Young Investigator Abstracts


 
Resampling-based methods for multiple hypothesis testing in
neurophysiological data 

Asohan Amarasingham(1), Matthew Harrison(2), Gyuri Buzsaki(1)
(1) Rutgers University
(2) Carnegie Mellon University

Many problems in neurophysiological data analysis motivate the use of
resampling-based hypothesis tests. This is because some of the
uncertainties intrinsic to the experimental setup can cast doubt on a
parametric framework, and also because many hypotheses of experimental
interest can naturally be translated into resampling schemes. One
example is the use of trial shuffling to test hypotheses about
differences in spiking activity across experimental conditions, as in
analyses of the peri-stimulus time histogram (PSTH). Another example
is the use of spike jitter to detect the existence of finely timed
spiking relationships in spike trains, over and above the cruder
effects of slowly-varying firing rates.

Here, we point out that resampling-based techniques can be extended in
a simple way to accomodate controls over simultaneous inference. Such
extensions have been described previously, but appear to be somewhat
overlooked in many applications to neurophysiology. The need for such
controls arises frequently because many statistics of interest are
multiple-valued and indexed by continuous variables such as time,
space, or other stimulus-dependent dimensions. For example, in PSTH
analysis, it would be useful to confirm differences across two
experimental conditions, but also to identify the subset of times at
which the neurons' responses differ.

We illustrate these ideas in the context of several typical problems
in multi-unit data analysis, such as i) identifying differences in the
PSTH across experimental conditions, ii) separating the effects of
firing rate from correlations, and iii) detecting monosynaptic
interactions in a population of neurons, along with their variation as
a function of time and/or space. We draw examples from simulated spike
trains that are characterized by nontrivial trial-to-trial variability
as well as from experimental multi-unit data recorded from the
hippocampus and prefrontal cortex of behaving rats in a delayed
alternation task. We also provide some discussion of the mathematical
assumptions underlying such controls, and some guidance in their
interpretation.

Email: asohan@andromeda.rutgers.edu

---------------------------------------------------------
---------------------------------------------------------
 
Spectral analysis of extracellular neuronal signals: 
A tale of two oscillations

Izhar Bar-Gad
Bar-Ilan University

Spectral analysis of neuronal data is an increasingly common
tool in uncovering activity patterns characteristic of both normal
behavior and abnormal pathological states. However, analysis of the
same oscillatory phenomenon may yield vastly different results based
on the type of neurophysiological signal studied. Processing of
multi-electrode extracellular recordings yields multiple
representations of the original recorded signal: local field potential
(low-pass filtered signal), background activity (rectified high-pass
filtered signal), multi-unit and single unit activity (spike trains of
multiple and single neurons respectively). These signals reflect
different properties of the population of neurons surrounding the
electrodes. Comparative spectral assessment of these signals allows
the formation of a generalized view of the oscillatory processes
affecting single neurons and their surrounding networks. These
spectral analysis methods are demonstrated via the analysis of
oscillatory activity characteristic to Parkinson's disease in
human patients undergoing deep brain stimulation (DBS) electrode
implantation surgery. This analysis yields new insights regarding the
nature of the oscillatory processes and their relation to both anatomy
and clinical manifestation of the disease.

Email: bargad@gmail.com

---------------------------------------------------------
---------------------------------------------------------
 
The distance between the (initial) eye and arm position is coded in
PPC using a gain field

Steve W.C. Chang and Lawrence H. Snyder
Department of Anatomy & Neurobiology
Washington University School of Medicine
St. Louis, MO

To examine the frame of reference used by the neurons in the posterior
parietal cortex during reaching movement planning, we collected
neuronal data while monkeys planned reaches to targets at different
locations using either the contralateral or ipsilateral arm.  Target,
and starting eye and arm positions were varied across trials.  Then,
we fit these neuronal data to a non-linear model, in which cell
activity depended on target location with respect to the weighted eye
and arm positions.  The model also contained linear gain field terms
for eye and arm positions to account for activity modulation that is
independent from tuning shifts.  We asked how data from each cell
would fit this general model.

