Fitting neuron models to spike trains.  
Jump to Full Text  
MedLine Citation:

PMID: 21415925 Owner: NLM Status: PubMednotMEDLINE 
Abstract/OtherAbstract:

Computational modeling is increasingly used to understand the function of neural circuits in systems neuroscience. These studies require models of individual neurons with realistic inputoutput properties. Recently, it was found that spiking models can accurately predict the precisely timed spike trains produced by cortical neurons in response to somatically injected currents, if properly fitted. This requires fitting techniques that are efficient and flexible enough to easily test different candidate models. We present a generic solution, based on the Brian simulator (a neural network simulator in Python), which allows the user to define and fit arbitrary neuron models to electrophysiological recordings. It relies on vectorization and parallel computing techniques to achieve efficiency. We demonstrate its use on neural recordings in the barrel cortex and in the auditory brainstem, and confirm that simple adaptive spiking models can accurately predict the response of cortical neurons. Finally, we show how a complex multicompartmental model can be reduced to a simple effective spiking model. 
Authors:

Cyrille Rossant; Dan F M Goodman; Bertrand Fontaine; Jonathan Platkiewicz; Anna K Magnusson; Romain Brette 
Related Documents
:

15584935  Estimation of parameters for a mathematical model of growth hormone secretion. 22355615  Unified regression model of binding equilibria in crowded environments. 21647935  Twosample densitybased empirical likelihood tests for incomplete data in application ... 19152475  Computer modelling of ciliary motility. 15974515  Assessing performance of multihospital organizations: a measurement approach. 8870965  Retrieval of clinical science information using an interactive activation and competiti... 
Publication Detail:

Type: Journal Article Date: 20110223 
Journal Detail:

Title: Frontiers in neuroscience Volume: 5 ISSN: 1662453X ISO Abbreviation: Front Neurosci Publication Date: 2011 
Date Detail:

Created Date: 20110318 Completed Date: 20110714 Revised Date: 20130813 
Medline Journal Info:

Nlm Unique ID: 101478481 Medline TA: Front Neurosci Country: Switzerland 
Other Details:

Languages: eng Pagination: 9 Citation Subset:  
Affiliation:

Laboratoire Psychologie de la Perception, CNRS, Université Paris Descartes Paris, France. 
Export Citation:

APA/MLA Format Download EndNote Download BibTex 
MeSH Terms  
Descriptor/Qualifier:


Grant Support  
ID/Acronym/Agency:

240132//European Research Council 
Comments/Corrections 
Full Text  
Journal Information Journal ID (nlmta): Front Neurosci Journal ID (publisherid): Front. Neurosci. ISSN: 16624548 ISSN: 1662453X Publisher: Frontiers Research Foundation 
Article Information Download PDF Copyright © 2011 Rossant, Goodman, Fontaine, Platkiewicz, Magnusson and Brette. openaccess: Received Day: 29 Month: 10 Year: 2010 Accepted Day: 13 Month: 1 Year: 2011 Electronic publication date: Day: 23 Month: 2 Year: 2011 collection publication date: Year: 2011 Volume: 5Elocation ID: 9 ID: 3051271 PubMed Id: 21415925 DOI: 10.3389/fnins.2011.00009 
Fitting Neuron Models to Spike Trains  
Cyrille Rossant^{1}^{2}  
Dan F. M. Goodman^{1}^{2}  
Bertrand Fontaine^{1}^{2}  
Jonathan Platkiewicz^{3}  
Anna K. Magnusson^{4}^{5}  
Romain Brette^{1}^{2}*  
^{1}Laboratoire Psychologie de la Perception, CNRS, Université Paris DescartesParis, France 

^{2}Equipe Audition, Département d'Etudes Cognitives, Ecole Normale SupérieureParis, France 

^{3}Institut des Systèmes Intelligents et de Robotique, Université Pierre et Marie CurieParis 06Paris, France 

^{4}Center for Hearing and Communication Research, Karolinska InstitutetStockholm, Sweden 

^{5}Department of Clinical Neuroscience, Karolinska University HospitalStockholm, Sweden 

