HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS		QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Abstract Full Text (PDF) All Versions of this Article: 96/2/746    most recent 00059.2006v1 Alert me when this article is cited Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via ISI Web of Science (10) Citing Articles via Google Scholar Google Scholar Articles by Eggermont, J. J. Search for Related Content PubMed PubMed Citation Articles by Eggermont, J. J.

Properties of Correlated Neural Activity Clusters in Cat Auditory Cortex Resemble Those of Neural Assemblies

Departments of Physiology and Biophysics and Psychology, University of Calgary, Calgary, Alberta, Canada

Submitted 18 January 2006; accepted in final form 3 May 2006

	ABSTRACT

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
Spiking activity was recorded from cat auditory cortex using multi-electrode arrays. Cross-correlograms were calculated for spikes recorded on separate microelectrodes. The pair-wise cross-correlation matrix was constructed for the peak values of the correlograms. Hierarchical clustering was performed on the cross-correlation matrix for six stimulus conditions. These were silence, three multi-tone stimulus ensembles with different spectral densities, low-pass amplitude-modulated noise, and Poisson-distributed click trains that each lasted 15 min. The resulting neuron clusters reflect patches in cortex of up to several mm² in size that expand and contract in response to different stimuli. Cluster positions and size were very similar for spontaneous activity and multi-tone stimulus-evoked activity but differed between those conditions and the noise and click stimuli. Cluster size was significantly larger in posterior auditory field (PAF) compared with primary auditory cortex (AI), whereas the fraction of common spikes (within a 10-ms window) across all electrode activity participating in a cluster was significantly higher in AI compared with PAF. Clusters crossed area boundaries in <5% of the cases were simultaneous recording were made in AI and PAF. Clusters are therefore similar to but not synonymous with the traditional view of neural assemblies. Common-spike spectrotemporal receptive fields (STRFs) were obtained for common-spike activity and all-spike activity within a cluster. Common-spike STRFs had higher signal-to-noise ratio than all-spike STRFs and showed generally spectral and temporal sharpening. The coincident and noncoincident output of the clusters could potentially act in parallel and may serve different modes of stimulus coding.

	INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
The frequency of sound is represented in the form of frequency-place maps in many auditory cortical areas. However, other representations in the auditory cortex may take the form of clusters rather than maps. A potential example is the spatial representation of sound (Cohen and Knudsen 1999). More complex perceptual properties have been linked to distributed neural representations called neural assemblies (Harris 2005). Do the properties of neural clusters have similarities to those attributed to neural assemblies? Comparison of the properties of clustered neural activity with those attributed to neural assemblies will be facilitated if the same criteria are used to define neural clusters and neural assemblies.

Neural assemblies have been defined as "a group of neurons [that are] at least transiently working together as indicated by correlation of unit activity" (Gerstein and Kirkland 2001). In visual cortex, cells with 0.5-mm separation showed the highest correlation among cells with similar receptive fields and similar connectivity from the lateral geniculate nucleus (Lampl et al. 1999). This suggests that connectivity is a dominant factor in neural assembly formation. Correlation and connectivity may ensure efficient propagation of neural activity through downstream stages in the neural pathway (Kimpo et al. 2003; Kistler and Gerstner 2002; Reyes 2003).

It is common to think about a neural assembly as widely distributed in cortical space, potentially extending over various subdivisions of cortex (Singer and Gray 1995). Connections over large spatial divisions of auditory cortex are provided by the thalamic cell axonal divergence and convergence, often estimated to be between 2 and 5 mm at the cortical level (Lee et al. 2004a) and intracortically through horizontal fibers (Clarke et al. 1993; Wallace et al. 1991). In visual cortex, the spatially periodic effects of the patchy connections of the horizontal fibers have been shown by cross-correlation (Ts'o et al. 1986). These cortico-cortical connections are for a sizable part heterotopic, i.e., do connect cell groups with characteristic frequencies (CFs) differing by more than one octave (Lee et al. 2004b).

Neural assembly membership is expected to be stimulus dependent and context specific and may reflect the number and functional strength of its common inputs under different conditions (Usrey and Reid 1999). It is, however, likely that at any point in time several, spatially overlapping, neural assemblies are active. In response to external events, a group of neurons forming a dynamical cell assembly may spontaneously organize itself temporarily by correlated firing of their spiking activity.

Neural assemblies, thus defined, may potentially be probed using micro-electrode arrays that allow recording from a set of relatively widely spaced neurons. These neurons could participate in one or more neural assemblies. The quantification of the correlation in spiking activity occurring between pairs of such widely spaced neurons thus becomes crucial in defining membership of neural assemblies. The stimulus will be one of the dominant sources of neural correlation because of its common input character. Although it is common to correct for stimulus-induced correlations by using shift predictors or joint PST histogram techniques (Eggermont 1994), the brain does not have that luxury but may exploit this stimulus-dependent correlation to change the extent and structure of the neural assemblies.

