# Shifting Spike Times or Adding and Deleting Spikes—How Different Types of Noise Shape Signal Transmission in Neural Populations

- Sergej O Voronenko
^{3, 4}Email author, - Wilhelm Stannat
^{3, 5}Email author and - Benjamin Lindner
^{3, 4}Email author

**5**:1

https://doi.org/10.1186/2190-8567-5-1

© S.O. Voronenko et al.; licensee Springer 2015

**Received: **15 April 2014

**Accepted: **17 November 2014

**Published: **12 January 2015

## Abstract

We study a population of spiking neurons which are subject to independent noise processes and a strong common time-dependent input. We show that the response of output spikes to independent noise shapes information transmission of such populations even when information transmission properties of single neurons are left unchanged. In particular, we consider two Poisson models in which independent noise either (i) adds and deletes spikes (AD model) or (ii) shifts spike times (STS model). We show that in both models *suprathreshold stochastic resonance* (SSR) can be observed, where the information transmitted by a neural population is increased with addition of independent noise. In the AD model, the presence of the SSR effect is robust and independent of the population size or the noise spectral statistics. In the STS model, the information transmission properties of the population are determined by the spectral statistics of the noise, leading to a strongly increased effect of SSR in some regimes, or an absence of SSR in others. Furthermore, we observe a high-pass filtering of information in the STS model that is absent in the AD model. We quantify information transmission by means of the lower bound on the mutual information rate and the spectral coherence function. To this end, we derive the signal–output cross-spectrum, the output power spectrum, and the cross-spectrum of two spike trains for both models analytically.

### Keywords

Time-dependent input Population coding Common noise Shifting of spikes Addition and deletion of spikes Mutual information Suprathreshold stochastic resonance## 1 Introduction

Neurons in the sensory periphery encode information about continuous time-dependent signals in sequences of action potentials. Hereby, upon repeated presentation of a stimulus, the response of the neuron is not perfectly reproducible but exhibits trial-to-trial variability. Processes, leading to such variability, are termed *noise* and can have various origins [1, 2]. How such noise processes affect the transmission of time-dependent signals in neurons can be studied in the framework of information theory [3, 4]. Within this framework, it has been shown, for instance, that the presence of noise can enhance the transmission of weak (subthreshold) signals in single neurons and neural models [5–7], an effect known as *stochastic resonance* and also observed outside biology [8, 9]. At the level of neural population coding, noise can also have a beneficial role for the transmission of strong (suprathreshold) signals [10, 11] by means of *suprathreshold stochastic resonance* (SSR), the mechanism of which is quite distinct from that of conventional stochastic resonance despite the similarity in their naming. Additionally, noise not only impacts the total transmitted information, but it also affects which frequencies of the sensory signal are preferably encoded by a neural system. The suppression of information about the input signal in certain frequency bands can be regarded as a form of information filtering [12–16]. Put differently, we may ask whether the neural system is preferentially encoding slow (low-frequency) components of a signal or fast (high-frequency) components of a signal, which can be quantified by the coherence function, as described below.

*Drosophila*olfactory system [25]. In the present study, we consider ensembles of neurons receiving highly correlated noise input as sketched in Fig. 1.

This work is organized as follows: First, we describe the methods by which we will study the effect of noise on signal transmission in a population of spiking neurons. Second, we introduce two models where independent noise either adds and deletes spikes, or shifts spike times in the output spike trains. In Sect. 4, we then derive the spectral statistics for the two models. These derivations can be skipped upon the first reading. In Sect. 5, we summarize the derived spectral statistics and proceed to study the effect of independent noise on information filtering and the total transmission of information in neural populations. We conclude with a summary and a discussion of our results in Sect. 6.

## 2 Methods

### 2.1 Spike Train Statistics & Ensemble Averages

*μ*th stochastic point process can be described by the spike count . This function starts at 0 at and is incremented by 1 at each spike time , i.e. for , for , and so forth. Equivalently, the output of a stochastic point process can be described by the derivative of . This derivative is called the spike train and is given by a sum of delta functions,

of the individual output spike trains.