[editedby] Edited by: Henry Markram, Ecole Polytechnique Fédérale de Lausanne, Switzerland [editedby] Reviewed by: Werner Van Geit, Okinawa Institute of Science and Technology, Japan Correspondence: *Correspondence: romain.brette@ens.fr 
An increasing number of studies in systems neuroscience rely on computational modeling to understand how function emerges from the interaction of individual neurons. Although the neuron models used in these studies are usually well established models [e.g., Hodgkin–Huxley (HH), integrateandfire models, and variations], their parameters are generally gathered from a number of measurements in different neurons and preparations. For example, the model of Rothman and Manis (^{2003}), an established biophysical model of neurons in the cochlear nucleus (CN) of the auditory brainstem, includes ionic channels with properties measured in different neurons of the CN of guinea pigs, combined with a sodium channel and an Ih current derived from previous measurements in mammalian neurons in other areas (including the thalamus). This situation is hardly avoidable for practical reasons, but it raises two questions: (1) channel properties from heterogeneous neuron types, species or ages may not be compatible, and (2) there could be functionally relevant correlations between parameter values within the same neuron type (e.g., maximal conductances), which would be missed if the information about channels came from several independent neurons. Therefore, it seems desirable to obtain models which are fitted for specific neurons.
If these models are to be used for network modeling, then the main goal is to predict the spike trains in response to an arbitrary input. Recently, it was found that simple phenomenological spiking models, such as integrateandfire models with adaptation, can predict the response of cortical neurons to somatically injected currents with surprising accuracy in spike timing (Jolivet et al., ^{2004}; Brette and Gerstner, ^{2005}; Gerstner and Naud, ^{2009}; Kobayashi et al., ^{2009}). A number of techniques have been used to fit specific models to electrophysiological recordings, in particular in the context of the recent INCF Quantitative SingleNeuron Modeling competition (Jolivet et al., 2008), but they are not generic, which limits their practical applicability. This is a difficult optimization problem for two reasons. Firstly, because of the threshold property, any matching criterion between the target spike train and the model spike train is necessarily discontinuous with respect to the model parameters (Brette, ^{2004}), which discards many efficient optimization algorithms. Secondly, a single evaluation of the criterion for a given set of parameter values involves the simulation of a neuron model over a very long time. Therefore, this model fitting problem requires an optimization technique that is both flexible and very efficient.
We have developed a model fitting toolbox (Rossant et al., ^{2010}) for the spiking neural network Brian (Goodman and Brette, ^{2009}). Brian is an opensource simulator^{1} written in Python that lets the user define a model by directly providing its equations in mathematical form. To achieve efficiency, we used vectorization techniques and parallel computing. In particular, optimization can run on a graphics processing unit (GPU), an inexpensive chip available on most modern computers and containing multiple processor cores. In this review, we start by giving an overview of our optimization technique and we demonstrate its use on neural recordings in the barrel cortex and in the auditory brainstem. We also show how a complex multicompartmental model can be reduced to a simple effective spiking model.
Our technique is illustrated in Figure 1. The experimental data consists of currentclamp recordings, with fluctuating currents mimicking in vivo synaptic activity (Figure 1A). In this example (consisting of barrel cortex recordings from the INCF competition), the same current was injected several times in the same neuron (A1 and A2, regular spiking cell) and in different neurons (B1, fast spiking cell). Figure 1B shows the Python script used to run the optimization procedure. The spiking neuron model is defined by its mathematical equations (here, a simple integrateandfire model) and the parameter values are to be optimized so that the model spike trains are as close as possible to the recorded spike trains, which is assessed by a criterion. We chose the gamma factor (Jolivet et al., 2008), used in the INCF competition, which is based on the number of coincidences between the two spike trains, within a fixed temporal window (δ = 4 ms in our figures). Other criteria could be used, but in any case the criterion is a discontinuous function of the model parameters, because model spike trains are themselves discontinuous functions of the parameters. This is a strong constraint: it requires