Address for correspondence
Stefano Panzeri
University of Newcastle Upon Tyne Department of Psychology, Ridley Building Newcastle Upon Tyne, NE1 7RU, U.K. Email || Email ||Web Page||Reprints | S. Panzeri,
S. R. Schultz, A. Treves and E. T. Rolls
copyright ©1999 by authors. |

Abstract.Is the information transmitted by an ensemble of neurons determined solely by the number of spikes fired by each cell, or do correlations in the emission of action potentials also play a significant role? We derive a simple formula which enables this question to be answered rigorously for short timescales. The formula quantifies the corrections to the instantaneous information rate which result from correlations in spike emission between pairs of neurons. The mutual information that the ensemble of neurons conveys about external stimuli can thus be broken down into firing rate and correlation components. This analysis provides fundamental constraints upon the nature of information coding - showing that over short timescales, correlations cannot dominate information representation, that stimulus-independent correlations may lead to synergy (where the neurons together convey more information than they would considered independently), but that only certain combinations of the different sources of correlation result in significant synergy rather than in redundancy or in negligible effects. This analysis leads to a new quantification procedure which is directly applicable to simultaneous multiple neuron recordings.## Introduction

A controversy exists over the extent to which firing rates and correlations between the responses of different cells (such as synchronisation of action potentials) contribute to information representation and processing by neural ensembles. Encoding of information in the correlation between the firing of different neurons has been demonstrated in specialised sensory systems such as the salamander retina (Meister et al.; 1995) and the auditory localisation system of the barn owl (Carr and Konishi; 1990; Carr; 1993). For the mammalian cortex, the evidence is less clear. Several investigators (Singer et al.; 1997; Vaadia et al.; 1995; deCharms and Merzenich; 1996) have presented evidence of stimulus-related changes in the correlation of firing between small populations of cortical cells. This might imply that correlations among cortical neurons reveal a substantial capability of cortical neural ensembles to code information synergistically - that is, that the ensemble of neurons provides more information than the simple sum of the contributions of the individual cells. Another possibility is that correlations actually limit rather than improve the information capacity of the population (Gawne and Richmond; 1993; Shadlen and Newsome; 1998; Zohary et al.; 1994). A third possibility is that correlations have no major effect on the efficiency of neural codes (Amit; 1997; Golledge et al.; 1996; Rolls and Treves; 1998). What is clearly needed in order to reconcile these viewpoints is a rigorous quantitative methodology for addressing the roles of both correlations and rates in information encoding.Information theory (Cover and Thomas; 1991; Shannon; 1948), which has found much use in recent years in the analysis of recordings from single cells (Kitazawa et al.; 1998; Optican and Richmond; 1987; Rieke et al.; 1996; Rolls and Treves; 1998), provides the basis for such an approach. Ideally, it would be desirable to take any experiment in which population activity was recorded in response to a clearly identifiable stimulus which the cells participate in encoding, and determine how many bits of information were present in the firing rates, how many in coincident firing by pairs of neurons, etc. - to divide the information into components which are indicative of the information encoding mechanisms involved. By considering the limit of rapid information representation, in which the duration of the window of measurement is of the order of the mean inter-spike interval (ISI), it is possible to do just this. This short timescale limit is not only a convenient approach to a complex problem; it is likely to be of direct relevance to information processing by the brain, as there is substantial evidence that sensory information is transmitted by neuronal activity in very short periods of time. Single unit recording studies have demonstrated that the majority of information is often transmitted in windows as short as 20-50ms (Heller et al.; 1995; Macknik and Livingstone; 1998; Oram and Perrett; 1992; Rolls and Tovee; 1994; Tovée et al.; 1993; Werner and Mountcastle; 1965). Event related potential studies of the human visual system (Thorpe et al.; 1996) provide further evidence that the processing of information in a multiple stage neural system can be extremely rapid. Finally, if one wishes to assess the information content of correlational assemblies which may last for only a few tens of milliseconds (Singer et al.; 1997), then appropriately short measurement windows must be used.

The current work describes the use of information theory to quantify the relative contributions of both firing rates and correlations between cells to the total information conveyed.

We report an expansion of the expression for mutual information to second order in time, show that the second order terms break down into those dependent on rate and those dependent upon correlation, and demonstrate that this can form the basis of a procedure for quantifying the information conveyed by simultaneously recorded neuronal responses. We show that with pairs of cells the approach works well for time windows several hundred milliseconds long, the time range of validity decreasing approximately inversely with population size. The expansion for the information shows that in short but physiological timescales, firing rates dominate information encoding when cell assemblies of limited size are considered, since correlations begin to contribute only with subleading terms. We find that the redundancy of coding is dependent upon the specific combination of the correlation in the number of spikes fired in the time window, which in the limit of short time windows measures the probability of coincident firing and thus the extent of synchronisation, and the correlation among mean response profiles to different stimuli. Furthermore, we observe that even stimulus-independent correlations may in some circumstances lead to synergistic coding.

## THE INFORMATION CARRIED BY NEURONAL RESPONSES AND ITS SHORT TIMESCALE EXPANSION.

MethodsConsider a period of time beginning at

t_{0}, of (short) durationt, in which a stimulussis present. Let the neuronal population response during this time be described by the vector , the number of spikes fired by each cell. Typically each component will have value 0 or 1, and only rarely higher. We may alternately describe the response by the firing rate vector ; this is purely for mathematical convenience and does not imply that we assume a firing rate code. In a typical cortical neurophysiology experiment we might hope to have tens of such trials with the same stimulus identifications. The stimuli considered are purely abstract: the procedures detailed in this paper are applicable to a wide variety of experimental paradigms. However, it may help to conceptualise the stimulus as for example which object, of some set, is being viewed by the experimental subject; indeed, data from such an experiment is examined later in this paper. Consider the stimuli to be taken from a discrete set withSelements, each occurring with probabilityP(s). The probability of events with response is denoted as , and the joint probability distribution as .Following Shannon (1948), we can write down the mutual information provided by the responses about the whole set of stimuli as

This assumes that the true probabilities are available. In practice, however, we have only a small to moderate number of events from which to compute the frequency table, and as a result a bias is introduced which must be corrected for (Panzeri and Treves; 1996).Now, the information can be approximated by a power series

whereI_{t}refers to the instantaneous information rate andI_{tt}to the second time derivative of the information, the instantaneous information ``acceleration''. For short timescales, only the first order and second order terms survive - higher order terms in the series become negligible. The time derivatives of the information can be calculated by taking advantage of the short time limit, as we will explain.DEFINITIONS AND MEASUREMENT OF CORRELATIONS FOR SHORT TIMESCALES.