*ξ*, conditioned on a realization of

*s*and

*η*, whereas stands for the total expectation. Note that is still a random process, unless a realization of

*s*and

*η*is fixed. Below, when analyzing correlation functions, e.g. Eq. (7), we will also consider averages over products of spike trains , which in mathematical terms corresponds to

This applies analogously to averages over the processes and .

*only*with respect to the common noise , will be an important quantity in our calculations. It still depends on the independent noise and the signal and is difficult to determine in experiments. More accessible is the average over all noise sources by repeated trials with a frozen stimulus and summation over all spike trains. In this way, we obtain (apart from a normalization factor ) the population rate

### 2.2 Information Transmission & Spectral Statistics

*mutual information rate*

*R*[3], which is measured in bits per second. For Gaussian signals, a lower bound on the mutual information rate [4, 31, 32] is given by

From Eq. (10) we see that for
the cross-spectrum of two spike trains,
, appears in the denominator of the coherence function and gains significance as *N* becomes larger. Therefore, an essential theoretical problem is to calculate this cross-spectrum.

As outlined above, the coherence function allows one to estimate the total flow of information through the neural population. However, because enters in a monotonic fashion in Eq. (4), we can also regard the coherence as a frequency-resolved measure of information transfer. Reduction of the coherence in certain frequency bands can be regarded as a form of information filtering, which needs to be distinguished from power filtering. Hence, besides the lower bound , we will also inspect the frequency dependence of the coherence function.

## 3 Models

- (1)
Poisson statistics of spontaneous activity;

- (2)
high correlations among neurons due to strong common noise input;

- (3)
encoding of a sensory signal in the time-dependent population rate.

For simplicity, we consider a linear encoding of a weak time-dependent signal. This will allow us to use the lower bound on the mutual information rate as an approximation for the total transmitted information. Note that, although already a single Poisson process can show conventional stochastic resonance [35], with our linear encoding paradigm we exclude this possibility. In our models, the signal transmission in a single neuron is always degraded by noise.

In our theoretical model, we assume that all neurons fire to zeroth-order in complete synchrony and a weak noise input, which is independent for every neuron, leads to a decorrelation of the output spike trains. For simplicity, we assume that for each neuron the independent noise process and the sensory signal are additive. Both, the sensory signal and the independent noise signals, are modeled by Gaussian processes with unit variance and zero mean.

The considered models can be regarded as *inhomogeneous Poisson processes* [36], which are rate-modulated by a common signal
and an independent noise
. Such processes are examples of a *doubly stochastic process* [37] or a *Cox process* and are a special case of the *inhomogeneous Bernoulli process* [38]. The simplicity of the considered models will allow us to characterise the information transfer of weak time-dependent signals analytically. Note that the assumptions (1)–(3) made above describe, in good approximation, spiking in specific sensory systems, e.g. in tangential neurons of the fly visual system [20, 39, 40]. The additional modifications that make up the differences between our two models can be regarded as additional operations on the spike trains in the form of thinning (or the opposite of it) and the introduction of an operational time [37, 41].

Before we introduce in detail the two models sketched in Fig. 2, it is worth to note that, for weak stimuli and weak independent noise, these models possess the same signal–output cross-spectrum , the same power spectrum , and the same time-dependent output firing rate. Therefore, for the coherence function and the information rate are identical for both models. The models are mainly distinguished by how independent noise affects the spikes of the output spike trains, which results in different cross-spectra of two spike trains. This setup allows us to study how the response of spikes to noise affects information transmission in neural populations, while keeping all other potential influences on signal transmission unchanged.

### 3.1 Addition and Deletion Model (AD Model)

*t*. We generate a spike in the

*j*th time bin in the

*μ*th spike train, whenever the following condition is fulfilled:

*ξ*is uniformly distributed in and uncorrelated in time. The spikes are assigned the height such that the discrete spike train reads

where
is the Heaviside function (implementing the indicator function) and the second argument of
indicates the time-discretized version of the spike train. Here
is the midpoint of the time bin where the *k* th spike of the *μ* th spike train was generated. In the limit
the spike train
approximates the sum of *δ*-functions
given by Eq. (1).

*s*and . As we show explicitly in Appendix A in Eq. (57), averaging additionally over the independent noise and the signal, one finds in the limit of

Throughout the paper, we will consider the limit , such that we can neglect correction terms like the one in the above equation.