us to use a global optimization algorithm that does not directly use gradient information, for example genetic algorithms. These are computationally intensive, because the criterion must be evaluated on a large number of parameter values, and each evaluation consists of many operations (at least as many as the number of recording samples in the traces). Processing each set of parameters serially is computationally inefficient because (1) in Python, each instruction has a small fixed computational cost (the “interpretation overhead”) which adds up to a substantial cost if each set of parameters is processed separately, and (2) with serial computations we cannot make use of multiple processors. To maximize efficiency without compromising flexibility (the possibility of easily defining any model, as shown in Figure 1B), we developed vectorization techniques, which allows us to use the Brian simulator with minimal interpretation overhead (Brette and Goodman, ^{2010}), and to run the optimization algorithm in parallel. Vectorization consists in simultaneously simulating many neurons defined by the same model but with different parameters, using vector operations to update their variables (that is, using a single Python instruction to perform the same computation on multiple items of data).
Figure 1C illustrates one step of the optimization algorithm. In this figure, we describe the CMAES algorithm (Hansen and Ostermeier, ^{2001}), but other global optimization algorithms can be used (for example we used the particle swarm algorithm in Rossant et al., ^{2010}). A large number of neurons are simulated in a vectorized way: they are defined by the same model but their parameter values are different. The current state of the optimization procedure is specified by a (Gaussian) distribution of parameter values. Parameter values for all the neurons to be simulated are randomly drawn from this distribution. The neural responses to the fluctuating current are simulated and compared to the target experimental spike trains. The best neurons are selected, and their parameter values are used to update the parameter distribution. It is straightforward to simultaneously optimize models for different recordings (e.g., A1, A2, and B1 in Figure 1A): the neuron population is simply subdivided in as many groups as target recordings. When the number of simultaneously simulated neurons is greater than a few thousand neurons, the Brian simulator (which is written in Python, an interpreted language) performs only marginally worse than custom compiled C code (Brette and Goodman, ^{2010}). But this performance can be greatly increased by distributing the optimization over multiple processors.
The most computationally expensive part of the optimization is simulating the neuron population. Since the neurons do not communicate with each other, it is straightforward to distribute their simulations over several processors, as illustrated in Figure 2. One machine acts as a “master” and centralizes the optimization. An iteration starts with the master sending the current parameter distribution to all machines (“workers”). This is a negligible data transfer because the distribution is fully specified by the means and the covariance matrix. The workers independently draw parameter values from the distribution and simulate a population of neurons with these parameters. Each worker then evaluates the performance and selects the best neurons. The parameter values from these local selections are sent back to the master. Again, this is a small data transfer. The master collects all best neurons and updates the parameter distribution, which is sent back to the workers at the next iteration. Since the exchange of information between processors is minimal, the work can be efficiently distributed across several processors in a single machine, or across multiple machines connected over a local network or even over the internet. We use a Python package called Playdoh to distribute the optimization process^{2}, which is based on the standard Python multiprocessing package. Performance scales approximately linearly with the number of processors (Rossant et al., ^{2010}).
Graphics processing units are specialized chips available on most modern computers, which can be used for highperformance parallel computing (Owens et al., ^{2007}). These chips consist of up to 512 processor cores which run separate threads in parallel. Programming GPUs is rather specialized, and in order to obtain maximal performance the structure of the program has to be matched to the design of the hardware. For example, GPUs are designed for single input, multiple data (SIMD) parallelism, meaning that the same operation is simultaneously applied to many different items of data. Memory access is faster when threads read or write contiguous memory blocks. This design is very well adapted to vectorized operations, which we use in our optimization technique. Our model fitting toolbox automatically takes advantage of GPUs, as illustrated in Figure 3. The user writes Python code providing equations as strings, as shown in Figure 1, and the toolbox automatically generates GPU C++ code at runtime (using the techniques discussed in Goodman, ^{2010}), which is compiled and executed on the GPU using the PyCUDA package (Klöckner et al., ^{2009}). With 240core GPUs, we achieved a 50–80×speed improvement using a single GPU (and multiple GPUs can be installed in a single machine or over a cluster for further speed improvements).