There are two kinds of correlations that influence the information. These have been previously termed ``signal'' and ``noise'' correlations (Gawne and Richmond; 1993). They can be distinguished by separating the responses into ``signal'' (the average response to each stimulus) and ``noise'' (the variability of responses from the average to each stimulus). The correlations in the response variability represent the tendency of the cells to fire more (or less) than average when a particular event (e.g. a spike from another neuron) is observed in the same time window. For short time windows, this of course measures the extent of synchronisation of the cells. One way to introduce the parameters quantifying noise correlation in the short time limit is in terms of the conditional firing probabilities:

where is the mean rate of response of celli(amongCcells in total) to stimulussover all the trials in which that stimulus was present. It has been assumed in the above that the conditional probabilities (3) scale proportionally tot; this is the only assumption underlying our time expansion. It is a natural assumption, being violated only in the implausible case of spikes locked to one another with infinite time precision, but in any case it can be verified for any given dataset.For , Eq. 3 gives us the probability of coincident firing: cells

iandjboth fire spikes in the same time period; is the fraction of coincidences above (or below) that expected from uncorrelated responses, normalised to the number of coincidences in the uncorrelated case (which is , the bar denoting the average across trials belonging to stimuluss). is thus given by the following expression:

(4)

and is named the ``scaled cross correlation density'' (Aertsen et al.; 1989). It can vary from -1 to ; negative 's indicate anticorrelation, whereas positive 's indicate correlation. This measure of correlation is used because it has a number of advantages for short time windows over the more well-known Pearson correlation coefficient for this application as described in Appendix A.For

i=j, Eq. (3) provides instead the probability of observing a spike emission by celligiven that we have observed adifferentspike from the same celliduring the same time window. The ``scaled autocorrelation coefficient'' must be measured as

(5)

Notice that in the numerator of the above equation the subtraction of is needed in order to correctly quantify the number of occurrences of at least two spikes from the same cells, which was not necessary wheniandjrepresented different cells. The relation of this measure to the Fano factor is described in Appendix A. Since we are considering a single short time window, note that these correlations are observed over repeated trials with the same stimulus.In order to maintain homogeneity with respect to the noise correlation case, we choose to quantify the correlations in the signal, i.e. the correlations in the mean responses of the neurons across the set of stimuli, as a signal scaled crosscorrelation coefficient:

The definition is similar to that of scaled noise correlation; the main difference is that now the average is across stimuli: , not across trials. As before, varies between -1 and .RESPONSE PROBABILITY QUANTIFICATION.

If the firing rates conditional upon the firing of other neurons are non-divergent, as assumed in eq. (3), the

texpansion of response probabilities becomes essentially an expansion in the total number of spikes emitted by the population in response to a stimulus. The only responses with non-zero probabilities up to ordert^{k}are the responses with up tokspikes from the population; the only events with non-zero probability are therefore to second order intthose with no more than two spikes emitted in total:

where is the zero response (no cells fire), indicates a single spike fired by celli, and indicates a pair of spikes fired by cellsiandj. By inserting these into the expression for the amount of information, Eq.(1), and equating with Eq. 2, the information derivatives can be found. Note that interactions among more than two cells do not affect the second order probabilities (and thus will not affect the information up to second order).BIAS CALCULATION.

The information rate

I_{t}and each of the three components ofI_{tt}are affected by a systematic error when calculated from limited data samples. This is a problem general to all nonlinear functions of probabilities, but has been noted to be of particular concern for Shannon information, from whichI_{t}andI_{tt}derive. Furthermore, the bias in then-th derivative of the information has a 1/t^{n}dependence, making treatment of the bias problem even more crucial. The problem can be largely avoided by estimating the bias by a standard error propagation procedure and subtracting it from the calculated quantity. The procedure is detailed in Appendix B.INTEGRATE AND FIRE SIMULATION.

Correlated spike trains were simulated using a method similar to that used by Shadlen and Newsome (1998), to which we refer for a full discussion of the advantages and limitations of the model. In brief, each cell received 300 excitatory and 300 inhibitory inputs, each a Poisson process in itself, whose (possibly stimulus dependent) mean rate is constant across the set of inputs for any specific stimulus condition, and contributed a fixed quantity to the membrane potential, which decayed with a time constant of 20ms. When the membrane potential exceeded a threshold, it was reset to a baseline value. Crosscorrelation between cells was controlled through the proportion of inputs shared between cells (33% for Fig. 2b-c and 0%/90% for Fig. 2d-f). The spiking threshold was 14.5 times the magnitude of the quantal input above the baseline for the figures shown. The threshold and time constant parameters were chosen so as to guarantee the conservation of the response dynamic range, i.e. the neurons respond with approximately the same firing rate as their inputs over the range of cortical firing frequencies. Despite its simplicity, this ``Integrate and Fire'' model with balanced excitation and inhibition can account for several aspects of the firing statistics observed in responses of neurons across large regions of the cerebral cortex (Shadlen and Newsome; 1998).

By fixing the proportion of shared connections, but varying the mean firing rate according to which stimulus is present, rate coding in the presence of a fixed level of correlation can be examined. The mean firing rates chosen for each stimulus were extracted from data recorded from real inferior temporal cortical cells (see below). By holding the mean rate fixed (at the global mean corresponding to the previous case) but instead varying the proportion of shared connections according to the stimulus, one can examine a pure correlational code - thus allowing a clarification and test of the present information theoretic analysis.

ResultsTHE INFORMATION IN NEURONAL ENSEMBLE RESPONSES IN SHORT TIME PERIODS.

In sufficiently short time windows, at most two spikes are emitted from the population. Taking advantage of this, the response probabilities can be obtained explicitly in terms of pairwise correlations - triplet and higher order interactions do not contribute (see Methods). Here we report the result of the insertion of the response probabilities obtained in this limit into the Shannon information formula (1): exact expressions quantifying the impact of pairwise correlations on the information transmitted by groups of spiking neurons. The information depends upon both the noise correlations and the signal correlations .

