform. Spike signal-to-noise was defined as: spike amplitude/noise
amplitude. We calculated spike amplitude by subtracting the minimum voltage
of the mean waveform from its maximum voltage. For estimating noise, we
calculated the mean voltage in a 0.2 ms baseline period, starting 1 ms before
the trough of each spike. We then measured the standard deviation of these
measurements across all 100 spikes for a given neuron. Noise amplitude was
defined as four times this standard deviation. Behavioral Tasks and
{Firing-Rate} {HistogramsThe} behavioral tasks have been described in detail
previously ([Maimon and Assad, 2006a] and [Maimon and Assad, 2006b]), and a
complete description is provided in Supplemental Experimental Procedures. In
brief, monkeys pressed a lever in order to cause a moving dot to reverse its
direction of motion, or, in alternate blocks of trials, the monkey pressed
the lever in reaction to a computer-generated reversal of the dot's
trajectory. Overall, per neuron we analyzed an average of 199 trials in area
5 (min 78; max 554), 217 in {LIP} (min 120; max 458), and 227 in {MT}/{MST}
(min 119; max 495). Only correct trials were analyzed. In generating
peri-event time histograms ({PETHs}), spike trains from {MT}/{MST} and {LIP}
were aligned to the time of the dot's motion reversal, and spike trains from
area 5 were aligned to the time of the arm movement. Our results were not
affected by these differences in how we preprocessed the data (Figure S4).
Multiple {PETHs} were generated for each cell, one for each dot turn or arm
movement in each task. Overall, 98.7\% of {PETHs} had >35 trials and thus
provided a robust estimate of firing rate. Specifically, 1.3\% of {PETHs} had
<35 trials, 61.5\% of {PETHs} had 35--45 trials, 0.9\% of {PETHs} had 45--55
trials, 21.8\% of {PETHs} had 55--65 trials, and 14.6\% of {PETHs} had >65
trials. Gamma Descriptions of Interspike Interval {DistributionsTo} evaluate
the first-order statistics of cortical spike trains, we examined rate-parsed
interspike interval histograms (Figure 3). These were generated by applying
an algorithm developed by Softky and Koch (1993) to each {PETH} and related
set of rasters. For each trial, we uniformly multiplied the overall {PETH} by
a scalar, defined as the number of spikes on the currently analyzed trial
divided by the average number of spikes on a trial. In other words, we
assumed that a cell's firing rate evolved with the same time course on all
trials, but the overall gain changed on any given presentation. Next, we
associated each interspike interval with the trial-specific firing rate at
the central time point of that interval. Finally, we plotted an {ISI}
histogram for intervals associated with 0--5 Hz firing, a separate histogram
for 5--10 Hz firing, and so on, in 5 Hz steps, up to the maximum rate at
which the cell {fired.We} characterized {ISI} histograms by measuring their
coefficient of variation ({CV} = {SD}/mean) and also by fitting them to gamma
distributions. Gamma probability density functions are defined by a shape (k)
and scale (ฮธ) parameter, as {follows:where,We} estimated k and ฮธ using a
nonlinear least-squares procedure (trust-region algorithm) that minimized the
error between the gamma density function and bin values of the {ISI}
histogram (bin width: 1 ms). We excluded the first eight bins of the {ISI}
histogram in calculating the least-squares measure so as to minimize the
effects of the absolute refractory period and bursting on the shape of the
fitted distributions. As we only examined {ISI} histograms associated with
10--40 Hz spiking (see main text), an 8 ms exclusion zone appeared to provide
the best compromise between properly disregarding very short, burst-related
or refractory-related, intervals, but also including enough data to properly
estimate the overall shape of histograms. In early attempts, when we did not
exclude the first 8 ms or when we used a maximum-likelihood method, {ISI}
histograms with burst-related peak followed by a trough and then a second
peak (Figure {6A}) were inappropriately fit with k 1, i.e., as
{exponentials.We} generated a total of 3234 {ISI} histograms associated with
10--40 Hz spiking (539 cells ร 6 histograms per cell). 2424 of these
histograms (75\%) had >150 interspike intervals and thus passed our threshold
for fitting the histogram to a gamma distribution. We were most interested in
k from the gamma fits, as this parameter provided a measure of spike-train
regularity (see main text). Across all fits, k ranged between 0.25 and 21.3.
95\% confidence intervals for k had a mean width of 0.18 ({SD} 0.12),
indicating that our estimates were accurate. Only 1/2424 fits had a
95\\\\\\\\ confidence interval for k wider than 1. We should note that we
view gamma distributions as useful descriptors of {ISI} histograms; we do not
claim any functional significance to the use of this density function as
opposed to log-normal or inverse Gaussian distributions, for example, which
may provide better fits (Barbieri et al., 2001). Future work will need to
determine the process that gives rise to interspike intervals, thus
constraining the proper statistical description of {ISI}
{histograms.Importantly}, for individual neurons, there was a good match
between k and the qualitative appearance of {ISI} regularity in unprocessed
spike trains (Supplemental Data). This correspondence, combined with
extensive analysis of simulated spike trains (data not shown), makes us
confident that the results described are not an artifact of parsing
interspike intervals into separate rate-matched histograms. Furthermore, any
bias in the rate-parsing algorithm should have applied equally across
cortical areas, whereas we observed different results in each area (Figure
{4A}). Note also that spike-count variance-to-mean plots (Figure {3B}, right)
were calculated without reference to {ISI} histograms and thus provide an
independent line of evidence for {ISI} regularity. {AcknowledgmentsWe} thank
{J.H}.R. Maunsell, C. {Padoa-Schioppa}, R. Maimon, and P. Weir for helpful
comments on the manuscript and discussions. K. Irwin, T. Lafratta, M.
{LaFratta}, J. {LeBlanc}, and D. Averbuch provided technical assistance. This
work was supported by {NEI} {EY12106} and the {McKnight} Endowment Fund for
Neuroscience and an {NEI} Vision Core Grant, {EY12196}.%
}
\verb{doi}
\verb 10.1016/j.neuron.2009.03.021
\endverb
\field{number}{3}
\field{pages}{426\bibrangedash 440}
\field{title}{Beyond Poisson: Increased {Spike-Time} Regularity across
Primate Parietal Cortex}
\verb{url}
\verb http://dx.doi.org/10.1016/j.neuron.2009.03.021
\endverb
\field{volume}{62}
\field{journaltitle}{Neuron}
\field{month}{05}
\field{year}{2009}
\endentry
\entry{Berry97}{article}{}
\name{author}{3}{}{%
{{}%
{Berry}{B.}%
{Michael~J.}{M.~J.}%
{}{}%
{}{}}%
{{}%
{Warland}{W.}%
{David~K.}{D.~K.}%
{}{}%
{}{}}%
{{}%
{Meister}{M.}%
{Markus}{M.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{National Academy of Sciences}%
}
\keyw{taouali15}
\strng{namehash}{BMJWDKMM1}
\strng{fullhash}{BMJWDKMM1}
\field{abstract}{%
Assessing the reliability of neuronal spike trains is fundamental to an
understanding of the neural code. We measured the reproducibility of retinal
responses to repeated visual stimuli. In both tiger salamander and rabbit,
the retinal ganglion cells responded to random flicker with discrete, brief
periods of firing. For any given cell, these firing events covered only a
small fraction of the total stimulus time, often less than 5\%. Firing events
were very reproducible from trial to trial: the timing jitter of individual
spikes was as low as 1 msec, and the standard deviation in spike count was
often less than 0.5 spikes. Comparing the precision of spike timing to that
of the spike count showed that the timing of a firing event conveyed several
times more visual information than its spike count. This sparseness and
precision were general characteristics of ganglion cell responses, maintained
over the broad ensemble of stimulus waveforms produced by random flicker, and