Figure 4 shows the application of our procedure to an in vitro intracellular recording of a cortical pyramidal cell from the 2009 Quantitative SingleNeuron Modeling competition (challenge A, first trace). In these recordings, fluctuating currents were injected into the soma and the elicited spike trains were used to fit various models. The most accurate one on the training data was the MATmodel (Kobayashi et al., ^{2009}), which won the INCF competition: it predicted about 86% of reliable spikes (spikes that are repeatedly observed in different trials) with 4 ms precision. It is essentially an integrateandfire model with adaptive threshold: the threshold is a sum of two adaptive components, which increase by a fixed amount after each spike and relax to an equilibrium value with different time constants. Other sorts of integrateandfire models with adaptation (either as an adaptive current or an adaptive threshold) also performed very well (see also Rossant et al., ^{2010}). On the test data, the simpler adaptive integrateandfire model performed better than the MATmodel (79 vs. 66%), which indicates overfitting, but this is presumably because we had to split the competition traces into training and test traces, resulting in little available data for the fitting. Interestingly, more complex models did not perform better. In particular, the adaptive exponential integrateandfire model (Brette and Gerstner, ^{2005}; AdEx) did not give better results although spike initiation is more realistic (FourcaudTrocme et al., ^{2003}; Badel et al., ^{2008}). This surprising result is explained by the fact that the optimized slope factor parameter (δ_{T}) was very small, in fact almost 0 mV, meaning that spike initiation was as sharp as in a standard integrateandfire model. This is consistent with the fact that spikes are sharp at the soma (Naundorf et al., ^{2006}; McCormick et al., ^{2007}), sharper than expected from the properties of sodium channels alone, because of the active backpropagation of spikes from the axon initiation site (Hu et al., ^{2009}; Platkiewicz and Brette, ^{2010}). The Izhikevich (^{2003}) model, a twovariable model with the same qualitative properties as the AdEx model, did not perform as well. This could be because spike initiation is not sharp enough in this model (FourcaudTrocme et al., ^{2003}) or because it is based on the quadratic model, which approximates the response of conductancebased models to constant currents near threshold, while the recorded neurons were driven by current fluctuations. We also fitted the HH model to the response of a fast spiking cortical cell, by optimizing the maximal conductances and the capacitance. The performance was much worse than that of an integrateandfire model (35 vs. 80%), even though the number of free parameters was slightly larger. One possibility is that the channel properties in the HH models did not match those of the cells and should have been optimized as well – although this increases the number of free parameters and therefore the quality of the optimization. But a likely possibility is that the sharpness of spikes cannot be well reproduced by a singlecompartment HH model (Naundorf et al., ^{2006}), even though it can by reproduced by a more complex multicompartmental HH model (McCormick et al., ^{2007}; Hu et al., ^{2009}). This suggests that, as a phenomenological model of cortical neurons, the singlecompartment HH model might be less accurate than a simpler integrateandfire model.
We also applied our model fitting technique to a recording of a neuron in the anteroventral CN of the auditory brainstem (brain slice made from a P12 mouse, see Methods in Magnusson et al., ^{2008}). In this case, it appeared that simple models were less accurate than for cortical neurons (Figure 4B shows a fit to an adaptive exponential integrateandfire model Brette and Gerstner, ^{2005}). One reason might be that neurons in the CN are more specialized than cortical neurons, with specific morphology (Wu and Oertel, ^{1984}) and strong nonlinear conductances (Golding et al., ^{1999}), which make them very sensitive to coincidences in their inputs (McGinley and Oertel, ^{2006}). Interestingly, the performance of a biophysical HH model of CN neurons (Rothman and Manis, ^{2003}) was even worse, when we optimized the maximal conductances of the various ionic channels. It could be that other channel parameters should also be optimized (time constants, reversal potentials, etc.), or that spike initiation is not well reproduced in singlecompartment HH models.