In the left column of Fig. 3, we show how a sensory signal is encoded in the population firing rate of a population of five AD neurons and how the output spike trains of the neurons are modulated by independent noise.

### 3.2 Spike-Time-Shifting Model (STS Model)

*N*neurons of the population generate identical spike trains

*μ*th neuron, the times are transformed into new spike times via the transformation

with
defined in Eq. (11). For a given spike time
, we integrate the right hand side of Eq. (16), until the integral attains the value
[36]. The resulting integration boundary
is then the *k* th spike time of the *μ* th spike train
. In general, due to the different independent noise processes
, the output spike trains
will be different for each neuron. Hereby, each spike train is an inhomogeneous Poisson spike train with a time-dependent firing rate. The procedure described in this section is equivalent to the simulation of a perfect integrate-and-fire neuron with exponentially distributed thresholds [36]. The time *t* obtained after the transformation of the time axis *h* in Eq. (16) is also known as operational time [37, 41].

Although we do not model the underlying noise process explicitly, we think of the homogeneous spike trains in Eq. (15) as a result of a common noise process *ξ*, analogously to the AD model. By the average
, we will denote the average over different realizations of the homogeneous Poisson spike trains in Eq. (15).

which in the limit leads to the same mean firing rate as for the AD model Eq. (14) in the limit of .

A simulation of five spike trains of the STS population, driven by a common noise process *ξ*, a common signal *s*, and independent noise processes
, is shown Fig. 3e. Note that the modulation in Eq. (16) is very distinct from adding jitter to the single spike times, as is considered in [42–44], in that the modulation of the spike times presented here preserves the order of the spikes in each spike train. Other models that incorporate the deletion of spikes in a Poisson spike train [45] or a combination of deletion and shifting as in the thinning and shifting model [42, 44], differ from the models presented here in that the single spike trains of those models are homogeneous spike trains with constant rates. However, the models in the present paper are designed such that the single spike trains have a prescribed time-dependent firing rate
, which still depends on the realization of the signal *s* and the individual noise *η*. The cross-correlations between spike trains are a consequence of the different implementations of the time-dependent firing rate and are not prescribed a priori as in [42, 44, 45]. Even if the deletion or shifting of spikes in the thinning and shifting model is performed on a rate-modulated mother process, the resulting process would not be equivalent to the AD model or STS model, in which the addition and deletion of spikes and the shifting of spike times are not independent of the signal realization. In particular, the thinning and shifting model of a population of daughter processes for which the stimulus is solely encoded in the firing rate of the mother process cannot exhibit suprathreshold stochastic resonance.

### 3.3 Modeling the Common Signal and the Independent Noise Processes

*s*and the independent noise sources are modeled by Gaussian stochastic processes with zero mean and unit variance. For simplicity, we choose for both, the signal and the independent noise, a flat power spectrum,

where
and
are lower and upper cutoff frequencies, respectively. Throughout the paper, we will consider a finite upper cutoff frequency and a non-vanishing lower cutoff frequency. As we will show in our analytical calculation below, the cross-spectrum for two spike trains of the STS model is finite only for
. A realization of the common signal *s* is shown in Fig. 3a and 3d.

### 3.4 Simulations

*t*, the total simulation time

*T*, and the number of realizations used for the numerical averaging of the spectral statistics are reported in Table 1.

**Parameters used in numerical simulations**

Figure | Δ |
| |
---|---|---|---|

Fig. 3 | 1⋅10 | 1⋅10 | 5⋅10 |

Fig. 4 | 5⋅10 | 1⋅10 | 5⋅10 |

Fig. 5 | 1⋅10 | 1⋅10 | 5⋅10 |

Fig. 6a | 5⋅10 | 3⋅10 | 1⋅10 |

Fig. 6b | 3⋅10 | 2⋅10 | 1⋅10 |

Fig. 6c | 1⋅10 | 3⋅10 | 1⋅10 |

Fig. 6d | 3⋅10 | 1⋅10 | 2⋅10 |

Fig. 7a STS | 5⋅10 | 2⋅10 | 2⋅10 |

Fig. 7a AD | 5⋅10 | 3⋅10 | 1⋅10 |

Fig. 7b STS | 5⋅10 | 1⋅10 | 2⋅10 |

Fig. 7b AD | 1⋅10 | 3⋅10 | 1⋅10 |

Fig. 8 | 2⋅10 | 1⋅10 | 2⋅10 |

Fig. 9 | 2⋅10 | 1⋅10 | – |

## 4 Derivation of Spectral Measures

### 4.1 Input–Output Cross-spectrum

*s*and are Gaussian processes with unit variance and zero mean, which leads to

which is equal for both models.