over a range of contrasts. Thus, the responses of retinal ganglion cells are
not properly described by a firing probability that varies continuously with
the stimulus. Instead, these neurons elicit discrete firing events that may
be the fundamental coding symbols in retinal spike trains.%
}
\verb{doi}
\verb 10.1073/pnas.94.10.5411
\endverb
\field{number}{10}
\field{pages}{5411\bibrangedash 5416}
\field{title}{The structure and precision of retinal spikeโtrains}
\verb{url}
\verb http://dx.doi.org/10.1073/pnas.94.10.5411
\endverb
\field{volume}{94}
\field{journaltitle}{PNAS}
\field{month}{05}
\field{year}{1997}
\endentry
\entry{Baudot13}{article}{}
\name{author}{6}{}{%
{{}%
{Baudot}{B.}%
{Pierre}{P.}%
{}{}%
{}{}}%
{{}%
{Levy}{L.}%
{Manuel}{M.}%
{}{}%
{}{}}%
{{}%
{Marre}{M.}%
{Olivier}{O.}%
{}{}%
{}{}}%
{{}%
{Monier}{M.}%
{Cyril}{C.}%
{}{}%
{}{}}%
{{}%
{Pananceau}{P.}%
{Marc}{M.}%
{}{}%
{}{}}%
{{}%
{Fr\'{e}gnac}{F.}%
{Yves}{Y.}%
{}{}%
{}{}}%
}
\keyw{area-v1, bicv-sparse, natural-scenes, taouali15}
\strng{namehash}{BP+1}
\strng{fullhash}{BPLMMOMCPMFY1}
\verb{doi}
\verb 10.3389/fncir.2013.00206
\endverb
\field{title}{Animation of natural scene by virtual eye-movements evokes high
precision and low noise in {V1} neurons}
\verb{url}
\verb http://dx.doi.org/10.3389/fncir.2013.00206
\endverb
\field{volume}{7}
\field{journaltitle}{Frontiers in Neural Circuits}
\field{year}{2013}
\endentry
\entry{Haider10}{article}{}
\name{author}{7}{}{%
{{}%
{Haider}{H.}%
{B.}{B.}%
{}{}%
{}{}}%
{{}%
{Krause}{K.}%
{M~R.}{M.~R.}%
{}{}%
{}{}}%
{{}%
{Duque}{D.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Yu}{Y.}%
{Y.}{Y.}%
{}{}%
{}{}}%
{{}%
{Touryan}{T.}%
{J.}{J.}%
{}{}%
{}{}}%
{{}%
{Mazer}{M.}%
{J~A.}{J.~A.}%
{}{}%
{}{}}%
{{}%
{McCormick}{M.}%
{D~A.}{D.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{HB+1}
\strng{fullhash}{HBKMRDAYYTJMJAMDA1}
\field{number}{1}
\field{pages}{107\bibrangedash 121}
\field{title}{Synaptic and network mechanisms of sparse and reliable visual
cortical activity during nonclassical receptive field stimulation}
\field{volume}{65}
\field{journaltitle}{Neuron}
\field{year}{2010}
\endentry
\entry{Borst99}{article}{}
\name{author}{2}{}{%
{{}%
{Borst}{B.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Theunissen}{T.}%
{F~E.}{F.~E.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Nature Publishing Group}%
}
\strng{namehash}{BATFE1}
\strng{fullhash}{BATFE1}
\field{number}{11}
\field{pages}{947\bibrangedash 957}
\field{title}{Information theory and neural coding}
\field{volume}{2}
\field{journaltitle}{Nature {N}euroscience}
\field{year}{1999}
\endentry
\entry{Churchland10}{article}{}
\true{moreauthor}
\name{author}{10}{}{%
{{}%
{Churchland}{C.}%
{M~M.}{M.~M.}%
{}{}%
{}{}}%
{{}%
{Byron}{B.}%
{M~Y.}{M.~Y.}%
{}{}%
{}{}}%
{{}%
{Cunningham}{C.}%
{J~P}{J.~P.}%
{}{}%
{}{}}%
{{}%
{Sugrue}{S.}%
{L~P}{L.~P.}%
{}{}%
{}{}}%
{{}%
{Cohen}{C.}%
{M~R.}{M.~R.}%
{}{}%
{}{}}%
{{}%
{Corrado}{C.}%
{G~S.}{G.~S.}%
{}{}%
{}{}}%
{{}%
{Newsome}{N.}%
{W~T.}{W.~T.}%
{}{}%
{}{}}%
{{}%
{Clark}{C.}%
{A~M.}{A.~M.}%
{}{}%
{}{}}%
{{}%
{Hosseini}{H.}%
{P.}{P.}%
{}{}%
{}{}}%
{{}%
{Scott}{S.}%
{B~B.}{B.~B.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Nature Publishing Group}%
}
\strng{namehash}{CMM+1}
\strng{fullhash}{CMMBMYCJPSLPCMRCGSNWTCAMHPSBB+1}
\field{number}{3}
\field{pages}{369\bibrangedash 378}
\field{title}{Stimulus onset quenches neural variability: a widespread
cortical phenomenon}
\field{volume}{13}
\field{journaltitle}{Nature {N}euroscience}
\field{year}{2010}
\endentry
\entry{Nienborg12}{article}{}
\name{author}{3}{}{%
{{}%
{Nienborg}{N.}%
{H.}{H.}%
{}{}%
{}{}}%
{{}%
{Cohen}{C.}%
{M~R.}{M.~R.}%
{}{}%
{}{}}%
{{}%
{Cumming}{C.}%
{B~G.}{B.~G.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Annual Reviews}%
}
\strng{namehash}{NHCMRCBG1}
\strng{fullhash}{NHCMRCBG1}
\field{pages}{463\bibrangedash 483}
\field{title}{Decision-related activity in sensory neurons: correlations
among neurons and with behavior}
\field{volume}{35}
\field{journaltitle}{Annual review of neuroscience}
\field{year}{2012}
\endentry
\entry{Carandini04}{article}{}
\name{author}{1}{}{%
{{}%
{Carandini}{C.}%
{Matteo}{M.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Public Library of Science}%
}
\keyw{taouali15}
\strng{namehash}{CM1}
\strng{fullhash}{CM1}
\field{abstract}{%
The visual cortex responds to repeated presentations of the same stimulus
with high variability. Because the firing mechanism is remarkably noiseless,
the source of this variability is thought to lie in the membrane potential
fluctuations that result from summated synaptic input. Here this hypothesis
is tested through measurements of membrane potential during visual
stimulation. Surprisingly, trial-to-trial variability of membrane potential
is found to be low. The ratio of variance to mean is much lower for membrane
potential than for firing rate. The high variability of firing rate is
explained by the threshold present in the function that converts inputs into
firing rates. Given an input with small, constant noise, this function
produces a firing rate with a large variance that grows with the mean. This
model is validated on responses recorded both intracellularly and
extracellularly. In neurons of visual cortex, thus, a simple deterministic
mechanism amplifies the low variability of summated synaptic inputs into the
large variability of firing rate. The computational advantages provided by
this amplification are not known.%
}
\verb{doi}
\verb 10.1371/journal.pbio.0020264
\endverb
\field{number}{9}
\field{pages}{e264+}
\field{title}{Amplification of {Trial-to-Trial} Response Variability by
Neurons in Visual Cortex}
\verb{url}
\verb http://dx.doi.org/10.1371/journal.pbio.0020264
\endverb
\field{volume}{2}
\list{location}{1}{%
{Smith-Kettlewell Eye Research Institute, San Francisco, California, USA.
matteo@ski.org}%
}
\field{journaltitle}{PLoS Biol}
\field{month}{08}
\field{year}{2004}
\endentry
\entry{Mainen95}{article}{}
\name{author}{2}{}{%
{{}%
{Mainen}{M.}%
{Z.~F.}{Z.~F.}%
{}{}%
{}{}}%
{{}%
{Sejnowski}{S.}%
{Terrence~J.}{T.~J.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{American Association for the Advancement of Science}%
}
\keyw{taouali15}
\strng{namehash}{MZFSTJ1}
\strng{fullhash}{MZFSTJ1}
\field{abstract}{%
It is not known whether the variability of neural activity in the cerebral
cortex carries information or reflects noisy underlying mechanisms. In an
examination of the reliability of spike generation using recordings from
neurons in rat neocortical slices, the precision of spike timing was found to
depend on stimulus transients. Constant stimuli led to imprecise spike
trains, whereas stimuli with fluctuations resembling synaptic activity
produced spike trains with timing reproducible to less than 1 millisecond.
These data suggest a low intrinsic noise level in spike generation, which
could allow cortical neurons to accurately transform synaptic input into
spike sequences, supporting a possible role for spike timing in the
processing of cortical information by the neocortex.%
}
\verb{doi}
\verb 10.1126/science.7770778
\endverb
\field{number}{5216}
\field{pages}{1503\bibrangedash 1506}
\field{title}{Reliability of spike timing in neocortical neurons.}
\verb{url}
\verb http://dx.doi.org/10.1126/science.7770778
\endverb
\field{volume}{268}
\list{location}{1}{%
{Howard Hughes Medical Institute, Salk Institute for Biological Studies, La
Jolla, CA 92037, USA.}%
}
\field{journaltitle}{Science}
\field{month}{06}
\field{year}{1995}
\endentry
\entry{Schneidman98}{article}{}
\name{author}{3}{}{%
{{}%
{Schneidman}{S.}%
{Elad}{E.}%
{}{}%
{}{}}%
{{}%
{Freedman}{F.}%
{B.}{B.}%
{}{}%
{}{}}%
{{}%
{Segev}{S.}%
{I.}{I.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{SEFBSI1}
\strng{fullhash}{SEFBSI1}
\field{abstract}{%
This memory is embedded in the distribution of channel states in the spike
initiation site. The nature and resolution of this memory depend on the size
of the channel pool and on the kinetics and number of states of the channels.