To be confident that the observed changes in the cross-correlation matrix are functional, one has to eliminate other ways in which neural correlations can change. It is known that significant correlations can arise from covariation in neuronal excitability, i.e., co-modulated firing rates, and from covariation in response latency (Brody 1999a,b) as well as from trial-to-trial variability in the stimulus or the effect thereof (Ben-Shaul et al. 2001). The latter may also be a reflection of changes in the underlying assembly. It is therefore important to use experimental conditions that will not allow stimulus variability, e.g., resulting from movement of the animal, which can be accomplished by using anesthesia. It is, however, also important to minimize state changes in the brain that can be caused by fluctuating or slowly changing anesthesia levels (Erchova et al. 2002) and their effect on cross-correlation strength. This may be harder to accomplish because a change from a state showing periodic cross-correlograms in the auditory thalamus, likely caused by spindle activity due to the use of ketamine anesthesia, to one without these periodicities could be induced by stimulation with continuous dynamic ripple noise (Miller and Schreiner 2000). So, even in stable anesthetized animals, brain states are not invariant, and transitions may be induced by appropriate auditory stimulation. One can also posit that appropriate auditory stimulation breaks down those assemblies that during silence synchronize to produce spindle activity. The result could be changing pair correlations between the neurons recorded from. These changing correlation strengths may in turn cause neurons to associate in different assemblies.

To make putative neural assemblies last longer, and easier to detect, compared with their fleeting nature during wakefulness and behavior, we used long-lasting steady-state stimuli. We analyzed neural activity recorded simultaneously by two eight-electrode arrays, each having a size of 1.5 x 0.5 mm, from middle layers (III/IV) of three auditory cortical areas under several 15-min duration conditions. These included silence (spontaneous activity) and stimulation with Poisson-distributed click trains (Eggermont and Smith 1995), low-pass noise amplitude-modulated wide-band noise, and multi-frequency tone-pip ensembles with random frequency and presentation time and presentation rates of 20, 120, and 240 pips/s (Tomita and Eggermont 2005; Valentine and Eggermont 2004). The electrode arrays were either both in primary auditory cortex (AI) or one was in AI and the other in the anterior auditory field (AAF) or posterior auditory field (PAF). The matrix of pair-wise peak correlation values was subjected to a hierarchical clustering procedure. Coincident-firing spectrotemporal receptive fields (STRF) were subsequently constructed for the clusters obtained. This was an extension of the coincident-firing pair STRFs reported previously (Tomita and Eggermont 2005).

Our findings suggest that cortical area boundaries nearly always act as cluster boundaries. Clusters were in large majority contiguous in space. Different clusters were formed under different stimulus conditions. Several operations on the coincident output of the clusters, as reflected in the common-spike and all-spike STRFs, may serve different modes of stimulus coding.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
Animal preparation

All animals were deeply anesthetized with the administration of 25 mg/kg of ketamine hydrochloride and 20 mg/kg of sodium pentobarbital, injected intramuscularly. A mixture of 0.2 ml of acepromazine (0.25 mg/ml) and 0.8 ml of atropine methyl nitrate (25 mg/ml) was administered subcutaneously at 0.25 ml/kg body wt. Lidocaine (20 mg/ml) was injected subcutaneously prior to incision. The tissue overlying the right temporal lobe was removed, and the dura was resected to expose the area bounded by anterior and posterior ectosylvian sulci. The cat was then secured with one screw cemented on the head without any other restraint. The wound margins were infused every 2 h with lidocain, and additional acepromazine/atropine mixture was administered every 2 h. The ketamine dose to maintain a state of areflexive anesthesia was on average 7.4 mg/kg*h (range: 3–12 mg/kg*h).

The care and the use of animals reported in this study was approved (BI 2001–021) and reviewed on a yearly basis by the Life and Environmental Sciences Animal Care Committee of the University of Calgary. All animals were maintained and handled according to the guidelines set by the Canadian Council of Animal Care.

Acoustic stimulus presentation

Stimuli were generated in MATLAB and transferred to the DSP boards of a TDT-2 (Tucker Davis Technologies) sound-delivery system. Acoustic stimuli were presented in an anechoic room from a speaker system [Fostex RM765 in combination with a Realistic super-tweeter that produced a flat spectrum (±5 dB) 40 kHz measured at the cat's head] placed 30° from the midline into the contralateral field, 50 cm from the cat's left ear. Calibration and monitoring of the sound field was accomplished with a condenser microphone (Bruel and Kjaer 4134) placed above the animal's head, facing the speaker and a measuring amplifier (Bruel and Kjaer 2636). Prior to acute recordings, peripheral hearing sensitivity was determined using auditory brain stem response (ABR) thresholds (details in Noreña et al. 2003).