We found that cells in the posterior parietal cortex use diverse
frames of reference, ranging from pure eye-centered to pure
arm-centered, though there is a population bias for representing
targets in an eye-centered coordinate.  This bias was notably stronger
on ipsilateral limb compared to contralateral limb trials, suggesting
that contralateral limb target representation is more evolved toward
the motor domain.

Interestingly, for each cell, the gain fields for eye and arm
positions were negatively correlated.  The strengths of the eye and
arm gain fields were roughly matched in individual cells, suggesting
that it is the distance between the eye and arm, and not the
individual positions of the eye and arm per se, that is important.
Indeed, a model using a single gain field term for the distance
between eye and arm outperformed the more general model for the vast
majority of single neurons.

Previous studies have argued that neurons in the parietal reach region
and medial intraparietal area code targets with respect to eye
position (eye-centered).  Cells in dorsal premotor cortex have been
described as encoding a complex representation of target and arm
position in eye-centered coordinates and target position in
arm-centered coordinates.  We now describe a simpler representation in
parietal cortex, in which ~80% of neurons are best described as
encoding target location relative to a point intermediate between the
starting eye and arm position, gain-modulated by the distance between
the eye and arm.

Support Contributed By: NEI

Email: steve@eye-hand.wustl.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Relating correlation and firing rate in populations of spiking neurons

Brent Doiron, University of Pittsburgh
Jaime de la Rocha, New York University
Eric Shea-Brown, University of Washington
Kresimir Josic, University of Houston
Alex Reyes, New York University


Temporal correlations between spike trains from distinct neurons are
widely reported throughout cortical and subcortical areas.  However,
the impact that correlations have on stimulus coding remains unclear.
Using a pair of uncoupled neurons we study the transfer of input
current correlations to the correlations in output spike trains.  In
vitro experiments show that for fixed input correlation coefficient
the output correlation coefficient scales with the firing rates of the
neurons, and is surprisingly independent of the spike train variance
(de la Rocha et al., 2007; Shea-Brown et al., 2008).  Analysis of
integrate-and-fire model neurons allow for an explicit expression
relating the correlation coefficient of the input current and output
spike train correlation.  Using our theory we show that the
correlation-rate relationship obtained from experiments is robust to
input heterogeneities, and is a consequence of the threshold
nonlinearity inherent to spiking dynamics.

The link between correlation and firing rate has interesting
consequences for stimulus coding in populations of correlated spiking
neurons.  An immediate result is that a stimulus tuning of firing
rates implies a similar tuning in the spike correlation of a
population of neurons.  Calculating the Fisher information of a
population of correlation-tuned neurons we show how for certain
parameter regimes spike correlation can increase the stimulus coding
of the population.  Extending this result to a two-layer network we
show how the correlation-rate relation significantly enhances the
amount of propagated information over the network.  This is
accomplished by mitigating the sometimes deleterious impact of spike
correlations on coding with the known benefits of correlation on
propagating activity to downstream populations.

de la Rocha, Doiron, Shea-Brown, Josic, Reyes, Nature, 448:802-806,
2007 Shea-Brown, Josic, de la Rocha, Doiron, Phys. Rev. Lett., In
press, 2008.

Email: bdoiron@pitt.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Time Series Analyses of the Dependence Structure of Hybrid Data using
Mutual Information with an Example of Neurophysiological Data
 
Apratim Guha
Department of Statistics and Applied Probability, National University
of Singapore, Singapore

(joint work with Atanu Biswas, Applied Statistics Unit, Indian
Statistical Institute}

We propose a model that captures the inter-relationship between spike
trains emanating from neurons in the temporal cortex of mice, which is
a part of the auditory cortex, and local field potentials in the
vicinity. The need is a model which justifiably combines model for a
spike train with for the continuous local field potential
values. Towards that, we emphasize the requirement of verifying the
presence of a dependence structure for the hybrid data set: there are
numerous methods available for modelling univariate data, both
continuous and discrete, in neuroscience literature which produce
excellent results, and may be employed in the event the components of
the data set are independent. However, checking for dependence of the
components of a hybrid process is not an easy problem, as many of the
most commonly available methods, in particular the correlation
techniques, could produce erroneous results (Guha, 2005).