We hypothesize that the number of channels in the spike initiation zone may
be optimized in some sense to give the reliability and accuracy discussed
above, together with a short-term memory of the neuron's activity. In this
context, it is interesting to mention the work of Marder, Abbott, Turrigiano,
Liu, and Golowasch (1996) and Abbott et al. (1996), which demonstrates
activity-dependent long-term changes in the properties of intrinsic membrane
currents.%
}
\field{title}{Ion Channel Stochasticity May Be Critical in Determining the
Reliability and Precision of Spike Timing}
\verb{url}
\verb http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.29.6082
\endverb
\field{journaltitle}{Neural computation}
\field{year}{1998}
\endentry
\entry{Arieli96}{article}{}
\name{author}{4}{}{%
{{}%
{Arieli}{A.}%
{Amos}{A.}%
{}{}%
{}{}}%
{{}%
{Sterkin}{S.}%
{Alexander}{A.}%
{}{}%
{}{}}%
{{}%
{Grinvald}{G.}%
{Amiram}{A.}%
{}{}%
{}{}}%
{{}%
{Aertsen}{A.}%
{Ad}{A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{American Association for the Advancement of Science}%
}
\keyw{taouali15}
\strng{namehash}{AASAGAAA1}
\strng{fullhash}{AASAGAAA1}
\field{abstract}{%
Evoked activity in the mammalian cortex and the resulting behavioral
responses exhibit a large variability to repeated presentations of the same
stimulus. This study examined whether the variability can be attributed to
ongoing activity. Ongoing and evoked spatiotemporal activity patterns in the
cat visual cortex were measured with real-time optical imaging; local field
potentials and discharges of single neurons were recorded simultaneously, by
electrophysiological techniques. The evoked activity appeared deterministic,
and the variability resulted from the dynamics of ongoing activity,
presumably reflecting the instantaneous state of cortical networks. In spite
of the large variability, evoked responses in single trials could be
predicted by linear summation of the deterministic response and the preceding
ongoing activity. Ongoing activity must play an important role in cortical
function and cannot be ignored in exploration of cognitive processes.%
}
\verb{doi}
\verb 10.1126/science.273.5283.1868
\endverb
\field{number}{5283}
\field{pages}{1868\bibrangedash 1871}
\field{title}{Dynamics of Ongoing Activity: Explanation of the Large
Variability in Evoked Cortical Responses}
\verb{url}
\verb http://dx.doi.org/10.1126/science.273.5283.1868
\endverb
\field{volume}{273}
\list{location}{1}{%
{Department of Neurobiology, Weizmann Institute of Science, Post Office Box
26, Rehovot 76100, Israel.}%
}
\field{journaltitle}{Science}
\field{month}{09}
\field{year}{1996}
\endentry
\entry{Tsodyks99}{article}{}
\name{author}{4}{}{%
{{}%
{Tsodyks}{T.}%
{Misha}{M.}%
{}{}%
{}{}}%
{{}%
{Kenet}{K.}%
{Tal}{T.}%
{}{}%
{}{}}%
{{}%
{Grinvald}{G.}%
{Amiram}{A.}%
{}{}%
{}{}}%
{{}%
{Arieli}{A.}%
{A.}{A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{American Association for the Advancement of Science}%
}
\keyw{taouali15}
\strng{namehash}{TMKTGAAA1}
\strng{fullhash}{TMKTGAAA1}
\field{abstract}{%
The relation between the activity of a single neocortical neuron and the
dynamics of the network in which it is embedded was explored by single-unit
recordings and real-time optical imaging. The firing rate of a spontaneously
active single neuron strongly depends on the instantaneous spatial pattern of
ongoing population activity in a large cortical area. Very similar spatial
patterns of population activity were observed both when the neuron fired
spontaneously and when it was driven by its optimal stimulus. The evoked
patterns could be used to reconstruct the spontaneous activity of single
neurons.%
}
\verb{doi}
\verb 10.1126/science.286.5446.1943
\endverb
\field{number}{5446}
\field{pages}{1943\bibrangedash 1946}
\field{title}{Linking Spontaneous Activity of Single Cortical Neurons and the
Underlying Functional Architecture}
\verb{url}
\verb http://dx.doi.org/10.1126/science.286.5446.1943
\endverb
\field{volume}{286}
\field{journaltitle}{Science}
\field{month}{12}
\field{year}{1999}
\endentry
\entry{Kenet03}{article}{}
\name{author}{5}{}{%
{{}%
{Kenet}{K.}%
{Tal}{T.}%
{}{}%
{}{}}%
{{}%
{Bibitchkov}{B.}%
{Dmitri}{D.}%
{}{}%
{}{}}%
{{}%
{Tsodyks}{T.}%
{Misha}{M.}%
{}{}%
{}{}}%
{{}%
{Grinvald}{G.}%
{Amiram}{A.}%
{}{}%
{}{}}%
{{}%
{Arieli}{A.}%
{Amos}{A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Nature Publishing Group}%
}
\keyw{taouali15}
\strng{namehash}{KTBDTMGAAA1}
\strng{fullhash}{KTBDTMGAAA1}
\field{abstract}{%
Spontaneous cortical activity--ongoing activity in the absence of intentional
sensory input--has been studied extensively, using methods ranging from {EEG}
(electroencephalography), through voltage sensitive dye imaging, down to
recordings from single neurons. Ongoing cortical activity has been shown to
play a critical role in development, and must also be essential for
processing sensory perception, because it modulates stimulus-evoked activity,
and is correlated with behaviour. Yet its role in the processing of external
information and its relationship to internal representations of sensory
attributes remains unknown. Using voltage sensitive dye imaging, we
previously established a close link between ongoing activity in the visual
cortex of anaesthetized cats and the spontaneous firing of a single neuron.
Here we report that such activity encompasses a set of dynamically switching
cortical states, many of which correspond closely to orientation maps. When
such an orientation state emerged spontaneously, it spanned several
hypercolumns and was often followed by a state corresponding to a proximal
orientation. We suggest that dynamically switching cortical states could
represent the brain's internal context, and therefore reflect or influence
memory, perception and behaviour.%
}
\verb{doi}
\verb 10.1038/nature02078
\endverb
\field{number}{6961}
\field{pages}{954\bibrangedash 956}
\field{title}{Spontaneously emerging cortical representations of visual
attributes.}
\verb{url}
\verb http://dx.doi.org/10.1038/nature02078
\endverb
\field{volume}{425}
\list{location}{1}{%
{Department of Neurobiology, The Weizmann Institute of Science, Rehovot
76100, Israel. (tal@phy.ucsf.edu}%
}
\field{journaltitle}{Nature}
\field{month}{10}
\field{year}{2003}
\endentry
\entry{Ress03}{article}{}
\name{author}{2}{}{%
{{}%
{Ress}{R.}%
{David}{D.}%
{}{}%
{}{}}%
{{}%
{Heeger}{H.}%
{David~J.}{D.~J.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Nature Publishing Group}%
}
\keyw{taouali15}
\strng{namehash}{RDHDJ1}
\strng{fullhash}{RDHDJ1}
\field{abstract}{%
We used functional magnetic resonance imaging ({fMRI}) to measure activity in
human early visual cortex (areas V1, V2 and V3) during a challenging
contrast-detection task. Subjects attempted to detect the presence of slight
contrast increments added to two kinds of background patterns. Behavioral
responses were recorded so that the corresponding cortical activity could be
grouped into the usual signal detection categories: hits, false alarms,
misses and correct rejects. For both kinds of background patterns, the
measured cortical activity was retinotopically specific. Hits and false
alarms were associated with significantly more cortical activity than were
correct rejects and misses. That false alarms evoked more activity than
misses indicates that activity in early visual cortex corresponded to the
subjects' percepts, rather than to the physically presented stimulus.%
}
\verb{doi}
\verb 10.1038/nn1024
\endverb
\field{number}{4}
\field{pages}{414\bibrangedash 420}
\field{title}{Neuronal correlates of perception in early visual cortex.}
\verb{url}
\verb http://dx.doi.org/10.1038/nn1024
\endverb
\field{volume}{6}
\field{journaltitle}{Nature {N}euroscience}
\field{month}{04}
\field{year}{2003}
\endentry
\entry{Gur97}{article}{}
\name{author}{3}{}{%
{{}%
{Gur}{G.}%
{M.}{M.}%
{}{}%
{}{}}%
{{}%
{Beylin}{B.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Snodderly}{S.}%
{D.~Max}{D.~M.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{GMBASDM1}
\strng{fullhash}{GMBASDM1}
\field{abstract}{%
Response variability of neurons limits the reliability and resolution of
sensory systems. It is generally thought that response variability in the
visual system increases at cortical levels, but the causes of the variability
have not been identified. We have measured the response variability of
neurons in primary visual cortex (V1) of alert monkeys. We recorded from 80
single cells distributed over all V1 layers and from 8 parvocellular cells of
the lateral geniculate nucleus. All cells were stimulated with a bar of
near-optimal orientation, color, and dimensions while continuously monitoring
the eye movements of fixation. To minimize the effects of eye movements,
responses that occurred while the eye was relatively steady were selected for
analysis. The impulses elicited by each stimulus presentation were counted,
and the variance and coefficient of variation were computed. Both measures of
response variability were much lower than reported previously for V1 cells of
both alert and anesthetized monkeys. Our data show that fixational eye
movements cause a large component of response variance in alert monkeys.