### 4.2 Cross-spectrum for Two Spike Trains for the AD Model

*ξ*, the common signal

*s*, and the independent noise processes and . Employing Eq. (14), we can write the second term in Eq. (20) as

*t*. As described in Sect. 3.1, the values of realizations of the process

*ξ*at different times are independent of each other, which allows us to average both spike trains separately leading to

*τ*, the cross-spectrum is real-valued for all frequencies.

### 4.3 Cross-spectrum for Two Spike Trains for the STS Model

*μ*and

*ν*for the STS model. We first consider the autocorrelation function

*ξ*, independent noise processes and , and an input signal

*s*. Employing Eq. (17), the last term in the above equation can be written as

*g*is a sum of two integrals over Gaussian variables and therefore also a Gaussian variable. The average of the delta function over realizations of

*g*is then the probability that

*g*attains the value , and is given by

We note that the linear term in vanishes due to the zero mean of the Gaussian signal . Equivalently, all higher-order odd terms in in Eq. (37) vanish due to the Gaussian nature of the signal (except for the correction term due to realizations of signal and individual noise that lead to ). From Eqs. (36) and (37) it can be seen that for a vanishing lower cutoff frequency of the independent noise spectrum ( ), the variance diverges and as a consequence of this the cross-correlation between the two spike trains vanishes—only the part that is due to the signal (second term in Eq. (37)) still contributes.

In Fig. 4, the analytical result for the cross-spectrum for two spike trains of the STS model Eq. (38) is compared with simulations. As for the AD model the cross-spectrum of two spike trains is real valued. In contrast to the AD model Eq. (25), the cross-spectrum of two spike trains for the STS model Eq. (38) exhibits a strong decrease at high frequencies, while it approaches the spike train power spectrum Eq. (39) at low frequencies. Note that, although we derived only up to second-order in , the theory fits the simulation results very well even for .

### 4.4 Single Spike Train Power Spectrum

For and , the power spectrum is flat, as we would expect for homogeneous Poisson spike trains.

## 5 Information Transmission in Neural Populations

in the analytical calculations to obtain simpler expressions. In the subsequent sections, we will study information transmission in populations of AD neurons and STS neurons.

### 5.1 AD Population

The linear term in Eq. (44) is always positive. Hence, the population of AD neurons always profits from weak independent noise regardless of the specific choice of model parameters.

### 5.2 STS Population

where and are defined in Eq. (38). As for the AD model discussed above, we used that signal and noise have equal power-spectra . Due to the frequency dependence of the cross-spectrum , the coherence function also depends strongly on the frequency and exhibits a monotone increase as shown in Fig. 5. Thus, the population of STS neurons can be regarded as a high-pass filter of information, similar to that observed for heterogeneous short-term plasticity [16] or coding by synchrony [13, 15].

In order to understand the high-pass filter effect in the coherence function as well as the stochastic resonance effect discussed below, we note that the cross-correlations between different spike trains contribute largely to the sum’s output variability, in particular in the absence of intrinsic noise. This output variability is quantified by the output’s power spectrum and appears in the denominator of the coherence function. With individual intrinsic noise, spike times of different neurons are slightly shifted, drastically reducing cross-correlations at high frequencies and thus the amount of the signal-unrelated variability in these frequency bands. Therefore, the coherence function increases with frequency.

The lower bound on the mutual information rate for the STS population is compared with simulations for two sets of parameters in Fig. 6. We observe that for the given parameters the STS model shows a large SSR effect, while the AD model profits only weakly from additional noise.

*N*. The second-order term in Eq. (48) can attain both negative and positive values depending on the choice of the model parameters. The condition that the second-order term becomes negative and that the lower bound on the mutual information rate at is a decreasing function of reads

for which the lower bound on the mutual information rate is equal for both models. From the above equation we can see that whether the STS population or the AD population transmits more information for a given value of independent noise is mainly determined by the noise and signal cutoff frequencies and .