In this paper, we study and model the interdependence of the
components of hybrid data, of which some are continuous time series
and the rest are point processes. The concept of correlation is often
not suitable for assessing the dependence structure of hybrid data. We
employ mutual information towards assessing the dependence pattern of
hybrid data as a sensible alternative. We consider a time series of
bivariate hybrid (mixed discrete and continuous) data, and use the
concept of mutual information to describe the dependence between the
components of a series and also the crossed dependence between the two
series.

We model and analyze the hybrid time series data from neurons in the
temporal cortex of mice. We combine the idea of variable signal plus
ongoing activity (VSPOA) models (Chen et al, Chaos, 2006) with an
explicit stimuli model of Brown et al (Journal of Neuroscience, 1998)
to propose an ARMA-type model for the hybrid process. At its simplest
form, both process could be taken as AR(1)-type processes: we assume
that for the continuous process the observation at time t depends on
the value of the process at time (t-1), and the value of the discrete
process at times (t-1) and t; whereas the value of the discrete
process is assumed to be dependent on observations of both processes
at time (t-1).This cross dependence-structure is validated by mutual
information, which is also employed as an eyeball estimate of the
goodness of fit along with traditional measures. Variants of the model
are discussed and compared. Some simulation results are also given.
 
email:staag@nus.edu.sg

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Neural signals correlated with recognition memory in the macaque
hippocampus.

Michael J. Jutras (1,2,3*). Pascal Fries (4,5), 
Elizabeth A. Buffalo (1,2,3,6)
1 Yerkes National Primate Research Center, Atlanta, GA.
2 Neuroscience Program, Emory University, Atlanta, GA.
3 Center for Behavioral Neuroscience, Atlanta, GA.
4 F.C. Donders Centre for Cognitive Neuroimaging, Netherlands.
5 Department of Biophysics, Radboud University Nijmegen, Netherlands.
6 Department of Neurology, Emory University School of Medicine, Atlanta, GA.

Recognition memory is the ability to perceive recently encountered
items as familiar.  This ability is impaired following lesions to the
hippocampus in both humans and non-human primates.  However, despite
extensive research, there is currently little evidence for recognition
memory signals in the primate hippocampus.  We examined hippocampal
activity as monkeys performed the Visual Preferential Looking Task
(VPLT), a recognition memory task which is highly sensitive to lesions
of the hippocampus.  This task elicited a range of behavioral
performance that allowed us to examine the relationship between
trial-to-trial fluctuations in recognition memory performance and
neural signals recorded in the hippocampus.

In the VPLT, monkeys were allowed to freely view large stimuli (110),
which were presented twice each and remained on the screen as long as
the monkey continued to look at them, up to a maximum looking time of
5 seconds.  Recognition memory was assessed by comparing the amount of
time the monkey looked at each stimulus during the first and second
presentations.  Overall, both monkeys demonstrated significant
recognition memory as evidenced by a reduction in looking time during
the second presentation relative to the first presentation (p<.001;
median reduction in looking time: 67%).  However, monkeys displayed a
range of behavioral performance, showing good recognition memory for
some stimuli (as evidenced by a large reduction in looking time from
the first to the second presentation), and poor recognition memory for
other stimuli (as evidenced by a small reduction in looking time).

We recorded spikes and local field potentials (LFPs) from the
hippocampus in two monkeys performing the VPLT.  Out of 133
hippocampal multi-units, 94 (71%) responded significantly to stimulus
presentation relative to the baseline pre-stimulus period.  The
firing rates of 28% of these visually-responsive multi-units were
significantly modulated depending on whether the stimuli were novel
or repeated.  Importantly, this modulation in firing rate was
significantly correlated with trial-to trial fluctuations in
performance such that increased firing rate differences between novel
and repeated trials were associated with better memory performance.
We also found that visually responsive hippocampal neurons showed
increased gamma-band spike-field coherence during encoding that
predicted recognition memory performance.  Across the population of
221 multi-unit-LFP pairs, a trial-by trial analysis revealed that
enhanced gamma-band synchrony during encoding was associated with
better !  memory performance.  Taken together, these data suggest
that memory formation in the hippocampus relies on a combination of
firing rate changes at the single-cell level and altered patterns of
synchronization at the population level.