In the short timescale limit, the first (

I_{t}) and second (I_{tt}) information derivatives suffice to fully describe the information kinetics. The instantaneous information rate is

Note that the expression forI_{t}is just a generalisation to the population level (a simple sum) of the expression previously derived for single cells (Bialek et al.; 1991; Skaggs et al.; 1993). The expression for the instantaneous information acceleration breaks up into three terms

The first of these terms is all that survives if there is no noise correlation at all. Thus therate componentof the information is given by the sum ofI_{t}(which is always greater than or equal to zero) and of the first term ofI_{tt}(which is instead always less than or equal to zero). The second term is non-zero if there is some correlation in the variance to a given stimulus, even if it is independent of which stimulus is present; this term thus represents the contribution ofstimulus-independent noise correlationto the information. The third component ofI_{tt}represents the contribution ofstimulus-modulated noise correlation, as it becomes non-zero only for stimulus-dependent correlations. We refer to these last two terms ofI_{tt}together as the correlational components of the information.The result just described shows that the short timescale limit allows a rigorous quantification of the effect of correlations on the information conveyed by neuronal ensembles. The price to pay is the limited temporal range of applicability, as it formally requires that the mean number of spikes in the considered time window be small. What must now be addressed is the actual temporal range of applicability of the approach. The range of validity of the (order

t^{2}) approximation will depend on how well the information time dependence fits a quadratic approximation (e.g., for a quadratic function, the Taylor expansion to the second order would of course be exact for the entiretrange). Since the assumption about the number of spikes will be broken first by the stimulus that gives the maximal response, a good scale for comparison with the range of validity is the minimum mean interspike interval to any stimulus. We studied the range of validity of the approximation, in the case of cells firing with Poisson statistics by direct calculation of the information; and in the more general case of up to four Integrate and Fire simulated neurons by a ``brute-force'' calculation, using many trials of simulated data. Fig. 1a shows the accuracy of the approximation for a pair of Poisson-simulated cells with mean-rate characteristics of real inferior temporal (IT) cells (Booth and Rolls; 1998); the range of validity in this case would appear to be, being conservative, about 6-8 times the minimum mean interspike interval. The approximation is better for smaller samples of cells. Scaling considerations suggest that the range of applicability should shrink as the inverse of the number of cells for larger populations. This expectation is roughly confirmed by the results in Fig. 1b, where five Poisson cells (again with typical cortical firing rates) are analysed. Integrate and Fire simulations confirm that this estimate of the range of validity is relatively robust to neuronal statistics and conserved across a wide range of response correlation values. Fig. 1c shows an estimate of the allowed ensemble size versus the time window for cells with mean rates to stimuli extracted from the cortical data of (Booth and Rolls; 1998). We conclude that the analysis is pertinent for timescales relevant to neuronal coding for ensembles of up to around 10-15 cells with firing rates similar to those in inferior temporal cortex, and for even larger ensembles of cells with lower firing rates, such as medial temporal lobe cells.

Figure 1.Range of validity of the short time-window approximation for the information. (a) The accuracy of the approximation shown for two Poisson-simulated cells with mean firing rate distributions to a set of ten simulated stimuli identical to those of a pair of inferior temporal cells from the dataset of (Booth and Rolls; 1998). The true information (solid line) is compared with the first (dashed) and second (dot-dashed) order series approximations. The vertical dotted line indicates the mean interspike interval for the stimulus which evokes the most spikes (providing a scale of comparison for the range of validity). An extensive number of simulations with Integrate and Fire neurons shows that this result, pictured in the simple Poisson case, is relatively robust to neuronal firing statistics. (b) As with (a), but for five instead of two simulated Poisson cells. (c) The allowable ensemble size while limiting the error due to the approximation to five percent, for the dataset, based upon 1/Cscaling.

RATE AND CORRELATION COMPONENTS OF THE INFORMATION.

So what is the impact of the second order terms on the information conveyed by the ensemble of neurons? Fig. 2 provides an illustration of their effect. In the case of non-interacting cells with zero autocorrelation, the information is carried entirely in the rate component of the information. This was exemplified by simulating a quintuplet of cells, each of which fired spikes according to a Poisson process (at any instant in time there is a given probability of firing which is constant in time throughout the experimental trial), independently of other cells, with a mean rate which was different for each of ten stimuli. The mean rate profiles of each of the five cells were taken from real cells described in (Booth and Rolls; 1998); these real cells were the ones analysed as simultaneously recorded pairs later in the paper, facilitating comparison with the real data examined. The result is shown in Fig. 2a, which, like the other graphs, shows separately the contributions of the three terms to the total information. The last two of these (the contributions of correlations) are negligible in the example of Fig. 2a, since the spike trains were by design uncorrelated.

By simulating spike trains with the Integrate and Fire model, we were able to examine situations in which the correlational components of the information were not negligible. The first such case considered was that of a quintuplet of neurons which had a large amount of common input (one third of their connections were shared). These neurons fired spikes in response to a balance of excitatory and inhibitory input (see Methods), with a membrane decay time of 20 milliseconds leading to a small amount of correlated firing, as shown by the crosscorrelogram in Fig. 2b. Ten stimuli again induced different mean firing rate responses from the cells. The correlation is not stimulus-dependent, and therefore the third component of

I_{tt}is still zero in this case. The second component ofI_{tt}, representing the effect of stimulus-independent correlation, can if required be again broken up into auto- (i=j) and cross-correlation () parts. The elements corresponding to autocorrelation are always positive for autocorrelations between -1 and 0 (as observed for the IT cells), whereas the crosscorrelation component is positive when, for positive noise correlation, the signal correlation is negative, i.e. when the cells anti-covary in their stimulus response profiles. In the case shown in Fig. 2c, this term is positive, and leads to a modest increase in the total information that can be transmitted.To model a situation where stimulus-dependent correlations conveyed information, we generated simulated data using the Integrate and Fire model for another quintuplet of cells which had a stimulus-dependent fraction of common input. This might correspond to a situation where cells transiently participate in different neuronal assemblies depending on stimulus conditions. There were again ten stimuli, but this time the mean spike emission rate to each stimulus was constant, at approximately 20 Hz, the global mean firing rate for the previous case. One of these stimuli simply resulted in independent input to each of the model cells, whereas each of the other nine stimuli resulted in an increase (to 90%) in the amount of shared input between one pair of cells (chosen at random from the ensemble such that each stimulus resulted in a different pair being correlated). The response of one such pair to changes in the amount of common input is shown in Fig. 2d. Panel (i) shows the fraction of shared connections as a function of time; panels (ii) and (iii) show the resulting membrane potentials and spike trains from the pair of neurons. During the high correlation state, there was on average 7 times higher probability of a coincidence in any 10ms period than chance. This crosscorrelation is also evident in the crosscorrelograms shown in Fig. 2e. The results are given in Fig. 2,,,f: all terms but the third of