*N*, all parameters as in Fig. 6d). In this situation, we consider the low-pass filtered summed output of the population for different levels of the intrinsic noise. Without intrinsic noise (Fig. 9a), the output, i.e. the sum of

*N*perfectly synchronized spike trains, does not resemble the input signal very much. It is important to note that according to Eq. (19) and Eq. (9) an average over many such runs would yield a time series that tracks the input signal closely. However, single runs (red, black, green) in the absence of the intrinsic noise are strongly unreliable. The right amount of intrinsic noise (used in Fig. 9b) desynchronizes the

*N*spike trains, reduces cross-correlations at high frequencies, and thus reduces output variability due to the common noise. Consequently, different realizations of the process for a frozen input signal look more similar and track the input signal reliably (cf. Fig. 9b). However, if we increase intrinsic noise to much higher levels, as in Fig. 9c, this noise itself starts to contribute significantly to the output variability and the reliability of signal transmission is diminished again.

## 6 Summary and Conclusions

In this paper, we investigated how the effect of noise on the output spikes influences information transmission properties of Poisson neurons. In particular, we considered two populations with strong common input, where in one case weak independent noise added and deleted spikes, while in the other it shifted spikes. In the limit of a weak sensory signal, we analytically derived the spectral statistics of both models and studied information filtering and the emergence of suprathreshold stochastic resonance (SSR). We showed that, even when single neurons of the AD model and STS model cannot be distinguished by their response statistics, the different effects of independent noise on spikes lead to qualitative and quantitative differences in information transmission on a population level.

In the AD model, the presence of the SSR effect is robust—whenever we consider a population with , a small amount of intrinsic noise has a beneficial effect on the signal transmission. In the STS model, the information transmission properties of the population are determined by the cutoff frequencies of the noise. Depending on the specific parameters, one finds a pronounced SSR in some regimes (exceeding the effect in the AD model by far) or no SSR effect in other regimes. Furthermore, we observe a high-pass filtering of information in the STS model that is absent in the AD model.

There are a number of studies that explored theoretically the case of weakly correlated neurons and employed perturbation methods to relate output spike train correlations to input correlations [46–52]. In this paper, we have considered the opposite limit of strongly correlated spike trains that are only weakly *decorrelated* due to intrinsic noise sources. In this limit, we were not only able to derive comparatively simple expressions for the cross-correlation between two spike trains but were also able to explore analytically the consequences of these correlations for the transmission of time-dependent signals.

The question arises how the specific choice of the output, which is taken to be the sum of individual spike trains, affects the findings discussed above. The most general approach would be to study the multivariate mutual information between the input signal and the population of output spike trains. This quantity is hard to compute numerically and analytically, and its exact calculation is beyond the scope of this study. However, the mutual information between the input signal and the sum of outputs is a lower bound for the full multivariate mutual information, because the summation can only degrade the information content contained in the entire set of the output spike trains. Additionally, for vanishing individual noise, , all output spike trains are identical and the information content of the population does not differ from the information content of the sum of identical spike trains. Therefore, if the mutual information between the input signal and the summed output increases with individual noise, i.e. exhibits suprathreshold stochastic resonance, the full multivariate mutual information increases as well.

The mutual information between the input signal and the summed output has been estimated here by its lower bound . In our setting with a weak signal that is encoded in the firing rate of the Poisson process, we expect that this bound is rather tight. In fact, for a single inhomogeneous Poisson process, the mutual information and its lower bound coincide in leading-order of the signal amplitude [53].

In this study, we inspected two simple and abstract models for the effect of a weak noise on neural spikes and its consequences on signal transmission by neural populations. We would like to emphasize that the pure limits of an AD model or an STS model approximate the behavior of biophysical neuron models. On one hand, it is plausible that in an excitable neuron model, in which the crossing of a threshold may be aided or prevented by a weak driving, addition and deletion of spikes as in our AD model can be observed. Stochastic oscillators, on the other hand, display a shifting of spike times due to a weak driving, as described by the phase response curve [54]. In between these limits, we expect a combination of both, addition and deletion as well as shifting of spikes. Indeed, such a combination has been observed experimentally [55]. Hence, a generalization of our framework to a Poisson process that includes both effects and allows one to tune gradually between the pure AD and STS models inspected in this paper would be certainly worth additional efforts in a future study.