Email: mjutras@emory.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Approximate methods for state-space models

Shinsuke Koyama
Department of Statistics and Center for the Neural Basis of 
Cognition
Carnegie Mellon University

A number of important data analysis problems in neuroscience can be solved 
using state-space models. The optimal estimate of the state is its 
conditional expectation given the observation histories, but this 
expectation is computationally demanding when nonlinearities are present. 
Various authors have therefore used Gaussian approximations to posterior 
densities that appear in the formulation. In the first part of the talk we 
investigate this approach, showing that the errors introduced by the 
approximation are not compounded across time. We then consider 
second-order expansions, and show that they can provide second-order 
accuracy in state estimates---but that no additional accuracy is possible 
by higher-order approximations. We discuss implementation of these methods 
and illustrate by decoding multielectrode motor cortical data.

In the second part of this talk, we develop fast methods for computing the 
exact MAP (posterior modal) path of state-space models. If the state 
dynamics are linear and driven by innovations with a log-concave density, 
and the observation model satisfies certain standard assumptions, then the 
optimization problem is strictly concave and may be solved rapidly with 
Newton's method because the Hessian of the loglikelihood is block 
tridiagonal. We describe applications of this approach to decoding 
problems.

Finally, we have developed real-time software which implements our 
decoding methods in conjunction with a brain-machine interface. We hope to 
be able to show results from a monkey using these methods to control 
3-dimensional robotic movement.

Parts of this work have been done in collaboration with Rob Kass, Lucia 
Castellanos Perez-Bolde, Cosma Rohilla Shalizi, Liam Paninski, Meel 
Velliste, Andrew Whitford in the Schwartz lab.


Email: koyama@stat.cmu.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Representational Similarity Analysis -- a framework for relating
computational theory, brain activity, and behavior

Nikolaus Kriegeskorte

A fundamental challenge for systems neuroscience is to quantitatively
relate its three major branches of research: behavioral
experimentation, brain-activity experimentation, and computational
modeling. We suggest relating the three branches by quantitatively
comparing representational dissimilarity matrices. For each pair of
experimental conditions (e.g. each pair of stimuli), the
representational dissimilarity matrix contains an entry reflecting the
dissimilarity of the activity patterns associated with the two
conditions. Intuitively, the dissimilarity matrix encapsulates the
information carried by a given representation in a brain or
computational model.

Using the representational dissimilarity matrix as the signature of
each representation allows quantitative comparisons between
representations without the need for a spatial correspondency mapping
(defining, for example, which neuron corresponds to which unit of a
computational model). We can relate representations measured in
biological brains to representations in computational network
models. We can also relate representations between different brain
regions, as well as between different individuals and
species. Moreover, we can relate different modalities of brain
activity measurement (e.g. single-cell recording and fMRI) for a given
brain representation. Our approach requires fewer assumptions than
formal information-theoretic alternatives and can handle designs with
large numbers of conditions (e.g. many stimuli).

We demonstrate this method, called ``representational similarity
analysis'', by relating object representations between monkey and
human IT (measured with single-cell recording and fMRI, respectively)
and several computational models. The representational dissimilarity
matrices are simultaneously related via second-level application of
multidimensional scaling and tested for relatedness and distinctness
using randomization and bootstrap techniques.

Email: kriegeskorten@mail.nih.gov

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Large-scale parameter estimation and dynamic source localization for
the magnetoencephalography (MEG) inverse problem

Camilo Lamus (1,2)
Simona Temereanca (1,3)
C.J. Long (1,3)
Matti S. Hdmdldinen (3,4)
E.N Brown (1,2,4)
P.L. Purdon (1,2,3)
1 Dept. of Brain and Cognitive Sciences, Massachusetts Institute 
of Technology, Cambridge, MA
2 Department of Anaethesia and Critical Care, Massachusetts 
General Hospital, Boston, MA.
3 MGH/MIT/HMS Martinos Center for Biomedical Imaging, Charlestown, MA.
4 Harvard-MIT Division of Health Science and Technology, 
Cambridge, MA