I_{tt}are essentially zero, and information transmission is in this case almost entirely due to stimulus-dependent correlations. The total amount of information that could be conveyed, even with this much shared input, was modest in comparison to that conveyed by rates dependent on the stimuli, at the same mean firing rate. The total information increased slightly if for the ``high correlation'' state the spike trains were nearly perfectly correlated; if the ``low correlation'' state corresponded not to chance, but to actual anticorrelation, then it is possible that even more information could be conveyed.

Figure 2.By simulating a quintuplet of cells, we were able to examine situations in which each of the three information components was prominent. In each case shown, 100 trials of data were generated for each different stimulus. In the case of Poisson cells, all information is carried in the mean firing rates. This is shown in (a), in which the total information is compared with that from the rates only (the contributions ofI_{t}and the first term ofI_{tt}added together) and the additional information due to stimulus independent correlations and stimulus dependent correlations. (b) and (c) show the result of an Integrate and Fire simulation in which common input is introduced by sharing of 1/3 of the connections for each pair of cells, giving the crosscorrelogram shown in (b). The total information (thicker solid line) and components of the information are shown versus the width of the measurement window in (c). The rate component is still dominant, but the stimulus-independent correlational component has a non negligible effect. (d)-(f) show a situation in which correlational information dominates: with a fixed mean firing rate, two of the five simulated cells (chosen randomly for that stimulus) increase their correlation by increasing the number of shared connections while the other two remained randomly correlated. The effect of this on cell spiking activity is shown in (d): panel (i) shows the fraction of shared connections, while panels (ii) and (iii) show the membrane potential and spike emission of the simulated cells. (e) shows the crosscorrelograms corresponding to the low and high correlation states. The result of this is seen in (f): information due to correlations, although modest in magnitude, in this constructed case dominates the total information, as discussed in the text.

To illustrate the impact of the population size and of the overall average firing rate, we performed more simulations using again the spike trains with the Integrate and Fire model with a time constant of 20 ms and 30% shared connections, as in Fig. 2b,c. In Fig. 3a we tested the effects of population size by computing the information from pairs of cells instead of from five cells, by extracting pairs from the set of five cells and then averaging across pairs. The effect of reducing the size of the population from five to two cells was to reduce the total information by a factor of about 2. It is of course expected that the amount of information depends on the number of cells in the population. It is also shown with this particular set of generated spikes that the relative amount of information in the different components is approximately similar. In Fig. 3b we show the effects of applying the methods to cells with lower firing rates. The lower firing rates were produced simply by dividing the rates by three relative to those used in Fig. 2c. (This places the firing rates of the cells in the same regime as that of primate hippocampal pyramidal neurons in vivo, which fire at lower rates than cortical visual cells (Rolls, Robertson and Georges-François; 1997).) The information was computed from the responses of the population of five simulated cells. The effect of dividing the overall firing rate by three was to reduce the information by approximately two. If only the first derivative was important, we would expect the total information to decrease linearly when decreasing the rate. The sublinear decrease is mostly due to the rate component of the second derivative, which is a negative contribution, larger for higher rates and much smaller for lower rates. Similarly, the correlational component of the information is also much less important for lower rates, as evident from our analysis.

To illustrate the effects of precision of synchrony, we produced spike trains by simulating a correlational assembly with a constant firing rate of 20 Hz to all stimuli, and a percentage of shared connections of either 0% or 90% to different stimuli. In order to increase the precision of the synchrony with respect to Fig. 2d-f, we decreased the membrane time constant to the value of 1 ms. This gives, in the high correlation state, a very precise 1 ms synchrony between the spike trains, as shown by the crosscorrelogram in Fig. 3c. The information, shown in Fig. 3d, is fully conveyed by the correlational component with no information in the firing rates, correctly reflecting the way in which the spikes were generated. The actual information with the 1 ms precision is higher by a factor of 1.5 with respect with the 20 ms membrane time constant (Fig. 2f). Thus the precision of spike synchrony has an impact on the correlational part of the information, as predicted by our analysis.

Figure 3.Illustration of the effect of changing population size (a), firing rates (b), and precision of synchrony (c-d). All symbols and conventions are as in Fig. 2. (a) shows the effect of reducing the population size from five to two cells in the Integrate and Fire simulation with 30% common inputs of Fig. 2b-c. The total information is significantly reduced. In (b) we reduced by a factor of three the firing rates of the Integrate and Fire simulation with 30% common inputs of Fig. 2b-c. The total information is reduced, and also the relative contribution of correlation becomes much less important at low rates. In (c-d) we simulated a correlational assembly with a constant firing rate of 20 Hz to all stimuli, and a percentage of shared connections of either 0% or 90% to different stimuli, as in Fig. 2d-f. The membrane time constant is set to 1 ms. The crosscorrelogram in the 90% shared connection state is plotted in (c). The information, shown in (d), is fully conveyed by the stimulus dependent correlational component and, because of the increased precision of synchrony, is higher than in the case of the 20 ms membrane time constant (Fig. 2f).

APPLICATION OF THE METHOD TO REAL NEURONAL DATA

To demonstrate that the technique is applicable to real data, we applied the information component analysis to three pairs of cortical cells recorded simultaneously from the same electrode from the dataset of (Booth and Rolls; 1998). This data came from an experiment in which the responses of neurons in the inferior temporal cortex of the macaque monkey were recorded while one of ten different objects was being viewed. The firing rates of these neurons were found by Booth and Rolls (1998) to convey information about which object was present, regardless of viewing angle. Four different views of each object were each presented five times, resulting in a total of twenty trials for each object. The median firing rate to the best stimulus of the cells used in our analysis was 44 Hz. The same data was used to estimate realistic firing rates for all the simulations described in this paper.