Moreover, the reliability of V1 neurons is not obviously degraded compared
with lateral geniculate nucleus cells. The high reliability of neurons in
alert monkeys is consistent with expectations from conventional biophysical
models, and it suggests that activity in a modest number of neurons may
suffice to form a perceptual decision.%
}
\field{number}{8}
\field{pages}{2914\bibrangedash 2920}
\field{title}{Response variability of neurons in primary visual cortex (V1)
of alert monkeys.}
\verb{url}
\verb http://www.jneurosci.org/cgi/content/abstract/17/8/2914
\endverb
\field{volume}{17}
\field{journaltitle}{The Journal of Neuroscience}
\field{month}{04}
\field{year}{1997}
\endentry
\entry{Churchland11}{article}{}
\name{author}{6}{}{%
{{}%
{Churchland}{C.}%
{A~K.}{A.~K.}%
{}{}%
{}{}}%
{{}%
{Kiani}{K.}%
{R.}{R.}%
{}{}%
{}{}}%
{{}%
{Chaudhuri}{C.}%
{R.}{R.}%
{}{}%
{}{}}%
{{}%
{Wang}{W.}%
{X.}{X.}%
{}{}%
{}{}}%
{{}%
{Pouget}{P.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Shadlen}{S.}%
{M~N.}{M.~N.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{CAK+1}
\strng{fullhash}{CAKKRCRWXPASMN1}
\field{number}{4}
\field{pages}{818\bibrangedash 831}
\field{title}{Variance as a signature of neural computations during decision
making}
\field{volume}{69}
\field{journaltitle}{Neuron}
\field{year}{2011}
\endentry
\entry{Ponce13}{article}{}
\name{author}{5}{}{%
{{}%
{{Ponce-Alvarez}}{P.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Thiele}{T.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Albright}{A.}%
{T~D}{T.~D.}%
{}{}%
{}{}}%
{{}%
{Stoner}{S.}%
{G~R.}{G.~R.}%
{}{}%
{}{}}%
{{}%
{Deco}{D.}%
{G.}{G.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{National Acad Sciences}%
}
\strng{namehash}{PATAATDSGRDG1}
\strng{fullhash}{PATAATDSGRDG1}
\field{number}{32}
\field{pages}{13162\bibrangedash 13167}
\field{title}{Stimulus{-}dependent variability and noise correlations in
cortical MT neurons}
\field{volume}{110}
\field{journaltitle}{PNAS}
\field{year}{2013}
\endentry
\entry{Ponce10}{article}{}
\name{author}{3}{}{%
{{}%
{{Ponce-Alvarez}}{P.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Kilavik}{K.}%
{B~E.}{B.~E.}%
{}{}%
{}{}}%
{{}%
{Riehle}{R.}%
{A.}{A.}%
{}{}%
{}{}}%
}
\strng{namehash}{PAKBERA1}
\strng{fullhash}{PAKBERA1}
\field{number}{1-2}
\field{pages}{351\bibrangedash 365}
\field{title}{Comparison of local measures of spike time irregularity and
relating variability to firing rate in motor cortical neurons}
\field{volume}{29}
\field{journaltitle}{Journal of Computational Neuroscience}
\field{year}{2010}
\endentry
\entry{Goris14}{article}{}
\name{author}{3}{}{%
{{}%
{Goris}{G.}%
{Robbe L.~T.}{R.~L.~T.}%
{}{}%
{}{}}%
{{}%
{Movshon}{M.}%
{J.~Anthony}{J.~A.}%
{}{}%
{}{}}%
{{}%
{Simoncelli}{S.}%
{Eero~P.}{E.~P.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Nature Publishing Group, a division of Macmillan Publishers Limited. All
Rights Reserved.}%
}
\keyw{taouali15}
\strng{namehash}{GRLTMJASEP1}
\strng{fullhash}{GRLTMJASEP1}
\verb{doi}
\verb 10.1038/nn.3711
\endverb
\field{number}{6}
\field{pages}{858\bibrangedash 865}
\field{title}{Partitioning neuronal variability}
\verb{url}
\verb http://www.cns.nyu.edu/pub/lcv/goris13-preprint.pdf
\endverb
\field{volume}{17}
\field{journaltitle}{Nature {N}euroscience}
\field{month}{06}
\field{year}{2014}
\endentry
\entry{Consul73}{article}{}
\name{author}{2}{}{%
{{}%
{Consul}{C.}%
{P~C.}{P.~C.}%
{}{}%
{}{}}%
{{}%
{Jain}{J.}%
{G~C.}{G.~C.}%
{}{}%
{}{}}%
}
\list{publisher}{2}{%
{Taylor}%
{Francis Group}%
}
\strng{namehash}{CPCJGC1}
\strng{fullhash}{CPCJGC1}
\field{number}{4}
\field{pages}{791\bibrangedash 799}
\field{title}{A generalization of the Poisson distribution}
\field{volume}{15}
\field{journaltitle}{Technometrics}
\field{year}{1973}
\endentry
\entry{Lambert92}{article}{}
\name{author}{1}{}{%
{{}%
{Lambert}{L.}%
{D.}{D.}%
{}{}%
{}{}}%
}
\list{publisher}{2}{%
{Taylor}%
{Francis}%
}
\strng{namehash}{LD1}
\strng{fullhash}{LD1}
\field{number}{1}
\field{pages}{1\bibrangedash 14}
\field{title}{Zero{-}inflated Poisson regression, with an application to
defects in manufacturing}
\field{volume}{34}
\field{journaltitle}{Technometrics}
\field{year}{1992}
\endentry
\entry{Miaou94}{article}{}
\name{author}{1}{}{%
{{}%
{Miaou}{M.}%
{S.}{S.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{MS1}
\strng{fullhash}{MS1}
\field{number}{4}
\field{pages}{471\bibrangedash 482}
\field{title}{The relationship between truck accidents and geometric design
of road sections: Poisson versus negative binomial regressions}
\field{volume}{26}
\field{journaltitle}{Accident Analysis and Prevention}
\field{year}{1994}
\endentry
\entry{Shankar97}{article}{}
\name{author}{3}{}{%
{{}%
{Shankar}{S.}%
{V.}{V.}%
{}{}%
{}{}}%
{{}%
{Milton}{M.}%
{J.}{J.}%
{}{}%
{}{}}%
{{}%
{Mannering}{M.}%
{F.}{F.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{SVMJMF1}
\strng{fullhash}{SVMJMF1}
\field{number}{6}
\field{pages}{829\bibrangedash 837}
\field{title}{Modeling accident frequencies as zero-altered probability
processes: an empirical inquiry}
\field{volume}{29}
\field{journaltitle}{Accident Analysis and Prevention}
\field{year}{1997}
\endentry
\entry{Mccullagh89}{book}{}
\name{author}{2}{}{%
{{}%
{McCullagh}{M.}%
{P.}{P.}%
{}{}%
{}{}}%
{{}%
{Nelder}{N.}%
{J~A.}{J.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{2}{%
{London England Chapman}%
{Hall 1983.}%
}
\strng{namehash}{MPNJA1}
\strng{fullhash}{MPNJA1}
\field{title}{Generalized linear models}
\field{year}{1989}
\endentry
\entry{Wedderburn74}{article}{}
\name{author}{1}{}{%
{{}%
{Wedderburn}{W.}%
{R~W~M.}{R.~W.~M.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Biometrika Trust}%
}
\strng{namehash}{WRWM1}
\strng{fullhash}{WRWM1}
\field{number}{3}
\field{pages}{439\bibrangedash 447}
\field{title}{Quasi likelihood functions, generalized linear models, and the
GaussNewton method}
\field{volume}{61}
\field{journaltitle}{Biometrika}
\field{year}{1974}
\endentry
\entry{Byers03}{article}{}
\name{author}{4}{}{%
{{}%
{Byers}{B.}%
{A~L.}{A.~L.}%
{}{}%
{}{}}%
{{}%
{Allore}{A.}%
{H.}{H.}%
{}{}%
{}{}}%
{{}%
{Gill}{G.}%
{T~M.}{T.~M.}%
{}{}%
{}{}}%
{{}%
{Peduzzi}{P.}%
{P~N.}{P.~N.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{BALAHGTMPPN1}
\strng{fullhash}{BALAHGTMPPN1}
\field{number}{6}
\field{pages}{559\bibrangedash 564}
\field{title}{Application of negative binomial modeling for discrete
outcomes: a case study in aging research}
\field{volume}{56}
\field{journaltitle}{Journal of clinical epidemiology}
\field{year}{2003}
\endentry
\entry{Poch96}{article}{}
\name{author}{2}{}{%
{{}%
{Poch}{P.}%
{M.}{M.}%
{}{}%
{}{}}%
{{}%
{Mannering}{M.}%
{F.}{F.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{American Society of Civil Engineers}%
}
\strng{namehash}{PMMF1}
\strng{fullhash}{PMMF1}
\field{number}{2}
\field{pages}{105\bibrangedash 113}
\field{title}{Negative binomial analysis of intersection accident
frequencies}
\field{volume}{122}
\field{journaltitle}{Journal of Transportation Engineering}
\field{year}{1996}
\endentry
\entry{Anscombe50}{article}{}
\name{author}{1}{}{%
{{}%
{Anscombe}{A.}%
{F~J}{F.~J.}%
{}{}%
{}{}}%
}
\strng{namehash}{AFJ1}
\strng{fullhash}{AFJ1}
\field{pages}{358\bibrangedash 382}
\field{title}{Sampling theory of the negative binomial and logarithmic series
distributions}
\field{journaltitle}{Biometrika}
\field{year}{1950}
\endentry
\entry{Bliss53}{article}{}
\name{author}{2}{}{%
{{}%
{Bliss}{B.}%
{C~I.}{C.~I.}%
{}{}%
{}{}}%
{{}%
{Fisher}{F.}%
{R~A.}{R.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{JSTOR}%
}
\strng{namehash}{BCIFRA1}
\strng{fullhash}{BCIFRA1}
\field{number}{2}
\field{pages}{176\bibrangedash 200}
\field{title}{Fitting the negative binomial distribution to biological data}
\field{volume}{9}
\field{journaltitle}{Biometrics}
\field{year}{1953}
\endentry
\entry{Scott12}{article}{}
\name{author}{2}{}{%
{{}%
{Scott}{S.}%
{James}{J.}%
{}{}%
{}{}}%
{{}%
{Pillow}{P.}%
{Jonathan~W}{J.~W.}%
{}{}%
{}{}}%
}
\strng{namehash}{SJPJW1}
\strng{fullhash}{SJPJW1}
\field{pages}{1898\bibrangedash 1906}
\field{title}{Fully Bayesian inference for neural models with
negative-binomial spiking}
\field{journaltitle}{Advances in neural information processing systems}
\field{year}{2012}
\endentry
\entry{Polson12}{article}{}
\true{moreauthor}
\name{author}{2}{}{%
{{}%
{Polson}{P.}%
{Nicholas~G}{N.~G.}%
{}{}%
{}{}}%
{{}%
{Scott}{S.}%
{James~G}{J.~G.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{International Society for Bayesian Analysis}%
}
\strng{namehash}{PNGSJG+1}
\strng{fullhash}{PNGSJG+1}
\field{number}{4}
\field{pages}{887\bibrangedash 902}
\field{title}{On the half-Cauchy prior for a global scale parameter}
\field{volume}{7}
\field{journaltitle}{Bayesian Analysis}
\field{year}{2012}
\endentry
\entry{Nawrot08}{article}{}
\name{author}{6}{}{%
{{}%
{Nawrot}{N.}%
{Martin~P.}{M.~P.}%
{}{}%
{}{}}%
{{}%
{Boucsein}{B.}%
{Clemens}{C.}%
{}{}%
{}{}}%
{{}%
{Rodriguez~Molina}{R.~M.}%
{Victor}{V.}%