## Appendix A: Mean Firing Rate of the AD and STS Model

*s*and

*η*are Gaussian processes with unit variance and zero mean. For the other terms in Eq. (49), we will show that they are of higher-order in Δ

*t*, , and . For the second term in Eq. (49), we can find an upper bound using the Cauchy–Schwarz inequality

A similar estimation leads to the same formula for the STS model.

## Appendix B: Probability for Synchronous Spikes in the AD Model

*t*in two spike trains and is given by

## Appendix C: Simplification of the Cross-correlation Function for the STS Model

used in the calculation of the cross-correlation function in Eq. (34).

## Appendix D: Variance of the Integrated Independent Noise

## Appendix E: Single Spike Train Power Spectrum

where we assume that the spike trains are stationary. Since a single spike train is considered and the average is taken over one independent noise process *η*, we will drop the subscript employed previously.

*t*and is the probability to find two spikes separated by the time interval

*τ*with . Analogously to Eq. (23) we find

*r*, which is defined in Eq. (11). For the probability of asynchronous spikes with we find

*η*. The above equation can be rewritten as

*W*, is Fourier transformed into

The order of the correction term in the above equation is proportional to the square root of the probability that , which has been calculated in Appendix A Eq. (52).

## Declarations

### Acknowledgements

This work was funded by the BMBF (FKZ:01GQ1001A and FKZ:01GQ1001B).