Dynamic estimation methods based on linear state-space models have
been applied to the inverse problem in magnetoencephalography (MEG),
and can improve source localization compared with static methods by
incorporating temporal continuity as a constraint.  The efficacy of
these methods is influenced by how well the state-space model
approximates the dynamics of the underlying brain current sources.
While some components of the state-space model can be defined from
brain anatomy and knowledge of the noise structure, parameters
governing the temporal evolution of underlying current sources,
i.e. the state noise covariance matrix, are unknown and must be chosen
on an ad-hoc basis or estimated from data. We applied the
Expectation-Maximization (EM) algorithm to estimate the state noise
covariance matrix and dipole sources in an MEG state-space model. This
estimation problem is a challenging one due to the size of the state
space, typically involving > 5,000 sources to accurately represent the
cortical surface.  We previously demonstrated in small-scale
simulation and ROI-based studies (100's of dipoles) that the
resulting source estimates were superior to those provided by static
Minimum Norm Estimate (MNE) or dynamic Fixed-Interval Smoother (FIS)
employing ad hoc parameter selection. In this work, we extend our
previous result to demonstrate that the EM algorithm outperforms the
MNE and FIS methods both in a large-scale simulation study and in an
analysis of human mu rhythm data.

Email: lamus@mit.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Markov Models for Neuronal Spike Trains

Jeffrey Liebner 
Lafayette College

Sequences of action potentials from neurons, known as spike trains,
can be modeled as point processes.  One of the simplest ways to model
spike times is to assume that they follow a Poisson process.
However, this assumption is insufficient when modeling within-trial
firing behavior.  Generally, the firing behavior of a neuron is
dependent upon the past spiking history of the neuron.  Neurons have
a refractory period during which, after each spike, the probability
of firing is greatly reduced.  In addition, during repeated trials of
an experiment, the firing rate of a neuron may exhibit
trial-dependent, slowly-varying changes.  This produces excess
trial-to-trial variation, beyond what would be expected by random
chance.  Such trial dependent effects include latency and
excitability.  Furthermore, the firing patterns of neurons may be
correlated with outside sources, such as oscillatory behavior in the
brain or the firing of neighboring neurons.  These behaviors of
neurons lead to significant deviations from the simple Poisson
process.

In order to incorporate the past spiking behavior of the neuron, a
Markov model, namely the inhomogeneous Markov interval process, has
previously been proposed.  However, in order to accommodate the other
effects that are seen, this model has been expanded, allowing it to
account for latency and excitability effects.  The phase of any
recorded oscillatory behavior is also included as a predictor of
spiking probability, with the phase being is estimated by smoothing
and the Hilbert transform.  A Gibbs sampling procedure is utilized to
estimate the effect of each aspect.  This procedure also applies the
mechanism of Bayesian Adaptive Regression Splines (BARS) to obtain
smoothed estimates of each effect.

The model has been utilized to analyze the variations in the firing
behavior of neurons in the locust brain.  The work of Mark Stopfer at
the National Institutes of Health has revealed variations in the
neuronal response of locusts over repeated experimentation, especially
with regard to excitability and relation to the oscillations of local
field potential.  By applying the model to data supplied by the
Stopfer lab, these variations were able to be quantified.  This
presentation will discuss the proposed Markov model and its
applications to neuronal data.

Email: liebnerj@lafayette.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Wide-band local field potential phase predicts time-dependent single
unit firing in awake auditory cortex

Robert C. Liu and Edgar Galindo-Leon
Department of Biology, Emory University 

Single units (SU) are considered the gold standard for studying neural
activity.  However, these can be difficult to isolate on electrodes, a
problem further exacerbated by modern electrophysiology's move
towards multielectrode paradigms.  Furthermore, SU's by definition
represent only the activity of individual neurons, whereas there is an
increasing interest in how populations of neurons behave.  These are
some of the reasons local field potentials (LFP) have gained
prominence as a tool for monitoring local neural networks in vivo.
LFP's are believed to reflect slow, non-spiking neural events such as
synaptic input and voltage-gated membrane fluctuations that are
synchronized across a nearby population of geometrically-aligned
neurons.  But how does the activity of this local neural network
relate to the spiking of one of its constituent SU's?