{}{}%
{}{}}%
{{}%
{Riehle}{R.}%
{Alexa}{A.}%
{}{}%
{}{}}%
{{}%
{Aertsen}{A.}%
{Ad}{A.}%
{}{}%
{}{}}%
{{}%
{Rotter}{R.}%
{Stefan}{S.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{NMP+1}
\strng{fullhash}{NMPBCRMVRAAARS1}
\field{abstract}{%
We propose a method for the time-resolved joint analysis of two related
aspects of single neuron variability, the spiking irregularity measured by
the squared coefficient of variation ({CV}(2)) of the {ISIs} and the
trial-by-trial variability of the spike count measured by the Fano factor
({FF}). We provide a calibration of both estimators using the theory of
renewal processes, and verify it for spike trains recorded in vitro. Both
estimators exhibit a considerable bias for short observations that count less
than about 5-10 spikes on average. The practical difficulty of measuring the
{CV}(2) in rate modulated data can be overcome by a simple procedure of spike
train demodulation which was tested in numerical simulations and in real
spike trains. We propose to test neuronal spike trains for deviations from
the null-hypothesis {FF}={CV}(2). We show that cortical pyramidal neurons,
recorded under controlled stationary input conditions in vitro, comply with
this assumption. Performing a time-resolved joint analysis of {CV}(2) and
{FF} of a single unit recording from the motor cortex of a behaving monkey we
demonstrate how the dynamic change of their quantitative relation can be
interpreted with respect to neuron intrinsic and extrinsic factors that
influence cortical variability in vivo. Finally, we discuss the effect of
several additional factors such as serial interval correlation and refractory
period on the empiric relation of {FF} and {CV}(2).%
}
\verb{doi}
\verb 10.1016/j.jneumeth.2007.10.013
\endverb
\field{number}{2}
\field{pages}{374\bibrangedash 390}
\field{title}{Measurement of variability dynamics in cortical spike trains}
\verb{url}
\verb http://dx.doi.org/10.1016/j.jneumeth.2007.10.013
\endverb
\field{volume}{169}
\list{location}{1}{%
{Bernstein Center for Computational Neuroscience Freiburg, Germany;
Neurobiology \& Biophysics, Faculty of Biology, University of Freiburg,
Germany.}%
}
\field{journaltitle}{Journal of Neuroscience Methods}
\field{month}{04}
\field{year}{2008}
\endentry
\entry{Atkinson82}{article}{}
\name{author}{1}{}{%
{{}%
{Atkinson}{A.}%
{A~C}{A.~C.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{JSTOR}%
}
\strng{namehash}{AAC1}
\strng{fullhash}{AAC1}
\field{pages}{1\bibrangedash 36}
\field{title}{Regression diagnostics, transformations and constructed
variables}
\field{journaltitle}{Journal of the Royal Statistical Society. Series B
(Methodological)}
\field{year}{1982}
\endentry
\entry{Treves91}{article}{}
\name{author}{2}{}{%
{{}%
{Treves}{T.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Rolls}{R.}%
{E~T.}{E.~T.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Informa UK Ltd UK}%
}
\strng{namehash}{TARET1}
\strng{fullhash}{TARET1}
\field{number}{4}
\field{pages}{371\bibrangedash 397}
\field{title}{What determines the capacity of autoassociative memories in the
brain ?}
\field{volume}{2}
\field{journaltitle}{Network: Computation in Neural Systems}
\field{year}{1991}
\endentry
\entry{Levy96}{article}{}
\name{author}{2}{}{%
{{}%
{Levy}{L.}%
{W~B}{W.~B.}%
{}{}%
{}{}}%
{{}%
{Baxter}{B.}%
{R~A.}{R.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{MIT Press}%
}
\strng{namehash}{LWBBRA1}
\strng{fullhash}{LWBBRA1}
\field{number}{3}
\field{pages}{531\bibrangedash 543}
\field{title}{Energy efficient neural codes}
\field{volume}{8}
\field{journaltitle}{Neural Computation}
\field{year}{1996}
\endentry
\entry{Baddeley96}{article}{}
\name{author}{1}{}{%
{{}%
{Baddeley}{B.}%
{R.}{R.}%
{}{}%
{}{}}%
}
\strng{namehash}{BR1}
\strng{fullhash}{BR1}
\field{pages}{560\bibrangedash 561}
\field{title}{An efficient code in V1?}
\field{volume}{381}
\field{journaltitle}{Nature}
\field{year}{1996}
\endentry
\entry{Shannon48}{article}{}
\name{author}{1}{}{%
{{}%
{Shannon}{S.}%
{C~E.}{C.~E.}%
{}{}%
{}{}}%
}
\keyw{communication entropy information shannon toread}
\strng{namehash}{SCE1}
\strng{fullhash}{SCE1}
\field{pages}{379\bibrangedash 423, 623\bibrangedash 656}
\field{title}{A Mathematical Theory of Communication}
\field{volume}{27}
\field{journaltitle}{The Bell System Technical Journal}
\field{year}{1948}
\warn{\item Invalid format of field 'month'}
\endentry
\entry{Simoncelli98}{article}{}
\name{author}{2}{}{%
{{}%
{Simoncelli}{S.}%
{E~P.}{E.~P.}%
{}{}%
{}{}}%
{{}%
{Heeger}{H.}%
{D~J.}{D.~J.}%
{}{}%
{}{}}%
}
\keyw{Time Factors}
\strng{namehash}{SEPHDJ1}
\strng{fullhash}{SEPHDJ1}
\field{abstract}{%
Electrophysiological studies indicate that neurons in the middle temporal
(MT) area of the primate brain are selective for the velocity of visual
stimuli. This paper describes a computational model of MT physiology, in
which local image velocities are represented via the distribution of MT
neuronal responses. The computation is performed in two stages, corresponding
to neurons in cortical areas V1 and MT. Each stage computes a weighted linear
sum of inputs, followed by rectification and divisive normalization. V1
receptive field weights are designed for orientation and direction
selectivity. MT receptive field weights are designed for velocity (both speed
and direction) selectivity. The paper includes computational simulations
accounting for a wide range of physiological data, and describes experiments
that could be used to further test and refine the model.%
}
\field{issn}{0042-6989}
\field{number}{5}
\field{pages}{743\bibrangedash 761}
\field{title}{A model of neuronal responses in visual area MT.}
\field{volume}{38}
\field{journaltitle}{Vision {R}esearch}
\field{year}{1998}
\warn{\item Invalid format of field 'month'}
\endentry
\entry{Rust06}{article}{}
\name{author}{4}{}{%
{{}%
{Rust}{R.}%
{N~C.}{N.~C.}%
{}{}%
{}{}}%
{{}%
{Mante}{M.}%
{V.}{V.}%
{}{}%
{}{}}%
{{}%
{Simoncelli}{S.}%
{E~P.}{E.~P.}%
{}{}%
{}{}}%
{{}%
{Movshon}{M.}%
{J~A.}{J.~A.}%
{}{}%
{}{}}%
}
\list{language}{1}{%
{eng}%
}
\keyw{Visual Cortex}
\strng{namehash}{RNCMVSEPMJA1}
\strng{fullhash}{RNCMVSEPMJA1}
\field{abstract}{%
Neurons in area MT (V5) are selective for the direction of visual motion. In
addition, many are selective for the motion of complex patterns independent
of the orientation of their components, a behavior not seen in earlier visual
areas. We show that the responses of MT cells can be captured by a
linear-nonlinear model that operates not on the visual stimulus, but on the
afferent responses of a population of nonlinear V1 cells. We fit this cascade
model to responses of individual MT neurons and show that it robustly
predicts the separately measured responses to gratings and plaids. The model
captures the full range of pattern motion selectivity found in MT. Cells that
signal pattern motion are distinguished by having convergent excitatory input
from V1 cells with a wide range of preferred directions, strong motion
opponent suppression and a tuned normalization that may reflect suppressive
input from the surround of V1 cells.%
}
\verb{doi}
\verb 10.1038/nn1786
\endverb
\field{issn}{1097-6256}
\field{number}{11}
\field{pages}{1421\bibrangedash 1431}
\field{title}{How MT cells analyze the motion of visual patterns.}
\verb{url}
\verb http://www.ncbi.nlm.nih.gov/pubmed/17041595
\endverb
\field{volume}{9}
\field{journaltitle}{Nature {N}euroscience}
\field{year}{2006}
\warn{\item Invalid format of field 'month'}
\endentry
\entry{Wang12}{article}{}
\name{author}{3}{}{%
{{}%
{Wang}{W.}%
{H~X.}{H.~X.}%
{}{}%
{}{}}%
{{}%
{Heeger}{H.}%
{D~J.}{D.~J.}%
{}{}%
{}{}}%
{{}%
{Landy}{L.}%
{M~S.}{M.~S.}%
{}{}%
{}{}}%
}
\keyw{Sensory Thresholds}
\strng{namehash}{WHXHDJLMS1}
\strng{fullhash}{WHXHDJLMS1}
\field{abstract}{%
First-order (contrast) surround suppression has been well characterized both
psychophysically and physiologically,but relatively little is known as to
whether the perception of second-order visual stimuli exhibits analogous
center--surround interactions. Second-order surround suppression was
characterized by requiring subjects to detect second-order modulation in
stimuli presented alone or embedded in a surround.Both contrast- (CM) and
orientation-modulated (OM) stimuli were used. For most subjects and both OM
and CM stimuli, second-order surrounds caused thresholds to be higher,
indicative of second-order suppression. For CM stimuli, suppression was
orientation-specific, i.e., higher thresholds for parallel than for
orthogonal surrounds. However, the evidence for orientation specificity of
suppression for OM stimuli was weaker. These results suggest that
normalization, leading to surround suppression, operates at multiple stages
in cortical processing.%
}
\field{issn}{1878-5646}
\field{pages}{192\bibrangedash 200}
\field{title}{Responses to second order texture modulations undergo surround
suppression.}
\field{volume}{62}
\field{journaltitle}{Vision {R}esearch}