## Authors’ Affiliations

## References

- Tuckwell HC:
*Stochastic Processes in the Neuroscience*. 1989.View ArticleMATHGoogle Scholar - Faisal AA, Selen LPJ, Wolpert DM:
**Noise in the nervous system.***Nat Rev Neurosci*2008,**9:**292.View ArticleGoogle Scholar - Shannon R:
**The mathematical theory of communication.***Bell Syst Tech J*1948,**27:**379.MathSciNetView ArticleMATHGoogle Scholar - Rieke F, Warland D, de Ruyter van Steveninck R, Bialek W:
*Spikes: Exploring the Neural Code*. MIT Press, Cambridge; 1999.MATHGoogle Scholar - Wiesenfeld K, Moss F:
**Stochastic resonance and the benefits of noise: from ice ages to crayfish and SQUIDs.***Nature*1995,**373:**33.View ArticleGoogle Scholar - Hänggi P:
**Stochastic resonance in biology.***ChemPhysChem*2002,**21:**285-290.View ArticleGoogle Scholar - McDonnell MD, Ward LM:
**The benefits of noise in neural systems: bridging theory and experiment.***Nat Rev Neurosci*2011,**12:**415.View ArticleGoogle Scholar - Gammaitoni L, Hänggi P, Jung P, Marchesoni F:
**Stochastic resonance.***Rev Mod Phys*1998,**70:**223.View ArticleGoogle Scholar - Lindner B, Garcia-Ojalvo J, Neiman A, Schimansky-Geier L:
**Effects of noise in excitable systems.***Phys Rep*2004,**392:**321.View ArticleGoogle Scholar - Stocks NG:
**Suprathreshold stochastic resonance in multilevel threshold systems.***Phys Rev Lett*2000,**84:**2310.View ArticleGoogle Scholar - Stocks NG, Mannella R:
**Generic noise-enhanced coding in neuronal arrays.***Phys Rev E*2001.,**64:**Article ID 030902Google Scholar - Chacron MJ, Doiron B, Maler L, Longtin A, Bastian J:
**Non-classical receptive field mediates switch in a sensory neuron’s frequency tuning.***Nature*2003,**423:**77.View ArticleGoogle Scholar - Middleton JW, Longtin A, Benda J, Maler L:
**Postsynaptic receptive field size and spike threshold determine encoding of high-frequency information via sensitivity to synchronous presynaptic activity.***J Neurophysiol*2009,**101:**1160.View ArticleGoogle Scholar - Lindner B, Gangloff D, Longtin A, Lewis JE:
**Broadband coding with dynamic synapses.***J Neurosci*2009,**29:**2076.View ArticleGoogle Scholar - Sharafi N, Benda J, Lindner B:
**Information filtering by synchronous spikes in a neural population.***J Comput Neurosci*2013,**34:**285.MathSciNetView ArticleMATHGoogle Scholar - Droste F, Schwalger T, Lindner B:
**Interplay of two signals in a neuron with heterogeneous synaptic short-term plasticity.***Front Comput Neurosci*2013.,**7:**Article ID 86Google Scholar - Knight BW:
**Dynamics of encoding in a population of neurons.***J Gen Physiol*1972,**59:**734.View ArticleGoogle Scholar - Gerstner W, Kistler WM:
*Spiking Neuron Models*. Cambridge University Press, Cambridge; 2002.View ArticleMATHGoogle Scholar - Alonso JM, Usrey WM, Reid RC:
**Precisely correlated firing in cells of the lateral geniculate nucleus.***Nature*1996,**383:**815.View ArticleGoogle Scholar - Warzecha AK, Kretzberg J, Egelhaaf M:
**Temporal precision of the encoding of motion information by visual interneurons.***Curr Biol*1998,**8:**359-368.View ArticleGoogle Scholar - Trong PK, Rieke F:
**Origin of correlated activity between parasol retinal ganglion cells.***Nat Neurosci*2008,**11**(11):1343-1351.View ArticleGoogle Scholar - Churchland M, Byron M, Cunningham J, Sugrue LP, Cohen MR, Corrado GS, Newsome WT, Clark AM, Hosseini P, Scott BB, Bradley DC, Smith MA, Kohn A, Movshon JA, Armstrong KM, Moore T, Chang SW, Snyder LH, Lisberger SG, Priebe NJ, Finn IM, Ferster D, Ryu SI, Santhanam G, Sahani M, Shenoy KV:
**Stimulus onset quenches neural variability: a widespread cortical phenomenon.***Nat Neurosci*2010,**13:**369-378.View ArticleGoogle Scholar - Steinmetz PN, Roy A, Fitzgerald PJ, Hsiao SS, Johnson KO, Niebur E:
**Attention modulates synchronized neuronal firing in primate somatosensory cortex.***Nature*2000,**404:**187-190.View ArticleGoogle Scholar - Stopfer M, Bhagavan S, Smith BH, Laurent G:
**Impaired odour discrimination on desynchronization of odour-encoding neural assemblies.***Nature*1997,**390:**70-74.View ArticleGoogle Scholar - Kazama H, Wilson RI:
**Origins of correlated activity in an olfactory circuit.***Nat Neurosci*2009,**12**(9):1136-1144.View ArticleGoogle Scholar - Poulet JFA, Petersen CCH:
**Internal brain state regulates membrane potential synchrony in barrel cortex of behaving mice.***Nature*2008,**454:**881.View ArticleGoogle Scholar - Gentet LJ, Avermann M, Matyas F, Staiger JF, Petersen CCH:
**Membrane potential dynamics of gabaergic neurons in the barrel cortex of behaving mice.***Neuron*2010,**65**(3):422-435.View ArticleGoogle Scholar - Binder MD, Powers RK:
**Relationship between simulated common synaptic input and discharge synchrony in cat spinal motoneurons.