This question was investigated in the auditory cortex of awake,
head-restrained mice listening to species-specific communication
calls.  Well-isolated SU's were recorded (300-6000 Hz) on a
high-impedance (4-6 Mohm) tungsten electrode simultaneously with
LFP's that were separately filtered and recorded (2-100 Hz).  A
recent study of LFP's in anesthetized rat barrel cortex found that
the wide-band LFP phase (extracted by Hilbert transform) just prior to
whisker stimulation was correlated with the magnitude of the multiunit
response, integrated over 50 ms (Haslinger et al, 2006).  In our
analysis, we also decomposed the LFP into amplitude and phase
components, but focused instead on the temporal trajectories of their
distributions around a SU spike.  We then applied a Bayesian algorithm
to predict the post-stimulus time histogram of the SU's response to
the communication calls.

Our results show that the Hilbert phase is a better predictor of the
time course of the SU response than either the Hilbert amplitude or
the raw LFP.  Moreover, since even the LFP phase distributions derived
from spontaneous periods of spiking could predict the stimulus-driven
SU response, it appears that the SU's relationship with its local
neural network is relatively stable during these different states.
Finally, our results indicate that on average, the local network
activity seems to follow SU firing, suggesting that the SU's in our
recordings tend to drive LFP fluctuations.  In summary, our analysis
methods provide a new tool for in-depth characterization of the
relationship between SU and LFP signals.

References
Haslinger R. et al, J Neurophysiol. 96(3):1658-63, 2006.

Email: robert.liu@emory.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Trajectory prediction in rats using an analysis of empirical firing
distributions of hippocampal place cells

Michael Prerau, Uri Eden
Boston University
Program in Neuroscience

It has been well established that certain hippocampal place cells fire
differentially in the same spatial region under different contextual
conditions. Previous analyses make several assumptions about place
field data that greatly reduce their ability to describe stochastic
structure and neural dynamics associated with these cells.  Common
assumptions about the receptive field characteristics of these
neurons hold that the number of counts in any interval has either a
normal or inhomogeneous Poisson distribution with means determined by
a Gaussian spatial receptive field shape. An analysis of the
empirical data shows that in the majority of cases, these assumptions
do not hold true. The conventional approach to determining whether a
neuron fires differentially under different condition using a low
resolution two-way ANOVA, relies extensively on this erroneous
assumption.  Another widespread assumption that is often not borne
out by the data is that the firing properties of t!  hese neurons are
stationary.

We present a Bayesian framework to characterize empirical
distributions of neural firing. This method offers greater spatial
resolution then previous methods and can capture the complex
probability distributions that compose place cell receptive
fields. The Bayesian framework allows us to compare the neuronal
activity under different contexts as well as to compute the likelihood
of newly observed data under the distributions for each context. Our
method also provides a quantitative metric of the degree of
differential firing shown in a given neuron.  One can expand this
method to perform population analyses and decode or predict context
from population activity. We demonstrate the method by predicting the
direction that a rat will turn on a T-maze from firing on the maze
stem.

Email: prerau@bu.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Bayesian Estimation of the time-varing Rate and Irregularity of
Neuronal Spiking

Takeaki Shimokawa and Shigeru Shinomoto
Graduate School of Science, Kyoto University,
Sakyo-ku, Kyoto 606-8502, Japan

We have recently revealed that there is an aspect of time-local firing
irregularity that is invariant with the time and the firing rate for
individual cortical neurons in vivo[1,2].  Conversely, Davies et al
reported that the time-local firing irregularity varied significantly
according to behavioral contexts in some other cortical area[3].  We
wish to investigate here the cause of apparent disagreement of these
conclusions by thoroughly examining how easily the firing patterns are
modified with the firing rate.  For this purpose, we develop here the
Bayes method that allows us to estimate both the instantaneous rate
and irregularity for a given spike sequence: We first consider the
process of generating spikes under a given rate and irregularity, and
then invert the conditional probability distribution to infer the rate
and the irregularity from the data, with the prior distributions
penalyzing the large gradient of the parameters.  The penalty
hyperparameters are determined by maximizing the (marginal)
likelihood, which can be carried out numerically using the
expectation-maximization (EM) method and the Kalman filter under the
assumption that the distributions of the two parameters are Gaussian.
We applied the present Bayesian estimation method to the
experimentally recorded spike data taken from Neural Signal
Archive[4], and revealed that there is a systematic correlation
between firing rate and irregularity, and that the degree of the
variability in the firing irregularity greatly depends on the cortical
areas.