\field{year}{2012}
\warn{\item Invalid format of field 'month'}
\endentry
\entry{Voges10neurocomp}{article}{}
\name{author}{2}{}{%
{{}%
{Voges}{V.}%
{N.}{N.}%
{}{}%
{}{}}%
{{}%
{Perrinet}{P.}%
{L.}{L.}%
{}{}%
{}{}}%
}
\keyw{area-v1, association\_field, assofield, decoding, perrinet11sfn}
\strng{namehash}{VNPL1}
\strng{fullhash}{VNPL1}
\field{abstract}{%
We study cortical network dynamics for a spatially embedded network model. It
represents, in terms of spatial scale, a large piece of cortex allowing for
long-range connections, resulting in a rather sparse connectivity. The
spatial embedding also permits us to include distance-dependent conduction
delays. We use two different types of conductance-based {I\&F} neurons as
excitatory and inhibitory units, as well as specific connection
probabilities. In order to remain computationally tractable, we reduce neuron
density, modelling part of the missing internal input via external poissonian
spike trains. Compared to previous studies, we observe significant changes in
the dynamical phase space: Altered activity patterns require another
regularity measures than the coefficient of variation. Hence, we compare
three different regularity measure on the basis of artificial
inter-spike-interval distributions. We identify two types of mixed states,
where different phases coexist in certain regions of the phase space. More
notably, our boundary between high and low activity states depends
predominantly on the relation between excitatory and inhibitory synaptic
strength instead of the input rate.%
}
\verb{doi}
\verb 10.1016/j.jphysparis.2009.11.004
\endverb
\field{issn}{0928-4257}
\field{number}{1-2}
\field{pages}{51\bibrangedash 60}
\field{title}{Phase space analysis of networks based on biologically
realistic parameters}
\verb{url}
\verb http://dx.doi.org/10.1016/j.jphysparis.2009.11.004
\endverb
\field{volume}{104}
\field{journaltitle}{Journal of Physiology-Paris}
\field{month}{01}
\field{year}{2010}
\endentry
\entry{Voges12}{article}{}
\name{author}{2}{}{%
{{}%
{Voges}{V.}%
{N.}{N.}%
{}{}%
{}{}}%
{{}%
{Perrinet}{P.}%
{L.}{L.}%
{}{}%
{}{}}%
}
\keyw{dynamical-model, kaplan13, lateral\_connections, patch-based,
recurrent\_neural\_networks}
\strng{namehash}{VNPL1}
\strng{fullhash}{VNPL1}
\field{abstract}{%
Most studies on the dynamics of recurrent cortical networks are either based
on purely random wiring or neighborhood couplings. Neuronal cortical
connectivity, however, shows a complex spatial pattern composed of local and
remote patchy connections. We ask to what extent such geometric traits
influence the '' idle'' dynamics of two-dimensional (2d) cortical network
models composed of conductance-based integrate-and-fire (iaf) neurons. In
contrast to the typical 1 mm2 used in most studies, we employ an enlarged
spatial set-up of 25 mm2 to provide for long-range connections. Our models
range from purely random to distance-dependent connectivities including
patchy projections, i.e., spatially clustered synapses. Analyzing the
characteristic measures for synchronicity and regularity in neuronal spiking,
we explore and compare the phase spaces and activity patterns of our
simulation results. Depending on the input parameters, different dynamical
states appear, similar to the known synchronous regular '' {SR}'' or
asynchronous irregular '' {AI}'' firing in random networks. Our structured
networks, however, exhibit shifted and sharper transitions, as well as more
complex activity patterns. Distance-dependent connectivity structures induce
a spatio-temporal spread of activity, e.g., propagating waves, that random
networks cannot account for. Spatially and temporally restricted activity
injections reveal that a high amount of local coupling induces rather
unstable {AI} dynamics. We find that the amount of local versus long-range
connections is an important parameter, whereas the structurally advantageous
wiring cost optimization of patchy networks has little bearing on the phase
space.%
}
\verb{doi}
\verb 10.3389/fncom.2012.00041
\endverb
\field{issn}{1662-5188}
\field{title}{Complex dynamics in recurrent cortical networks based on
spatially realistic connectivities}
\verb{url}
\verb http://invibe.net/LaurentPerrinet/Publications/Voges12
\endverb
\field{volume}{6}
\field{journaltitle}{Frontiers in Computational Neuroscience}
\field{year}{2012}
\endentry
\entry{Haykin99}{book}{}
\name{author}{1}{}{%
{{}%
{Haykin}{H.}%
{S.}{S.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Prentice Hall}%
}
\strng{namehash}{HS1}
\strng{fullhash}{HS1}
\field{title}{. Neural Networks: A Comprehensive Foundation}
\field{year}{1999}
\endentry
\entry{Teich97}{article}{}
\name{author}{5}{}{%
{{}%
{Teich}{T.}%
{M.~C.}{M.~C.}%
{}{}%
{}{}}%
{{}%
{Heneghan}{H.}%
{C.}{C.}%
{}{}%
{}{}}%
{{}%
{Lowen}{L.}%
{S.~B.}{S.~B.}%
{}{}%
{}{}}%
{{}%
{Ozaki}{O.}%
{T.}{T.}%
{}{}%
{}{}}%
{{}%
{Kaplan}{K.}%
{E.}{E.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{TMCHCLSBOTKE1}
\strng{fullhash}{TMCHCLSBOTKE1}
\field{abstract}{%
We used a variety of statistical measures to identify the point process that
describes the maintained discharge of retinal ganglion cells ({RGC}'s) and
neurons in the lateral geniculate nucleus ({LGN}) of the cat. These measures
are based on both interevent intervals and event counts and include the
interevent-interval histogram, rescaled range analysis, the event-number
histogram, the Fano factor, Allan factor, and the periodogram. In addition,
we applied these measures to surrogate versions of the data, generated by
random shuffling of the order of interevent intervals. The continuing
statistics reveal 1/f-type fluctuations in the data (long-duration power-law
correlation), which are not present in the shuffled data. Estimates of the
fractal exponents measured for {RGC}- and their target {LGN}-spike trains are
similar in value, indicating that the fractal behavior either is transmitted
form one cell to the other or has a common origin. The gamma-r renewal
process model, often used in the analysis of visual-neuron interevent
intervals, describes certain short-term features of the {RGC} and {LGN} data
reasonably well but fails to account for the long-duration correlation. We
present a new model for visual-system nerve-spike firings: a gamma-r renewal
process whose mean is modulated by fractal binomial noise. This fractal,
doubly stochastic point process characterizes the statistical behavior of
both {RGC} and {LGN} data sets remarkably well.%
}
\field{number}{3}
\field{pages}{529\bibrangedash 546}
\field{title}{Fractal character of the neural spike train in the visual
system of the cat.}
\verb{url}
\verb http://view.ncbi.nlm.nih.gov/pubmed/9058948
\endverb
\field{volume}{14}
\list{location}{2}{%
{Department of Electrical}%
{Computer Engineering, Boston University, Massachusetts 02215, USA.
teich@bu.edu}%
}
\field{journaltitle}{JOSA A}
\field{month}{03}
\field{year}{1997}
\endentry
\entry{Dayan01}{book}{}
\name{author}{2}{}{%
{{}%
{Dayan}{D.}%
{P.}{P.}%
{}{}%
{}{}}%
{{}%
{Abbott}{A.}%
{L.~F.}{L.~F.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{The MIT Press}%
}
\keyw{taouali15}
\strng{namehash}{DPALF1}
\strng{fullhash}{DPALF1}
\field{abstract}{%
Theoretical neuroscience provides a quantitative basis for describing what
nervous systems do, determining how they function, and uncovering the general
principles by which they operate. This text introduces the basic mathematical
and computational methods of theoretical neuroscience and presents
applications in a variety of areas including vision, sensory-motor
integration, development, learning, and memory. The book is divided into
three parts. Part I discusses the relationship between sensory stimuli and
neural responses, focusing on the representation of information by the
spiking activity of neurons. Part {II} discusses the modeling of neurons and
neural circuits on the basis of cellular and synaptic biophysics. Part {III}
analyzes the role of plasticity in development and learning. An appendix
covers the mathematical methods used, and exercises are available on the
book's Web site.%
}
\field{edition}{1st}
\field{howpublished}{Hardcover}
\field{title}{Theoretical Neuroscience: Computational and Mathematical
Modeling of Neural Systems}
\verb{url}
\verb http://www.amazon.com/exec/obidos/redirect?tag=citeulike07-20\&path=ASI
\verb N/0262041995
\endverb
\field{journaltitle}{Journal of Cognitive Neuroscience}
\field{month}{12}
\field{year}{2001}
\endentry
\entry{Cochran54}{article}{}
\name{author}{1}{}{%
{{}%
{Cochran}{C.}%
{W~G.}{W.~G.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{JSTOR}%
}
\strng{namehash}{CWG1}
\strng{fullhash}{CWG1}
\field{number}{4}
\field{pages}{417\bibrangedash 451}
\field{title}{Some methods for strengthening the common {$\chi$} 2 tests}
\field{volume}{10}
\field{journaltitle}{Biometrics}
\field{year}{1954}
\endentry
\entry{Yates99}{book}{}
\name{author}{3}{}{%
{{}%