***J Neurophysiol*2001,**86:**2266.Google Scholar - Neiman AB, Russell DF:
**Two distinct types of noisy oscillators in electroreceptors of paddlefish.***J Neurophysiol*2004,**92:**492.View ArticleGoogle Scholar - Warzecha AK, Rosner R, Grewe J:
**Impact and sources of neuronal variability in the fly’s motion vision pathway.***J Physiol*2013,**107**(1):26-40.Google Scholar - Bialek W, Deweese M, Rieke F, Warland D:
**Bits and brains—information-flow in the nervous-system.***Physica A*1993,**200:**581.View ArticleGoogle Scholar - Gabbiani F:
**Coding of time-varying signals in spike trains of linear and half-wave rectifying neurons.***Netw Comput Neural Syst*1996,**7:**61.View ArticleMATHGoogle Scholar - Press WH, Teukolsky SA, Vetterling WT, Flannery BP:
*Numerical Recipes: The Art of Scientific Computing*. 3rd edition. Cambridge University Press, Cambridge; 2007.MATHGoogle Scholar - Gardiner CW:
*Handbook of Stochastic Methods*. 1985.Google Scholar - Bezrukov SM, Vodyanoy I:
**Stochastic resonance in non-dynamical systems without response thresholds.***Nature*1997,**385:**319-321.View ArticleGoogle Scholar - Gabbiani F, Cox SJ:
*Mathematics for Neuroscientists*. Academic Press, San Diego; 2010.MATHGoogle Scholar - Cox DR, Isham V:
*Point Processes*. Chapman & Hall, London; 1980.MATHGoogle Scholar - Feller W
**1.**In*An Introduction to Probability Theory and Its Applications*. 3rd edition. Wiley, New York; 1968.Google Scholar - Gestri G, Mastebroek HAK, Zaagman WH:
**Stochastic constancy, variability and adaptation of spike generation—performance of a giant-neuron in the visual-system of the fly.***Biol Cybern*1980,**38:**31.View ArticleMATHGoogle Scholar - de Ruyter van Steveninck RR, Lewen GD, Strong SP, Koberle R, Bialek W:
**Reproducibility and variability in neural spike trains.***Science*1997,**275:**1805.View ArticleGoogle Scholar - Feller W
**2.**In*An Introduction to Probability Theory and Its Applications*. 2nd edition. Wiley, New York; 1971.Google Scholar - Bauerle N, Grubel R:
**Multivariate counting processes: copulas and beyond.***ASTIN Bull*2005,**35**(2):379.MathSciNetView ArticleMATHGoogle Scholar - Neiman AB, Russell DF, Rowe MH:
**Identifying temporal codes in spontaneously active sensory neurons.***PLoS ONE*2011.,**6:**Article ID e27380Google Scholar - Trousdale J, Hu Y, Shea-Brown E, Josic K:
**A generative spike train model with time-structured higher order correlations.***Front Comput Neurosci*2013.,**7:**Article ID 84Google Scholar - Kuhn A, Aertsen A, Rotter S:
**Higher-order statistics of input ensembles and the response of simple model neurons.***Neural Comput*2003,**15**(1):67-101.View ArticleMATHGoogle Scholar - Moreno-Bote R, Parga N:
**Auto- and crosscorrelograms for the spike response of leaky integrate-and-fire neurons with slow synapses.***Phys Rev Lett*2006.,**96:**Article ID 028101Google Scholar - de la Rocha J, Doiron B, Shea-Brown E, Josic K, Reyes A:
**Correlation between neural spike trains increases with firing rate.***Nature*2007,**448:**802.View ArticleGoogle Scholar - Shea-Brown E, Josić K, de la Rocha J, Doiron B:
**Correlation and synchrony transfer in integrate-and-fire neurons: basic properties and consequences for coding.***Phys Rev Lett*2008.,**100:**Article ID 108102Google Scholar - Ostojic S, Brunel N, Hakim V:
**How connectivity, background activity, and synaptic properties shape the cross-correlation between spike trains.***J Neurosci*2009,**29:**10234.View ArticleGoogle Scholar - Vilela RD, Lindner B:
**A comparative study of three different integrate-and-fire neurons: spontaneous activity, dynamical response, and stimulus-induced correlation.***Phys Rev E*2009.,**80:**Article ID 031909Google Scholar - Abouzeid A, Ermentrout B:
**Correlation transfer in stochastically driven neural oscillators over long and short time scales.***Phys Rev E*2011.,**84**(6): Article ID 061914Google Scholar - Schultze-Kraft M, Diesmann M, Grun S, Helias M:
**Noise suppression and surplus synchrony by coincidence detection.***PLoS Comput Biol*2013.,**9**(4): Article ID e1002904Google Scholar - Bialek W, Zee A:
**Coding and computation with neural spike trains.***J Stat Phys*1990,**59:**103.View ArticleGoogle Scholar - Ermentrout GB, Terman DH:
*Mathematical Foundations of Neuroscience*. Springer, New York; 2010.View ArticleMATHGoogle Scholar - Malyshev A, Tchumatchenko T, Volgushev S, Volgushev M:
**Energy-efficient encoding by shifting spikes in neocortical neurons.***Eur J Neurosci*2013,**38**(8):3181-3188.View ArticleGoogle Scholar

## Copyright

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.