[1] Shinomoto S., Shima K. and Tanji J. (2003) Differences in Spiking
Patterns Among Cortical Neurons. Neural Comput 15, 2823-2842.  [2]
Shinomoto S., Miyazaki Y., Tamura H. and Fujita I. (2005) Regional and
Laminar Differences in In Vivo Firing Patterns of Primate Cortical
Neurons. J Neurophysiol 94, 567-575.
[3] Davies R.M., Gerstein G.L. and Baker S.N. (2006) Measurement of
Time-Dependent Changes in the Irregularity of Neural Spiking. J
Neurophysiol 96, 906-918.
[4] Neural Signal Archive, available online at http://www.neuralsignal.org.

Email: shimokawa@ton.scphys.kyoto-u.ac.jp

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Uncovering higher-order synchrony within massively parallel spike trains

Benjamin Staude [1], Stefan Rotter [2,3], Sonja Gruen [1,4]
1. Unit of Statistical Neuroscience, Theoretical Neuroscience Group,
Brain Science Institute, RIKEN, Wako, Japan
2. Theory and Data Analysis, Institute for Frontier Areas of
Psychology and Mental Health, Freiburg, Germany
3. Bernstein Center for Computational Neuroscience Freiburg, Germany
4. Bernstein Center for Computational Neuroscience Berlin, Germany

The cell assembly hypothesis [1] postulates dynamically interacting
groups of neurons as building blocks of cortical information
processing. Synchronized spiking across large neuronal groups was
later suggested as a potential signature for active assemblies [2],
and analysis methods for assembly detection focussed on the estimation
of higher-order correlations among simultaneously recorded neurons
[3,4]. However, the number of parameters in presently available
techniques grows exponentially with the number of recorded neurons,
which requires vast sample sizes. As a consequence, most attempts to
detect active cell assemblies resort to pairwise interactions. These,
however, do not allow to infer on large synchronized neuronal pools,
and are insensitive for sparse synchronous events [5]. As massively
parallel extracellular recordings are becoming available, the limited
experimental evidence in favor of the cell assembly hypothesis has
been assigned, among other things, to a lack of suitable analysis
tools [6].

Here we present a novel procedure to detect synchronized spiking in
large neuronal pools that circumvents the need for extensive sample
sizes. Based on estimates of only a few low-order cumulants of the
superimposed and discretely sampled activity of all recorded neurons
(population spike counts) we devise a statistical test for the
presence of higher-order synchrony among the spike trains. The test
exploits the simple fact that absence of higher-order synchrony
imposes constraints on correlations of lower order, which can be
estimated via the respective cumulants of the population spike
counts. The inference of higher-order synchrony from measured
cumulants of lower order circumvents the need to estimate large
numbers of higher-order parameters and therefore is less susceptible
to the limited sample sizes from in vivo recordings than previous
approaches. The method is tested for correlated Poisson processes,
where correlations of various orders are induced by 'inserting'
appropriate patterns of near-synchronous spikes [7]. When applied to
simulated data, the test is surprisingly sensitive for higher-order
synchrony present in the data, even if their effects on pairwise
correlation coefficients c are very small (in the range of c ~ 0.01,
compare cf. [5]). The applicability of the method is illustrated by
estimates for the required sample sizes and the robustness against
deviations from the Poisson assumption.

Acknowledgments.  Funded by NaFoG Berlin, the German Ministry for
Education and Research (BMBF grants 01GQ01413 and 01GQ0420), and the
Stifterverband fur die Deutsche Wissenschaft

References.
1. Hebb. Organization of behavior. Wiley, 1949
2. Abeles. Local cortical circuits. Springer, 1982
3. Martignon et al. Biol Cyber 1995, 73:69-81
4. Nakahara & Amari. Neural Comput 14:2296-2316, 2002
5. Schneidman et al. Nature 440:1007-1012, 2006
6. Brown et al. Nat Neurosci 7:456-461, 2004
7. Ehm et al. EJS 1:473-495, 2007

Email: staude@brain.riken.jp

---------------------------------------------------------------------------
---------------------------------------------------------------------------
  
Model-based optimal inference of spike times and 
calcium dynamics given noisy and intermittent calcium-
fluorescence imaging

Joshua Vogelstein, Department of Neuroscience,
Johns Hopkins School of Medicine.
Liam Paninski, Department of Statistics, Columbia 
University.