{Yates}{Y.}%
{D.}{D.}%
{}{}%
{}{}}%
{{}%
{Moore}{M.}%
{D~S.}{D.~S.}%
{}{}%
{}{}}%
{{}%
{McCabe}{M.}%
{G~P.}{G.~P.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{W. H. Freeman}%
}
\strng{namehash}{YDMDSMGP1}
\strng{fullhash}{YDMDSMGP1}
\field{edition}{1st}
\field{title}{Practice Of Statistics Advanced Placement}
\field{year}{1999}
\endentry
\entry{Campbell07}{article}{}
\name{author}{1}{}{%
{{}%
{Campbell}{C.}%
{I.}{I.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Wiley Online Library}%
}
\strng{namehash}{CI1}
\strng{fullhash}{CI1}
\field{number}{19}
\field{pages}{3661\bibrangedash 3675}
\field{title}{Chi squared and Fisher Irwin tests of two by two tables with
small sample recommendations}
\field{volume}{26}
\field{journaltitle}{Statistics in {M}edicine}
\field{year}{2007}
\endentry
\entry{Amarasingham06}{article}{}
\name{author}{5}{}{%
{{}%
{Amarasingham}{A.}%
{Asohan}{A.}%
{}{}%
{}{}}%
{{}%
{Chen}{C.}%
{Ting-Li~L.}{T.-L.~L.}%
{}{}%
{}{}}%
{{}%
{Geman}{G.}%
{Stuart}{S.}%
{}{}%
{}{}}%
{{}%
{Harrison}{H.}%
{Matthew~T.}{M.~T.}%
{}{}%
{}{}}%
{{}%
{Sheinberg}{S.}%
{David~L.}{D.~L.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Society for Neuroscience}%
}
\keyw{taouali15}
\strng{namehash}{AACTLLGSHMTSDL1}
\strng{fullhash}{AACTLLGSHMTSDL1}
\field{abstract}{%
The variability of cortical activity in response to repeated presentations of
a stimulus has been an area of controversy in the ongoing debate regarding
the evidence for fine temporal structure in nervous system activity. We
present a new statistical technique for assessing the significance of
observed variability in the neural spike counts with respect to a minimal
Poisson hypothesis, which avoids the conventional but troubling assumption
that the spiking process is identically distributed across trials. We apply
the method to recordings of inferotemporal cortical neurons of primates
presented with complex visual stimuli. On this data, the minimal Poisson
hypothesis is rejected: the neuronal responses are too reliable to be fit by
a typical firing-rate model, even allowing for sudden, time-varying, and
trial-dependent rate changes after stimulus onset. The statistical evidence
favors a tightly regulated stimulus response in these neurons, close to
stimulus onset, although not further away.%
}
\verb{doi}
\verb 10.1523/jneurosci.2948-05.2006
\endverb
\field{number}{3}
\field{pages}{801\bibrangedash 809}
\field{title}{Spike count reliability and the Poisson hypothesis.}
\verb{url}
\verb http://dx.doi.org/10.1523/jneurosci.2948-05.2006
\endverb
\field{volume}{26}
\list{location}{1}{%
{Division of Applied Mathematics, Brown University, Providence, Rhode
Island 02912, USA. asohan@dam.brown.edu}%
}
\field{journaltitle}{The Journal of Neuroscience}
\field{month}{01}
\field{year}{2006}
\endentry
\entry{Jackson07}{article}{}
\name{author}{2}{}{%
{{}%
{Jackson}{J.}%
{Jadin}{J.}%
{}{}%
{}{}}%
{{}%
{Redish}{R.}%
{A.~David}{A.~D.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Wiley Subscription Services, Inc., A Wiley Company}%
}
\keyw{taouali15}
\strng{namehash}{JJRAD1}
\strng{fullhash}{JJRAD1}
\field{abstract}{%
The firing of place cells in the rodent hippocampus is reliable enough to
infer the rodent's position to a high accuracy; however, hippocampal firing
also reflects the stages of complex tasks. Theories have suggested that these
task-stage responses may reflect changes in reference frame related to
task-related subgoals. If the hippocampus represents an environment in
multiple ways depending on a task's demands, then switching between these
cell assemblies should be detectable as a switch in spatial maps or reference
frames. Place cells exhibit extreme temporal variability or "overdispersion,"
which Fenton et al. suggest reflects changes in active cell-assemblies. If
reference-frame switching exists, investigating the relationship of the
single cell variability described by Fenton and colleagues to network level
processes provides an entry point to understanding the relationship between
cell-assembly-like mechanisms and an animal's behavior. We tested the
cell-assembly explanation for overdispersion by recording hippocampal neural
ensembles from rats running three tasks of varying spatial complexity: linear
track ({LT}), cylinder-foraging ({CF}), and cylinder-goal ({CG}). Consistent
with the reports by Fenton and colleagues, hippocampal place cells showed
high variance in their firing rates across place field passes on the {CF} and
{CG} tasks. The directional firing of hippocampal place cells on {LT}
provided a test of the reference-frame hypothesis: ignoring direction
produced overdispersion similar to the {CF} and {CG} tasks; taking direction
into account produced a significant decrease in overdispersion. To directly
examine the possibility of a network modulation of cell-assemblies, we
clustered the firing patterns within each pixel and chained them together to
construct whole-environment spatial firing maps. Maps were internally
self-consistent, switching with mean rates of several hundred milliseconds.
There were significant increases in map-switching rates following
reward-related events on the {LT} and {CG} tasks, but not on the {CF} task.
Our results link single cell variability with network-level processes and
imply that hippocampal spatial representations are made up of multiple,
continuous sub-maps, the selection of which depends on the animal's goals
when reward is tied to the animal's spatial behavior. (c) 2007 {Wiley-Liss},
Inc.%
}
\verb{doi}
\verb 10.1002/hipo.20359
\endverb
\field{number}{12}
\field{pages}{1209\bibrangedash 1229}
\field{title}{Network dynamics of hippocampal cell-assemblies resemble
multiple spatial maps within single tasks.}
\verb{url}
\verb http://dx.doi.org/10.1002/hipo.20359
\endverb
\field{volume}{17}
\list{location}{1}{%
{Graduate Program in Neuroscience, University of Minnesota, Minnneapolis,
Minnesota; Department of Neuroscience, University of Minnesota, Minneapolis,
Minnesota}%
}
\field{journaltitle}{Hippocampus}
\field{year}{2007}
\endentry
\entry{Hoel43}{article}{}
\name{author}{1}{}{%
{{}%
{Hoel}{H.}%
{Paul~G.}{P.~G.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{HPG1}
\strng{fullhash}{HPG1}
\verb{doi}
\verb 10.1214/aoms/1177731457
\endverb
\field{number}{2}
\field{pages}{155\bibrangedash 162}
\field{title}{On Indices of Dispersion}
\verb{url}
\verb http://dx.doi.org/10.1214/aoms/1177731457
\endverb
\field{volume}{14}
\field{journaltitle}{The Annals of Mathematical Statistics}
\field{month}{06}
\field{year}{1943}
\endentry
\entry{Kathirgamatamby53}{article}{}
\name{author}{1}{}{%
{{}%
{Kathirgamatamby}{K.}%
{N.}{N.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Biometrika Trust}%
}
\keyw{taouali15}
\strng{namehash}{KN1}
\strng{fullhash}{KN1}
\field{number}{1/2}
\field{title}{Note on the Poisson Index of Dispersion}
\verb{url}
\verb http://www.jstor.org/stable/2333115
\endverb
\field{volume}{40}
\field{journaltitle}{Biometrika}
\field{year}{1953}
\endentry
\entry{Eden10}{article}{}
\name{author}{2}{}{%
{{}%
{Eden}{E.}%
{Uri~T.}{U.~T.}%
{}{}%
{}{}}%
{{}%
{Kramer}{K.}%
{Mark~A.}{M.~A.}%
{}{}%
{}{}}%
}
\keyw{taouali15}
\strng{namehash}{EUTKMA1}
\strng{fullhash}{EUTKMA1}
\field{abstract}{%
An important characterization of neural spiking is the ratio of the variance
to the mean of the spike counts in a set of intervals---the Fano factor. For
a Poisson process, the theoretical Fano factor is exactly one. For simulated
or experimental neural data, the sample Fano factor is never exactly one, but
often appears close to one. In this short communication, we characterize the
distribution of the Fano factor for a Poisson process, allowing us to compute
probability bounds and perform hypothesis tests on the distribution of
recorded neural spike counts. We show that for a Poisson process the Fano
factor asymptotically follows a gamma distribution with dependence on the
number of observations of spike counts, and that convergence to this
asymptotic distribution is fast. The analysis provides a simple method to
determine how close to 1 the computed Fano factor should be and to formally
test whether the observed variability in the spiking is likely to arise in
data generated by a Poisson process.%
}
\verb{doi}
\verb 10.1016/j.jneumeth.2010.04.012
\endverb
\field{number}{1}
\field{pages}{149\bibrangedash 152}
\field{title}{Drawing inferences from Fano factor calculations}
\verb{url}
\verb http://dx.doi.org/10.1016/j.jneumeth.2010.04.012
\endverb
\field{volume}{190}
\field{journaltitle}{Journal of Neuroscience Methods}
\field{month}{06}
\field{year}{2010}
\endentry
\entry{Jazayeri06}{article}{}
\name{author}{2}{}{%
{{}%
{Jazayeri}{J.}%
{Mehrdad}{M.}%
{}{}%
{}{}}%
{{}%
{Movshon}{M.}%
{J.~Anthony}{J.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Nature Publishing Group}%
}
\keyw{decoding, kaplan13, neural-decoding, population-coding, taouali15}
\strng{namehash}{JMMJA1}