Our model fitting tools can also be used to reduce a complex biophysical model to a simpler phenomenological one, by fitting the simple model to the spike train generated by the complex model in response to a fluctuating input. We show an example of this technique in Figure 5 where a multicompartmental model of a cortical neuron, used in a recent study of spike initiation (Hu et al., ^{2009}), is reduced to an integrateandfire model with adaptive threshold (Platkiewicz and Brette, ^{2010}). In this example, the simpler model predicted 90% of spikes (Γ = 0.90) with a 4ms precision. This surprising accuracy may be due to the fact that active backpropagation of action potentials from the initiation site to the soma makes spike initiation very sharp (McCormick et al., ^{2007}), as in an integrateandfire model. Indeed, it can be seen in Figure 8 of Platkiewicz and Brette (^{2010}), where threshold properties of this multicompartmental model were studied, that spikes are produced precisely when the membrane potential exceeds the (dynamic) threshold.
We have developed efficient parallel algorithms for fitting arbitrary spiking neuron models to electrophysiological data, which are now freely available as part of the Brian simulator^{3}. These algorithms can take advantage of GPUs, which are cheap pieces of hardware available on many desktop PCs and laptops. They can also run on the multiple computers running in a lab, without specific hardware or complex configuration. This computational tool can be used by modelers in systems neuroscience, for example, to obtain empirically validated models for their studies. Because the technique requires only a few minutes of currentclamp recording, another interesting application would be to examine diversity in neural populations, to examine for example the variability and correlations of parameters (maximum conductances, time constants, and so forth). Other model fitting techniques have been previously described by several authors, most of them based on maximum likelihood (Paninski et al., ^{2004}, ^{2007}; Huys et al., ^{2006}), but these are generally modelspecific whereas our approach is generic. Besides, maximum likelihood techniques are designed for cases when neuron responses are very variable between trials and can only be described in a probabilistic framework. On the contrary, the optimization approach we chose is best suited for currentclamp recordings in slices, in which neural responses are precisely reproducible (Mainen and Sejnowski, ^{1995}).
Our results confirm that integrateandfire models with adaptation are a good description of the response of cortical neurons to somatically injected currents. Complex multicompartmental models of cortical neurons could also be accurately reduced to such simple models. This is surprising for two reasons. Firstly, neurons have many ionic channels with specific properties and it would be surprising that they are not relevant for the input–output properties of neurons. However, it is known that detailed conductancebased models with widely diverse ion channel characteristics can in fact have the same properties at the neuron level (Goldman et al., ^{2001}). Our technique only produces a phenomenological or “effective” description of neural responses, without trying to explicitly model all the contributing ionic mechanisms. We can speculate that the presence of a variety of ionic channels makes it possible for the cell to tune its integrative properties through the action of modulation or intrinsic plasticity, which are not included in the effective description. Another important aspect to bear in mind is that we only addressed the response of neurons to somatically injected currents, while dendritic properties are certainly very important for the function of cortical neurons (Stuart et al., ^{1999}). Interestingly, simple models did not perform as well at predicting the spike trains of neurons in the auditory brainstem, presumably because they express strong nonlinear ionic channels, e.g., Ih (Rothman and Manis, ^{2003}). Secondly, spike initiation in integrateandfire models is very sharp, unlike in HH models, and this is known to impact the response of neuron models to fluctuating inputs (FourcaudTrocme et al., ^{2003}), so it might seem surprising that integrateandfire models predict the responses of cortical neurons so well. However, in real cortical neurons, spike initiation is in fact very sharp, unlike in singlecompartment HH models (Naundorf et al., ^{2006}; Badel et al., ^{2008}). This property results from the active backpropagation to the soma of spikes initiated in the axon hillock (McCormick et al., ^{2007}). Complex multicompartmental HH models can reproduce this property, as well as threshold variability (Hu et al., ^{2009}; Platkiewicz and Brette, ^{2010}), but singlecompartment ones cannot. This explains why integrateandfire models are surprisingly effective descriptions of spike initiation in cortical neurons – if adaptation is also included, in particular threshold adaptation (Platkiewicz and Brette, ^{2010}).