Each component of the information for the cells in this dataset were calculated for each pair of cells, and the standard error in the calculation of each information component was obtained by error propagation from the variances in the measurements of spike counts and coincidences over the trials. The average information encoding characteristics of the three pairs (which were qualitatively similar to each other) is shown in Fig. 3. In this example it is clear that the rate component is prominent in the information representation; however, the important point to gain from this result is that the information expansion provides a demonstrably practical method for analysing simultaneously recorded data with only a small number of experimental trials. The values of each component and an estimate of its error are available both for single small ensembles, and for the ``average picture'' obtained by recording large numbers of such small neuronal ensembles. This encourages us to think that it will be possible to analyse simultaneously recorded data in a systematic and rigorous quantitative manner. Of course, the populations of neurons that actually act effectively together are larger than can be studied using this, or any other presently known procedure. However, it is reasonable to assume that effects present in such large ensembles will be to some extent observable in the smaller ensembles that we can in practice record from and analyse.

Figure 4.Application of the procedure to real neurophysiological data. The information contained in the responses of three pairs of primate inferior temporal (IT) cortical neurons about viewed objects is broken down into three components: rate, stimulus-independent noise correlation, and stimulus-modulated noise correlation. The vertical dotted line again shows for timescale comparison the mean interspike interval to the maximally responsive stimulus.

DiscussionREDUNDANCY VERSUS SYNERGY.

The

redundancyof a code can be defined (and measured in bits) as the amount of information that would be obtained by adding the information obtained from each cell as if they were independentminusthat obtained by considering the whole neuronal ensemble (Rieke et al.; 1996). (If part of what two neurons are saying is the same, it does not help to say it twice.) For synergistic coding (in which knowledge of the states of a pair of neurons at the same time provides more information than one would obtain by considering each neuron alone), the value of this Shannon redundancy is of course negative.Eq. 9 shows that overall correlations in the distribution of mean responses alone can only lead to redundancy - since all (

i,j)contributions to the first term ofI_{tt}are negative, except that they are zero when there is no overall ``signal'' correlation in the mean response profiles (i.e. ). To have synergistic coding of information one needs correlations in the variability of the responses (to a given stimulus), i.e. non-zero parameters. Even when such ``noise'' correlations are independent of the stimuli, however, it is possible to have synergy. This can be demonstrated by considering the sign of the Shannon redundancy, obtained from Eq. 1 by subtracting the information conveyed by the population from the sum of that carried by each single cell. This shows four basic régimes of operation, illustrated in the two-cell example of Fig. 5. If the cells anti-covary in their response profiles to stimuli, the cell must have a positive noise correlation, above the boundary value depicted in Fig. 5(a), to obtain synergy; or if the cells do have positive signal correlation, then coincidences must be actively suppressed by a negative noise correlation stronger than the corresponding boundary value. When the signal and noise correlations have the same sign, one always obtains redundancy in the short timescale limit. Clearly, already with pairs of cell, the interplay between correlation in the noise and correlation in the signal introduces a potential for both redundancy and synergy.A simple example of the interplay between signal and noise correlation in a pair of cells is graphically depicted in Fig. 5(b). In this example there are three stimuli, A, B and C, which occur with equal probability. The first cell emits a different mean number of spikes to each stimulus: 1 to A, 2 to B, and 3 to C. Let us consider cases where the second cell either fires the same mean number of spikes to each stimulus (full correlation), or alternatively has the mean responses to A and C exchanged (anti-correlation with ). On any given occasion, noise adds 0 or spikes, with equal (1/3) probability, to the output of the cell. It is easy to check that the information carried by each cell alone is bits, which is less than half the maximum information (entropy) available in the stimulus set.

Fig. 5(b) illustrates the four different situations for the second cell, depending on the way in which the mean (signal) and variance (noise) of the number of spikes fired by it are related to those of the first cell. When the signal and noise are both correlated (or both anti-correlated), the joint probabilities of the numbers of spikes fired by each cell tend to bunch up along the diagonal, so that the total information from both cells is less than the sum of that obtained from each cell on its own. If however with correlated signal the noise is anti-correlated or if with correlated noise, the signal is anti-correlated, then the joint probabilities are more spread out. This means that the presented stimulus can be clearly identified on the basis of the joint response observed. If this spread is sufficient, it is possible to obtain synergy: the information calculated from the joint probability matrix exceeds the sum of that obtained for each cell individually. In this example, with the signal having (anti-correlation), if the noise was uncorrelated between the cells, then the redundancy can be easily seen to be 0.04 bits; it is only when the noise correlation is increased above that the coding is synergistic.

Examination of Eq.(9) and Fig. 5(c) reveals the total amount of redundancy (or synergy) to be much more sensitive to noise correlation when the signal correlation is high - e.g. the noise correlation leads to large redundancy if the cells are tuned to the same stimulus. If the signal correlation is small, the redundancy is close to zero, no matter how correlated the noise is. This explains why the impact of noise correlation on the performance of pools of neurons has been emphasised in experiments utilising simple one-dimensional discrimination tasks (in which neurons from a local pool have been found to be tuned to the same stimulus (Parker and Newsome; 1998; Zohary et al.; 1994)), while noise correlation has been described as less important for groups of neurons coding for complex stimuli (Gawne and Richmond; 1993; Gawne et al.; 1996), which tend to use a more distributed encoding. This study therefore shows rigorously that correlations do not necessarily invoke redundancy, and that it is not possible in general to estimate the `effective' number of neurons participating in encoding by simply measuring the noise correlation, as done by Zohary et al. (1994). Further, it predicts, since the signal correlation was reported to decrease towards zero when increasing the complexity of the stimulus set used to test the neurons (Gawne et al.; 1996; Rolls et al.; 1997), that noise correlation might have only a small impact in the encoding of large sets of natural stimuli - at least as far as stimulus-independent noise correlation is concerned. The effect of small stimulus-dependent correlations in the noise is considered below, and it should not be neglected.

The possibility of synergy with constant correlations has been raised previously (Oram et al.; 1998) as a phenomenological observation; the information expansion we have introduced places this phenomenon on a solid mathematical footing, and delineates, for short time windows, the exact boundaries of the regions of synergy and redundancy. The discussion by Oram et al. (1998) points out that correlations might lead to synergy when neurons are tuned to different stimuli, but it is not able to bridge between different encoding situations and predict the exact amounts of redundancy or synergy that occur, as our analysis does. The analysis of synergy presented here, unlike that of (Oram et al.; 1998), generalises to the case of stimulus-modulated correlations - it is enough to specify the stimulus dependence of correlations and take into account also the third term of

I_{tt}.