\strng{fullhash}{JMMJA1}
\field{abstract}{%
Sensory information is encoded by populations of neurons. The responses of
individual neurons are inherently noisy, so the brain must interpret this
information as reliably as possible. In most situations, the optimal strategy
for decoding the population signal is to compute the likelihoods of the
stimuli that are consistent with an observed neural response. But it has not
been clear how the brain can directly compute likelihoods. Here we present a
simple and biologically plausible model that can realize the likelihood
function by computing a weighted sum of sensory neuron responses. The model
provides the basis for an optimal decoding of sensory information. It
explains a variety of psychophysical observations on detection,
discrimination and identification, and it also directly predicts the relative
contributions that different sensory neurons make to perceptual judgments.%
}
\verb{doi}
\verb 10.1038/nn1691
\endverb
\field{number}{5}
\field{pages}{690\bibrangedash 696}
\field{title}{Optimal representation of sensory information by neural
populations}
\verb{url}
\verb http://dx.doi.org/10.1038/nn1691
\endverb
\field{volume}{9}
\list{location}{1}{%
{Center for Neural Science, 4 Washington Place, Room 809, New York
University, New York, New York 10003, USA.}%
}
\field{journaltitle}{Nature {N}euroscience}
\field{month}{05}
\field{year}{2006}
\endentry
\entry{Box94}{misc}{}
\name{author}{3}{}{%
{{}%
{Box}{B.}%
{George~EP}{G.~E.}%
{}{}%
{}{}}%
{{}%
{Jenkins}{J.}%
{Gwilym~M}{G.~M.}%
{}{}%
{}{}}%
{{}%
{Reinsel}{R.}%
{Gregory~C}{G.~C.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Prentice-Hall, Englewood Cliffs NJ}%
}
\strng{namehash}{BGEJGMRGC1}
\strng{fullhash}{BGEJGMRGC1}
\field{title}{Time Analysis, Forecasting and Control}
\field{year}{1994}
\endentry
\entry{Gourieroux84}{article}{}
\name{author}{3}{}{%
{{}%
{Gourieroux}{G.}%
{C.}{C.}%
{}{}%
{}{}}%
{{}%
{Monfort}{M.}%
{A.}{A.}%
{}{}%
{}{}}%
{{}%
{Trognon}{T.}%
{A.}{A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{JSTOR}%
}
\strng{namehash}{GCMATA1}
\strng{fullhash}{GCMATA1}
\field{pages}{701\bibrangedash 720}
\field{title}{Pseudo maximum likelihood methods: Applications to Poisson
models}
\field{journaltitle}{Econometrica: Journal of the Econometric Society}
\field{year}{1984}
\endentry
\entry{Lawless87}{article}{}
\name{author}{1}{}{%
{{}%
{Lawless}{L.}%
{J~F.}{J.~F.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Wiley Online Library}%
}
\strng{namehash}{LJF1}
\strng{fullhash}{LJF1}
\field{number}{3}
\field{pages}{209\bibrangedash 225}
\field{title}{Negative binomial and mixed Poisson regression}
\field{volume}{15}
\field{journaltitle}{Canadian {J}ournal of {S}tatistics}
\field{year}{1987}
\endentry
\entry{Dai13}{article}{}
\name{author}{3}{}{%
{{}%
{Dai}{D.}%
{H.}{H.}%
{}{}%
{}{}}%
{{}%
{Bao}{B.}%
{Y.}{Y.}%
{}{}%
{}{}}%
{{}%
{Bao}{B.}%
{M.}{M.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{DHBYBM1}
\strng{fullhash}{DHBYBM1}
\field{number}{1}
\field{pages}{21\bibrangedash 27}
\field{title}{Maximum likelihood estimate for the dispersion parameter of the
negative binomial distribution}
\field{volume}{83}
\field{journaltitle}{Statistics and {P}robability {L}etters}
\field{year}{2013}
\endentry
\entry{Piegorsch90}{article}{}
\name{author}{1}{}{%
{{}%
{Piegorsch}{P.}%
{W~W.}{W.~W.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{JSTOR}%
}
\strng{namehash}{PWW1}
\strng{fullhash}{PWW1}
\field{pages}{863\bibrangedash 867}
\field{title}{Maximum likelihood estimation for the negative binomial
dispersion parameter}
\field{journaltitle}{Biometrics}
\field{year}{1990}
\endentry
\entry{Clark89}{article}{}
\name{author}{2}{}{%
{{}%
{Clark}{C.}%
{S~J.}{S.~J.}%
{}{}%
{}{}}%
{{}%
{Perry}{P.}%
{J~N.}{J.~N.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{JSTOR}%
}
\strng{namehash}{CSJPJN1}
\strng{fullhash}{CSJPJN1}
\field{pages}{309\bibrangedash 316}
\field{title}{Estimation of the negative binomial parameter {$\kappa$} by
maximum quasi{-}likelihood}
\field{journaltitle}{Biometrics}
\field{year}{1989}
\endentry
\entry{Leon12}{article}{}
\name{author}{4}{}{%
{{}%
{Sanz-Leon}{S.-L.}%
{Paula}{P.}%
{}{}%
{}{}}%
{{}%
{Vanzetta}{V.}%
{I.}{I.}%
{}{}%
{}{}}%
{{}%
{Masson}{M.}%
{G~S.}{G.~S.}%
{}{}%
{}{}}%
{{}%
{Perrinet}{P.}%
{L.~U.}{L.~U.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{American Physiological Society}%
}
\keyw{bicv-sparse, kaplan13, motion-clouds, perrinetadamsfriston14,
sanz12jnp, vacher14}
\strng{namehash}{SLPVIMGSPLU1}
\strng{fullhash}{SLPVIMGSPLU1}
\field{abstract}{%
Choosing an appropriate set of stimuli is essential to characterize the
response of a sensory system to a particular functional dimension, such as
the eye movement following the motion of a visual scene. Here, we describe a
framework to generate random texture movies with controlled information
content, i.e., Motion Clouds. These stimuli are defined using a generative
model that is based on controlled experimental parametrization. We show that
Motion Clouds correspond to dense mixing of localized moving gratings with
random positions. Their global envelope is similar to natural-like
stimulation with an approximate full-field translation corresponding to a
retinal slip. We describe the construction of these stimuli mathematically
and propose an open-source Python-based implementation. Examples of the use
of this framework are shown. We also propose extensions to other modalities
such as color vision, touch, and audition.%
}
\verb{doi}
\verb 10.1152/jn.00737.2011
\endverb
\field{issn}{1522-1598}
\field{number}{11}
\field{pages}{3217\bibrangedash 3226}
\field{title}{Motion {C}louds: model{-}based stimulus synthesis of
natural{-}like random textures for the study of motion perception}
\verb{url}
\verb http://dx.doi.org/10.1152/jn.00737.2011
\endverb
\field{volume}{107}
\field{journaltitle}{Journal of Neurophysiology}
\field{month}{03}
\field{year}{2012}
\endentry
\entry{Scholl13data}{misc}{}
\name{author}{4}{}{%
{{}%
{Scholl}{S.}%
{B.}{B.}%
{}{}%
{}{}}%
{{}%
{Tan}{T.}%
{A~Y~Y.}{A.~Y.~Y.}%
{}{}%
{}{}}%
{{}%
{Corey}{C.}%
{J.}{J.}%
{}{}%
{}{}}%
{{}%
{Priebe}{P.}%
{N~J.}{N.~J.}%
{}{}%
{}{}}%
}
\strng{namehash}{SBTAYYCJPNJ1}
\strng{fullhash}{SBTAYYCJPNJ1}
\field{howpublished}{\url{ CRCNS.org. http://dx.doi.org/10.6080/K05Q4T1N}}
\field{title}{Spiking responses of LGN neurons in the mouse to drifting
gratings}
\field{year}{2013}
\endentry
\entry{Scholl13}{article}{}
\name{author}{4}{}{%
{{}%
{Scholl}{S.}%
{B.}{B.}%
{}{}%
{}{}}%
{{}%
{Tan}{T.}%
{A~Y~Y.}{A.~Y.~Y.}%
{}{}%
{}{}}%
{{}%
{Corey}{C.}%
{J.}{J.}%
{}{}%
{}{}}%
{{}%
{Priebe}{P.}%
{N~J.}{N.~J.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Soc Neuroscience}%
}
\strng{namehash}{SBTAYYCJPNJ1}
\strng{fullhash}{SBTAYYCJPNJ1}
\field{number}{26}
\field{pages}{10616\bibrangedash 10624}
\field{title}{Emergence of orientation selectivity in the Mammalian visual
pathway}
\field{volume}{33}
\field{journaltitle}{Journal of Neuroscience}
\field{year}{2013}
\endentry
\entry{Robinson63}{article}{}
\name{author}{1}{}{%
{{}%
{Robinson}{R.}%
{D~A.}{D.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{IEEE}%
}
\strng{namehash}{RDA1}
\strng{fullhash}{RDA1}
\field{number}{4}
\field{pages}{137\bibrangedash 145}
\field{title}{A method of measuring eye movemnent using a scieral search coil
in a magnetic field}
\field{volume}{10}
\field{journaltitle}{Bio-medical Electronics, IEEE Transactions on}
\field{year}{1963}
\endentry
\entry{Collewijn75}{article}{}
\name{author}{3}{}{%
{{}%
{Collewijn}{C.}%
{H.}{H.}%
{}{}%
{}{}}%
{{}%
{Mark}{M.}%
{F.}{F.}%
{Van~der}{V.~d.}%
{}{}}%
{{}%
{Jansen}{J.}%
{T~C.}{T.~C.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Elsevier}%
}
\strng{namehash}{CHVdMFJTC1}
\strng{fullhash}{CHVdMFJTC1}
\field{number}{3}
\field{pages}{447\bibrangedash IN5}
\field{title}{Precise recording of human eye movements}
\field{volume}{15}
\field{journaltitle}{Vision {R}esearch}
\field{year}{1975}
\endentry
\entry{Hays82}{article}{}
\name{author}{1}{}{%
{{}%
{Hays}{H.}%
{A~V.}{A.~V.}%
{}{}%
{}{}}%
}
\strng{namehash}{HAV1}
\strng{fullhash}{HAV1}
\field{number}{1}
\field{pages}{1\bibrangedash 10}
\field{title}{A {UNIX}-based multiple process system for real{-}time data
acquisition and control}
\field{volume}{2}
\field{journaltitle}{WESCON Conf Proc}
\field{year}{1982}
\endentry
\entry{Cavanaugh02}{article}{}
\name{author}{3}{}{%
{{}%
{Cavanaugh}{C.}%
{J~R.}{J.~R.}%
{}{}%
{}{}}%
{{}%
{Bair}{B.}%
{W.}{W.}%
{}{}%
{}{}}%
{{}%
{Movshon}{M.}%
{J~A.}{J.~A.}%
{}{}%
{}{}}%
}
\list{publisher}{1}{%
{Am Physiological Soc}%
}
\strng{namehash}{CJRBWMJA1}
\strng{fullhash}{CJRBWMJA1}
\field{number}{5}
\field{pages}{2530\bibrangedash 2546}
\field{title}{Nature and interaction of signals from the receptive field
center and surround in macaque V1 neurons}
\field{volume}{88}
\field{journaltitle}{Journal of Neurophysiology}
\field{year}{2002}
\endentry
\lossort
\endlossort
\endinput