Our model fitting toolbox can be extended in several ways. Different optimization algorithms could be implemented, but more interestingly different error criterions could be chosen. For example, one could fit the intracellular voltage trace rather than the spike trains, or try to predict the value of the spike threshold (Azouz and Gray, ^{2000}; Platkiewicz and Brette, ^{2010}). Finally, although our technique primarily applies to neural responses to intracellular current injection, it could in principle be applied also to extracellularly recorded responses to timevarying stimuli (e.g., auditory stimuli), if spike timing is reproducible enough, for example in the bushy cells of the CN (Louage et al., ^{2005}).
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Notes
fn1^{1}http://briansimulator.org
fn2^{2}http://code.google.com/p/playdoh/
fn3^{3}http://briansimulator.org
This work was supported by the European Research Council (ERC StG 240132) and by Vetenskapsrådet (grant no. 80326601).
References
Azouz R.,Gray C.. (Year: 2000). Dynamic spike threshold reveals a mechanism for synaptic coincidence detection in cortical neurons in vivo. Proc. Natl. Acad. Sci. U.S.A.97, 811010859358  
Badel L.,Lefort S.,Brette R.,Petersen C. C. H.,Gerstner W.,Richardson M. J. E.. (Year: 2008). Dynamic IV curves are reliable predictors of naturalistic pyramidalneuron voltage traces. J. Neurophysiol.99, 656–66610.1152/jn.01107.200718057107  
Brette R.. (Year: 2004). Dynamics of onedimensional spiking neuron models. J. Math. Biol.48, 38–5614685771  
Brette R.,Gerstner W.. (Year: 2005). Adaptive exponential integrateandfire model as an effective description of neuronal activity. J. Neurophysiol.94, 3637–364210.1152/jn.00686.200516014787  
Brette R.,Goodman D. F.. (Year: 2010). Vectorised algorithms for spiking neural network simulation. Neural. Comput. (in press).  
FourcaudTrocme N.,Hansel D.,van Vreeswijk C.,Brunel N.. (Year: 2003). How spike generation mechanisms determine the neuronal response to fluctuating inputs. J. Neurosci.23, 11628–1164014684865  
Fujino K.,Oertel D.. (Year: 2001). Cholinergic modulation of stellate cells in the mammalian ventral cochlear nucleus. J. Neurosci.21, 7372–738311549747  
Gerstner W.,Naud R.. (Year: 2009). How good are neuron models?Science326, 379–38010.1126/science.118193619833951  
Golding N. L.,Ferragamo M. J.,Oertel D.. (Year: 1999). Role of intrinsic conductances underlying responses to transients in octopus cells of the cochlear nucleus. J. Neurosci.19, 2897–290510191307  
Goldman M. S.,Golowasch J.,Marder E.,Abbott L. F.. (Year: 2001). Global structure, robustness, and modulation of neuronal models. J. Neurosci.21, 5229–523811438598  
Goodman D.,Brette R.. (Year: 2009). The Brian simulator. Front. Neurosci.3:219753088  
Goodman D. F. M.. (Year: 2010). Code generation: a strategy for neural network simulators. Neuroinformatics8, 183–19610.1007/s120210109082x20857234  
Hansen N.,Ostermeier A.. (Year: 2001). Completely derandomized selfadaptation in evolution strategies. Evol. Comput.9, 159–19511382355  
Hu W.,Tian C.,Li T.,Yang M.,Hou H.,Shu Y.. (Year: 2009). Distinct contributions of na(v)1.6 and na(v)1.2 in action potential initiation and backpropagation. Nat. Neurosci.12, 996–100219633666  
Huys Q. J. M.,Ahrens M. B.,Paninski L.. (Year: 2006). Efficient estimation of detailed singleneuron models. J. Neurophysiol.96, 872–89010.1152/jn.00079.200616624998  
Izhikevich E. M.. (Year: 2003). Simple model of spiking neurons. IEEE Trans. Neural Netw.14, 1569–157210.1109/TNN.2003.82044018244602  
Jolivet R.,Kobayashi R.,Rauch A.,Naud R.,Shinomoto S.,Gerstner W.. (Year: 2008). A benchmark test for a quantitative assessment of simple neuron models. J. Neurosci. Methods169, 417–42418160135  
Jolivet R.,Lewis T. J.,Gerstner W.. (Year: 2004). Generalized integrateandfire models of neuronal activity approximate spike trains of a detailed model to a high degree of accuracy. J. Neurophysiol.92, 959–97610.1152/jn.00190.200415277599  
Jolivet R.,Schrmann F.,Berger T. K.,Naud R.,Gerstner W.,Roth A.. (Year: 2008). The quantitative singleneuron modeling competition. Biol. Cybern.99, 417–42610.1007/s004220080261x19011928  
Klöckner A.,Pinto N.,Lee Y.,Catanzaro B.,Ivanov P.,Fasih A.,Sarma A. D.,Nanongkai D.,Pandurangan G.,Tetali P.. (Year: 2009). PyCUDA: GPU runtime code generation for highperformance computing. Arxiv preprint arXiv09113456  
Kobayashi R.,Tsubo Y.,Shinomoto S.. (Year: 2009). Madetoorder spiking neuron model equipped with a multitimescale adaptive threshold. Front. Comput. Neurosci.3:910.3389/neuro.10.009.200919668702  