Figure 5.Redundancy and synergy. (a) Consideration of a pair of neurons, in which the noise correlation is taken to be stimulus-independent, illustrates the range of possibilities. The vertical axis plots the noise correlation measured by the scaled cross-correlation density . The horizontal axis shows the correlation in the signal, also expressed in terms of the scaled crosscorrelation density . If the mean responses of each neuron to the various stimuli are uncorrelated, , then (no matter what the noise correlation ) the redundancy is exactly zero. This is the case along the vertical solid line. If the pair of neurons covary in the mean number of spikes they fire to each stimulus, then all else being equal, we have a redundant situation. If, however, there is enough anticorrelation in the noise, or vice versa there is anticorrelation in the signal and enough positive correlation in the noise, then the coding becomes synergistic. (b) Graphical depiction of the interplay between signal and noise correlation. Histograms show the probability of each cell emitting from zero to four spikes. The joint probability matrix of spike emissions from both cells is shown between the histograms for each cell. Contributions from each stimulus are labelled by shading the corresponding probability block. (c) The redundancy (divided by the square of the time window) corresponding to the above plane. Values below the zero axis indicate synergistic encoding.

A NULL HYPOTHESIS FOR THE ROLE OF CORRELATIONS IN THE CEREBRAL CORTEX

How might typical correlations among cells in a population scale up with the size of the population? Clearly, this is a question for experiments to address, in fact a most crucial question for those investigating correlations in neural activity. Such experiments need to be well designed, accurate and systematic. It is easy to see, from the analysis above, that with large populations even small correlations could produce extreme effects, resulting in either large redundancy or (perhaps less often) very substantial synergy. This is because, while with

Ccells there areCfirst order terms in the information, there are obviouslyC^{2}second order terms,C^{3}third order ones (which depend also on three-way correlations) and so on. Depending on tiny details of the correlational structure, successive terms can affect transmitted information in both directions. Within such broad a realm of possibilities, it is then of interest to try to formulate a sort of null hypothesis, that might provide at least a reference point against which to contrast any more structured candidate theory. One example is the scaling behaviour we might expect if the correlations were not playing any special role at all in the system or area being analysed. In this ``null'' hypothesis, the parameters would be expected to be small, that is to deviate from 0 (no correlations) only in so far as the set of stimuli used is limited (Gawne and Richmond; 1993; Rolls, Treves and Tovée; 1997); similarly the stimulus-dependent noise correlation would be small. The scaling behaviour corresponding to this null hypothesis can be examined by further expandingI_{tt}as a series in these new small parameters: at times of the order of the inter-spike interval, second order terms of order (redundant) and (synergistic) are introduced (the angular brackets indicating the average value). If we have a large enough population of cells, and and are not sufficiently small to counteract the additionalCfactor, these ``random'' redundancy and synergy contributions will be substantial. Obviously, in this situation our expansion would start to progressively fail in quantifying the information, as higher order terms in thetexpansion become more and more important in this case. But this clearly shows that a sufficiently large population of cells, which has not been designed to code stimuli in any particular cooperative manner, has the potential to provide large effects of redundancy or synergy, arising simply from random correlations among the firing of the different cells. This reinforces the need for systematic study of the magnitude and scaling of correlations in the cerebral cortex.CORRELATIONAL ASSEMBLIES AND MUTUAL INFORMATION

If cells participate in context-dependent correlational assemblies (Singer et al.; 1997), then a significant amount of information should be found in the third component of

I_{tt}, relative to the total information, when analysing data obtained from the appropriate experiments. The challenge for the establishment of correlational theories of neural coding has thus been laid down: to demonstrate quantitatively how substantial a proportion of the information about external correlates is provided by correlations between cells, given the large amount of information that has been shown in some neural systems to be coded by rate (Rolls and Treves; 1998). The second order series expansion we have described allows precisely this to be achieved for small ensembles of cells - for a time window of 20ms, an ensemble of about 10-15 cells which fired at a peak mean rate to a stimulus of around 50Hz (e.g. IT cells) could be analysed; with the same time window, 25-30 cells firing at a lower peak mean rate of around 20 Hz, (such as neurons from the medial temporal lobe) could be studied. Beyond this population size the information expansion can still be of use in picking up the correlational variables conveying most of the information in small subpopulations, and eliminating the irrelevant variables. Reduction of the response space of a large population of cells to a treatable size thus may be possible without loosing salient features.In order to test hypotheses about the role of correlations in solving the binding problem (Gray et al.; 1992; Singer et al.; 1997; von der Malsburg; 1995), as opposed to other solutions (Treisman and Gelade; 1980; Treisman; 1996), and about information coding in general (Vaadia et al.; 1995; deCharms and Merzenich; 1996), careful quantitative experimental studies of the correlations prevailing in the neural activity of different parts of the brain are needed. Data analyses based on the time-expansion approach then have the potential to elucidate the role of correlations in the encoding of information by cortical neurons.

AcknowledgementsWe thank R. Baddeley, A. Renart and D. Smyth for useful discussions. We thank M. Booth for kindly providing neurophysiological data. This research was supported by an E.C. Marie Curie Research Training Grant ERBFMBICT972749 (SP), a studentship from the Oxford McDonnell-Pew Centre for Cognitive Neuroscience (SRS), and by MRC Programme Grant PG8513790.

Appendix A. Correlation measures.The numerical value of the information calculated is independent of the correlation measure used, and we show here how to express the results using the Pearson correlation coefficients and Fano factors instead of scaled cross correlation coefficients. In the text, we chose to use the scaled cross-correlation measure of Eq. 3 because it produces a more compact mathematical formulation of what is addressed in this paper, and has useful scaling properties as the time window becomes small.