As recent advances in calcium sensing technologies enable us to 
simultaneously image many neurons, complementary analytical 
tools must also be developed to maximize the utility of this 
experimental paradigm. While the observations are fluorescence 
movies, the signals of interest are the spike trains and/or time-
varying intracellular calcium concentrations, [Ca2+].  Inferring 
the value of these "hidden states" is often problematic for a 
number of reasons: (i) observations are inherently both noisy 
and intermittent, (ii) the relationship between fluorescence, 
[Ca2+], and spike trains is nonlinear, and (iii) the parameters 
governing the dynamics of these states are typically unknown.
We develop a number of algorithms designed to optimally infer 
precise spike times and [Ca2+] dynamics, given the above 
features. In particular, we develop 3 iterative algorithms: (i) a 
log-barrier method for non-negative deconvolution, (ii) a 
"greedy" projection pursuit regression for further constraining 
the inferred spike train, and (iii) a particle filtering approach for 
computing the marginal distribution of [Ca2+] and spikes at 
each time step, conditioned on all the fluorescence 
observations.  Each algorithm recursively infers the underlying 
spike train, and then estimates the model parameters. All of 
these approaches facilitate incorporating stimulus information, 
potentially improving the inference fidelity over previously 
described methods. 

The methods also differ somewhat from one another. The first 
two assume a relatively simple observation and spiking model, 
and can therefore be made extremely efficient, costing O(T) time 
per iteration.  We use these fast methods to initialize the 
parameters for the more sophisticated, but more 
computationally expensive particle filtering, which confers 
several advantages over the fast methods.  First, nonlinear dye 
saturation is incorporated into the model.  Second, spike 
inference can consider refractoriness - or even coupling from 
other neurons - if that information is available.  Third, because 
we compute the marginal distribution of spikes at each time, 
conditioned on all the observations, noise is better accounted 
for, and errorbars on the estimates may be provided.  Fourth, 
the absolute [Ca2+] may be determined without additional 
experimentation or ratiometric sensors.  Finally, while this 
method is optimal for a set of model assumptions, many of 
these assumptions may be relaxed for greater accuracy if 
desired. To test the degree to which these algorithms 
successfully infer spike times from fluorescence movies, 
analysis of experimental data is currently underway.

Email: joshuav@jhu.edu

---------------------------------------------------------------------------
---------------------------------------------------------------------------
 
Reconstructing stereotyped movement by coupling
trajectory decoding and landmark-time decoding in the motor cortex

Wei Wu 
Department of Statistics
Florida State University

A number of decoding methods, varying from common linear Gaussian
models to more complicated point process frameworks, have been
developed to infer hand movement from neuronal firing activity in the
motor cortex.  Most of these methods focus on estimating subject's
hand trajectory in a continuous movement.  We recently proposed a
template-based time identification decoding approach with linear
convolution and dynamic programming.  We showed that if a stereotyped
movement is well represented by a sequence of targets (or landmarks),
then the main structure of the movement will be better addressed by
detecting the reaching times at those targets.

Both trajectory decoding and landmark-time decoding have advantages
respectively, whereas a coupling of these two different strategies
has not been examined.  Here we propose the synergy that comes from
combining these two approaches for a stereotyped movement under a
state-space framework, where the recordings were made in the arm area
of primary motor cortex in an awake behaving monkey using a
chronically implanted multi-electrode array.  We at first identify
the target times using an improved version of the template-based
method.  The identification is highly accurate with median error
around 60 ms.  Then we include the detected targets as a linear
control input in the kinematic model of the state-space formulation.
Such an inclusion is justified by the empirical linear relationship
between the kinematics and target position in the recorded data.
Experimental results show that the coupling model includes the
benefits from both approaches and significantly improves the d!
ecoding accuracy.
 
This work was done in collaboration with Zhiyi Chi and Nicholas
G. Hatsopoulos.

Email: wwu@stat.fsu.edu