Louage D. H. G.,van der Heijden M.,Joris P. X.. (Year: 2005). Enhanced temporal response properties of anteroventral cochlear nucleus neurons to broadband noise. J. Neurosci.25, 1560–157010.1523/JNEUROSCI.474204.200515703410  
Magnusson A. K.,Park T. J.,Pecka M.,Grothe B.,Koch U.. (Year: 2008). Retrograde GABA signaling adjusts sound localization by balancing excitation and inhibition in the brainstem. Neuron59, 125–13710.1016/j.neuron.2008.05.01118614034  
Mainen Z.,Sejnowski T.. (Year: 1995). Reliability of spike timing in neocortical neurons. Science268, 15037770778  
McCormick D. A.,Shu Y.,Yu Y.. (Year: 2007). Neurophysiology: Hodgkin and Huxley modelstill standing?Nature445, E12; discussion E2–E3.10.1038/nature0552317287765  
McGinley M. J.,Oertel D.. (Year: 2006). Rate thresholds determine the precision of temporal integration in principal cells of the ventral cochlear nucleus. Hear. Res.216–217, 52–6310.1016/j.heares.2006.02.00616597491  
Naundorf B.,Wolf F.,Volgushev M.. (Year: 2006). Unique features of action potential initiation in cortical neurons. Nature440, 1060–106310.1038/nature0461016625198  
Owens J. D.,Luebke D.,Govindaraju N.,Harris M.,Kruger J.,Lefohn A. E.,Purcell T. J.. (Year: 2007). A survey of generalpurpose computation on graphics hardware. Comput. Graph. Forum26, 80–113  
Paninski L.,Pillow J.,Lewi J.. (Year: 2007). Statistical models for neural encoding, decoding, and optimal stimulus design. Prog. Brain Res.165, 493–50710.1016/S00796123(06)65031017925266  
Paninski L.,Pillow J. W.,Simoncelli E. P.. (Year: 2004). Maximum likelihood estimation of a stochastic integrateandfire neural encoding model. Neural. Comput.16, 2533–256115516273  
Platkiewicz J.,Brette R.. (Year: 2010). A threshold equation for action potential initiation. PLoS Comput. Biol.6, e100085010.1371/journal.pcbi.100085020628619  
Rossant C.,Goodman D. F. M.,Platkiewicz J.,Brette R.. (Year: 2010). Automatic fitting of spiking neuron models to electrophysiological recordings. Front. Neuroinformatics4:210.3389/neuro.11.002.201020224819  
Rothman J. S.,Manis P. B.. (Year: 2003). The roles potassium currents play in regulating the electrical activity of ventral cochlear nucleus neurons. J. Neurophysiol.89, 3097–311310.1152/jn.00127.200212783953  
Stuart G.,Spruston N.,Hausser M.. (Year: 1999). Dendrites. New York: Oxford University Press  
Wu S. H.,Oertel D.. (Year: 1984). Intracellular injection with horseradish peroxidase of physiologically characterized stellate and bushy cells in slices of mouse anteroventral cochlear nucleus. J. Neurosci.4, 1577–15886726347 
Phenomenological spiking model  An effective neuron model with accurate input–output properties (the output being spike trains), but where the underlying biophysical mechanisms are not explicitly represented (e.g., integrateandfire model), as opposed to a biophysical model (e.g., Hodgkin–Huxley model). 
Python  Python is a highlevel programming language. Programs can be run from a script or interactively from a shell (as in Matlab). It is often used for providing a highlevelinterface to low level code. The Python community has developed a large number of third party packages, including NumPy and SciPy which are commonly used for efficient numerical and scientific computation. 
Vectorization  Vectorization is a technique for achieving computational efficiency in highlevel languages. It consists of replacing repeated operations by single operations on vectors (e.g., arithmetic operations) that are implemented in a dedicated efficient package (e.g., NumPy for Python). 
Graphics processing units  Graphics processing units or GPUs are specialized chips available on most modern computers (originally for graphics rendering), which can be used for highperformance parallel computing. These chips consist of up to 512 processor cores working in parallel. 
Global optimization  Global optimization is the minimization or maximization of a function of several variables, using algorithms such as genetic algorithms, particle swarm optimization, differential evolution. 
Romain Brette obtained a PhD in Computational Neuroscience from the Paris VI University in France. He did postdoctoral studies in Alain Destexhe's group in GifsurYvette (France) and Wulfram Gerstner's group in Lausanne (Switzerland). He is now an associate professor of computational neuroscience at Ecole Normale Supérieure, Paris. His group investigates spikebased neural computation, theory, and simulation of spiking neuron models, with a special focus on the auditory system.
Figures
Article Categories:
Keywords: python, spiking models, simulation, optimization, parallel computing. 
Previous Document: Novel drugeluting stents in the treatment of de novo coronary lesions.
Next Document: Doselevels and first signs of efficacy in contemporary oncology phase 1 clinical trials.