A widespread measure for crosscorrelation is the Pearson correlation coefficient , which normalises the number of coincidences above independence to the standard deviation of the number of coincidences expected if the cells were independent. The normalization used by the Pearson correlation coefficient quantifies the strength of correlations between neurons in a rate-independent way. However, it should be noted that the Pearson noise-correlation measure approaches zero at short time windows:

(10)

where is the standard deviation of the count of spikes emitted by celliin response to stimuluss.Under assumption 3, remains finite as , thus by using this measure we can keep the

texpansion of the information explicit. This greatly increases the amount of insight obtained from the series expansion.Similarly, an alternative to scaled autocorrelation density for the measure of autocorrelations is the so called ``Fano'' factor

F, that is the variance of the spike count divided by its mean (Rieke et al.; 1996). This measure is used in neurophysiology because for the renewal process, often used as a stochastic model of neuronal firing, the variance is proportional to the mean. (Fano factors lower than 1 indicate that the process is more regular than a Poisson process).Fgrows linearly withtfor short times: . Again, we prefer to the Fano factor in the information expansion becauseF-1 approaches zero for short times.To express the information derivatives in terms of Pearson correlation coefficients and Fano factor, is is enough to make the following simple substitutions in eq. (9):

(11)

We note that the ``scaled cross-correlation measure'' is sensitive to the mean firing rate, as the strength of neuronal interactions might be overemphasized at low rate (Aertsen et al.; 1989): it cannot be taken to be a linear measure of interaction strength. However, the value of the information transmitted by the the number of spikes simultaneously fired by each cell, and of each particular components, depends on the response probabilities, and not on the particular way chosen to quantify the correlations. Therefore the particular measure used for correlations is for this application ultimately a matter of mere notation.

Appendix B: Evaluation of the bias and of the variance of information derivatives

It is possible to analytically derive an estimate of the amount of the bias, which can then be subtracted to provide an unbiased estimate. This is done using the standard error propagation procedure (see e.g. Bevington and Robinson; 1992).

A function of the firing rates can be expanded about the mean rate as

(1)

where

(2)

Applying this to Eq. 8, we obtain

(3)

where the `hat' over theisummation indicates that it is only over the `relevant's,ipairs, i.e. those with non-zero underlying probability of spike emission. If the underlying probability is zero, then no finite sampling fluctuations are possible and thats,idoes not contribute to the bias.Obviously this correction as it stands cannot be applied to the components of

I_{tt}. A similar correction can be derived by the same method. This calculation as we shall see is slightly more involved. We will have to calculate the bias for each component ofI_{tt}separately, so we will consider a generic function of a set of (possibly correlated) random variables. Each random variable is the average (obtained on the basis of a limited number of trialsN) of a random variablex_{j}. We assume for the purposes of our analytical estimate that the number of trialsNis large but finite. In this case the independent variables fluctuate around their true value , and the fluctuations scale as 1/N. Therefore this derivation of the bias of each information component using error propagation is equivalent to the 1/Nexpansion of the bias of the full information derived e.g. in (Panzeri and Treves; 1996).Under these assumptions, the sampling bias in is:

where the (co)variances of the means of the random variables are

The expression for the variance of is similarly

To apply the above formalism to our cases of the information derivative components

I_{tt1}-I_{tt3}, the random variables used are the mean rates of the cells to each stimulus , and the variables , which are defined as

= (8) = (9)

represents the rate of coincidences between cells, whereas parameterises the rate of `autoincident' firing by the same cell (i.e. two different spikes fired by the one cell in the same small time window). The corresponding parameters could have been used in this derivation, but use of the variables in this section makes the bias derivation easier and more transparent. These random variables are each calculated on the basis ofN_{s}trials per stimulus.The leading contributions of these (co)variances can be calculated analytically in the short time window limit. They are:

Note that the leading order fluctuations are those in the number of coincidences (i.e. ). This provides a note of caution for measurers of correlation. For the estimate of the biases, which we are about to detail, in practice the variances may as well be computed numerically from the measurements as from these formulae; the formulae do however provide a check.Carrying out the differentiation, we obtain for the bias of the three separate components of

I_{tt}(denoted byI_{tt1(bias)},I_{tt2(bias)},I_{tt3(bias)}):

The `hats' in the summations of terms proportional to have the same meaning as before; those in the summations of terms proportional to similarly restrict the sum to cases with non-zero underlying probabilities of observing coincidences from cellsi,jin response to stimuluss(and similarly for the other terms).Note that the leading contribution to the

I_{tt}bias is fromI_{tt3}, which is proportional to 1/t^{2}, whereas the biases inI_{tt1}andI_{tt2}are only proportional to 1/t.We now have analytical expressions for the bias due to finite sampling in each of the components of

I_{tt}, as well asI_{t}. The bias estimate obtained from each of these is subtracted from the `raw' quantity. A more detailed study of the range of validity of the bias removal using simulated data can be found in (Schultz; 1998).We conclude by noting that the procedure used to count `bins' for the summation over `relevant bins' in the above equations was a `naive' counting procedure, in which we only add terms in which there is at least one spike (or coincidence if it is a sum over

iandj) in any of the trials. For sufficiently short time windows, and a small number of trials per stimulus, the bias correction fails. This occurs more evidently inI_{tt3}because of the 1/t^{2}dependence. Other non-naive counting procedures can be used to obtain more accurate estimates of the bias. By using a Bayesian counting procedure (exactly the same one described in (Panzeri and Treves; 1996)), it was possible to reduce somewhat the time at which the bias correction broke down, and obtain a more accurate bias estimation at very short times, at the expense of losing the property that the resulting information estimate is an upper bound on the information. In this procedure, the problem of the summations reduces to estimating the number of relevant bins. This is done by choosing a ``guess'' value for the number of relevant bins, and a prior probability function which has one (constant) value for each of the occupied bins and another (constant) value for each of the empty bins. The posterior probability distribution is then calculated and the posterior estimate of the number of relevant bins obtained. This procedure applies just as well toI_{tt3}as it did to the full information in the case described by Panzeri and Treves (1996). We do not describe this Bayesian counting procedure here, because it is exactly the one reported in Panzeri and Treves (1996); however, for values of firing rates in the range relevant for visual cortical cells, the use of a Bayesian counting procedure makes a difference only for time windows as short as 2-3 ms.

## References

Aertsen, A. M. H. J., Gerstein, G. L., Habib, M. K. and Palm, G. (1989).- Dynamics of neuronal firing correlation: modulation of ``effective connectivity'',
J. Neurophysiol.61: 900-917.

Amit, D. J. (1997).- Is synchronization necessary and is it sufficient?,
Behavioural and brain sciences20: 683.

Bevington, P. R. and Robinson, D. K. (1992).Data Reduction and error analysis for the physical sciences, McGraw-Hill, New York.

Bialek, W., Rieke, F., de Ruyter van Steveninck, R. R. and Warland, D. (1991).- Reading a neural code,
Science252: 1854-1857.

Booth, M. C. A. and Rolls, E. T. ( 1998).- View-invariant representations of familiar objects by neurons in the inferior temporal visual cortex,
Cerebral Cortex8: 1047-3211.

Carr, C. E. (1993).- Processing of temporal information in the brain,
Ann. Rev. Neurosci.16: 223-243.

Carr, C. E. and Konishi, M. (1990).- A circuit for detection of interaural time differences in the brainstem of the barn owl,
J. Neurosci.10: 3227-3246.

Cover, T. M. and Thomas, J. A. (1991)- .
Elements of information theory, John Wiley, New York.

deCharms, R. C. and Merzenich, M. M. ( 1996).- Primary cortical representation of sounds by the coordination of action potentials,
Nature381: 610-613.

Gawne, T. J., Kjaer, T. W., Hertz, J. A. and Richmond, B. J. (1996).- Adjacent visual cortical complex cells share about 20% of their stimulus-related information,
Cerebral Cortex6: 482-489.

Gawne, T. J. and Richmond, B. J. ( 1993).- How independent are the messages carried by adjacent inferior temporal cortical neurons?,
J. Neurosci.13: 2758-2771.

Golledge, D. R., Hildetag, C. C. and Tovée, M. J. (1996).- A solution to the binding problem?,
Current Biology6(9): 1092-1095.

Gray, C. M., Engel, A. K., König, P. and Singer, W. (1992).- Synchronization of oscillatory neuronal responses in cat striate cortex: Temporal properties,
Visual Neuroscience8: 337-347.

Heller, J., Hertz, J. A., Kjaer, T. W. and Richmond, B. J. (1995).- Information flow and temporal coding in primate pattern vision,
J. Comp. Neurosci.2: 175-193.

Kitazawa, S., Kimura, T. and Yin, P.-B. ( 1998).- Cerebellar complex spikes encode both destinations and errors in arm movements,
Nature392: 494-497.

Macknik, S. L. and Livingstone, M. S. ( 1998).- Neuronal correlates of visibility and invisibility in the primate visual system,
Nature Neuroscience1(2): 144-149.

Meister, M., Lagnado, L. and Baylor, D. A. ( 1995).- Concerted signalling by retinal ganglion cells,
Science270(17): 1207-1210.

Optican, L. M. and Richmond, B. J. ( 1987).- Temporal encoding of two-dimensional patterns by single units in primate inferior temporal cortex: III. Information theoretic analysis,
J. Neurophysiol.57: 162-178.

Oram, M. W., Földiák, P., Perrett, D. I. and Sengpiel, F. (1998).- The 'Ideal Homunculus': decoding neural population signals,
Trends in Neurosciences21(6): 259-265.

Oram, M. W. and Perrett, D. I. (1992)- . Time course of neuronal responses discriminating different views of face and head,
J. Neurophysiol.68: 70-84.

Panzeri, S. and Treves, A. (1996).- Analytical estimates of limited sampling biases in different information measures,
Network7: 87-107.

Parker, A. J. and Newsome, W. T. ( 1998).- Sense and the single neuron: probing the physiology of perception,
Annual Review of Neuroscience21: 227-277.

Rieke, F., Warland, D., de Ruyter van Steveninck, R. R. and Bialek, W. (1996).Spikes: exploring the neural code, MIT Press, Cambridge, MA.

Rolls, E. T. and Tovee, M. J. (1994)- . Processing speed in the cerebral cortex and the neurophysiology of visual masking,
Proc. R. Soc. London257: 9-15.

Rolls, E. T. and Treves, A. (1998).Neural networks and brain function, Oxford University Press, Oxford, U.K.

Rolls, E. T., Treves, A. and Tovée, M. J. ( 1997).- The representational capacity of the distributed encoding of information provided by populations of neurons in primate temporal visual cortex,
Exp. Brain Res.114: 149-162.

Schultz, S. R. (1998).Information encoding in the mammalian cerebral cortex. D.Phil. Thesis, Corpus Christi College, Oxford.

Shadlen, M. N. and Newsome, W. T. ( 1998).- The variable discharge of cortical neurons: implications for connectivity, computation and coding,
J. Neurosci.18(10): 3870-3896.

Shannon, C. E. (1948).- A mathematical theory of communication,
AT&T Bell Labs. Tech. J.27: 379-423.

Singer, W., Engel, A. K., Kreiter, A. K., Munk, M. H. J., Neuenschwander, S. and Roelfsema, P. (1997).- Neuronal assemblies: necessity, signature and detectability,
Trends in Cognitive Sciences1: 252-261.

Skaggs, W. E., McNaughton, B. L., Gothard, K. and Markus, E. (1993).- An information theoretic approach to deciphering the hippocampal code,
inS. Hanson, J. Cowan and C. Giles (eds),Advances in Neural Information Processing Systems, Vol. 5, Morgan Kaufmann, San Mateo, pp. 1030-1037.

Thorpe, S., Fize, D. and Marlot, C. ( 1996).- Speed of processing in the human visual system,
Nature381: 520-522.

Tovée, M. J., Rolls, E. T., Treves, A. and Bellis, R. P. (1993).- Information encoding and the response of single neurons in the primate temporal visual cortex,
J. Neurophysiol.70: 640-654.

Treisman, A. (1996).- The binding problem,
Current Opinion in Neurobiology6: 171-178.

Treisman, A. and Gelade, G. (1980).- A feature integration theory of attention,
Cogn. Psychol.12: 97-136.

Vaadia, E., Haalman, I., Abeles, M., Bergman, H., Prut, Y., Slovin, H. and Aertsen, A. (1995).- Dynamics of neuronal interactions in monkey cortex in relation to behavioural events,
Nature373: 515-518.

von der Malsburg, C. (1995).- Binding in models of perception and brain function,
Current Opinion in Neurobiology5: 520-526.

Werner, G. and Mountcastle, V. B. ( 1965).- Neural activity in mechanoreceptive cutaneous afferents: Stimulus-response relations, Weber functions, and information transmission,
J. Neurophysiol.28: 359-397.

Zohary, E., Shadlen, M. N. and Newsome, W. T. ( 1994).- Correlated neuronal discharge rate and its implication for psychophysical performance,
Nature370: 140-143.