Cortical frequency-tuning properties

Frequency-tuning curves were measured by randomly presenting 27 gamma-tone pips with frequencies covering five octaves (e.g., 1.25–40 kHz) in equal logarithmic steps and presented at eight different stimulus levels in 10-dB steps (e.g., 5–75 dB SPL) at a rate of 4/s such that each intensity-frequency combination was repeated five times. The envelope of the gamma tones is given by

(1)

with t in ms. The duration of the gamma tones at half-peak amplitude was 15 ms, and the envelope was truncated at 50 ms, where the amplitude is down by 64 dB compared with its peak value.

STRFs were obtained by presenting multi-frequency stimuli consisting of randomly presented gamma-tone pips. Here, tone pips for each of 81 frequencies in five octaves were randomly presented according to a Poisson process (Blake and Merzenich 2002), with similar average rate but different realization for each frequency. Three stimulus ensembles were used for which each tone pip frequency was presented at a rate of 0.25, 1.5, or 3 Hz so that the aggregate tone-pip rates were 20/s, 120/s, or 240/s.

Poisson-distributed click trains, with mean click rate of 8/s and dead time of 20 ms, and lasting 15 min were also presented with peak click levels equal to those for the gamma-tone pips. Wide-band noise (bandwidth: 40 kHz) modulated with a 30-Hz low-pass filtered noise and lasting 15 min was also presented.

Recording and spike separation procedure

Two arrays of eight electrodes (Frederic Haer) each with impedances between 1 and 2 M were used. The electrodes were arranged in a 4 x 2 configuration with inter-electrode distance within rows and columns equal to 0.5 mm. Each electrode array was oriented such that all electrodes were touching the cortical surface and then were manually and independently advanced using a Narishige M101 hydraulic microdrive (1 drive for each array). The depth of recording was between 700 and 1,200 µm, and thus the electrodes were likely in deep layer III or layer IV. The signals were amplified 10,000 times using a Frederic Haer HiZx8 set of amplifiers with filter cut-off frequencies set at 300 Hz and 5 kHz. The signals were processed by a TDT-Pentusa multi-channel data-acquisition system (filter bandwidth: 300 Hz to 10 kHz). Spike sorting was done off-line using a semi-automated procedure based on principal component analysis and K-means clustering implemented in MATLAB. The spike times and waveforms were stored. The multiple single-unit (MSU) data presented in this paper represent only well-separated single units that, because of their regular spike wave form, likely are dominantly from pyramidal cells. For statistical purposes, the separated single-unit spike trains were added again to form a multi-unit spike train, thereby eliminating potential contributions from thalamocortical afferents or fast spikes from interneurons.

Data analysis

FREQUENCY-TUNING. To assess frequency-tuning properties, the peak number of action potentials in the post stimulus time histogram (PSTH, 5-ms bins) calculated over the first 100 ms after gamma-tone presentation was estimated. The results were calculated per stimulus intensity and were combined into an intensity-frequency-rate profile from which tuning curves, rate-intensity functions, and iso-intensity rate-frequency contours could be derived (Eggermont 1996) using routines implemented in MATLAB. The frequency-tuning curve was defined for a firing rate at 25% of the maximum peak-firing rate (FRmax). This was 10–20% above the background firing-rate, but as the latter was dependent on the level of stimulus-induced suppression, the tuning curve criterion based on a percentage of peak firing rate was preferred over that based on increase over background activity. The values of firing rate given in this paper are in spikes/s. For that purpose, the number of spikes in a 5-ms bin is divided by the number of repetitions (5) times the bin size in seconds (0.005 s), i.e., by 0.025. This gives very high values of FRmax when the MSU spikes are well synchronized to the onset of stimulus. The threshold was determined from the tuning curve.

The STRF was determined by constructing PSTHs for each of the preceding gamma tones in a 100-ms window. For that purpose, each spike elicited was plotted several times in the appropriate frequency bins and in the 100-ms time window after the onset of each of the preceding gamma tones (because spike latency is a priori unknown). If the gamma tones had no effect on a spontaneously firing neuron, the entire matrix of 81 frequency bins by 50 (2-ms duration) time bins would be filled uniformly. If certain frequencies consistently produce excitation in a certain latency window, then this part of the frequency-time plane would receive more hits. In case certain frequencies produce consistently lower activity than average, this would be interpreted as inhibition (Tomita and Eggermont 2005; Valentine and Eggermont 2004).

For STRFs, typically obtained at 65 dB SPL, contour lines or color-coded images are plotted. STRF overlap was measured by taking the contour line representing 30% of the difference between the maximum response and the mean value of the STRF for each of the neurons, calculating the number of pixels for each neuron (each pixel defined as: 1 frequency bin by 1 2-ms time bin) as well as the number of overlapping pixels. The weighted STRF overlap was defined as [overlap size/sqrt(STRF1 size * STRF2 size)], which scaled between 0 (no overlap) and 1 (complete overlap). The minimum latency of the 30% excitation contour line, the temporal duration of the 30% excitation contour line, the frequency of the center of gravity, and the spectral extent of the 30% contour lines were determined for each stimulus.

Overlap of inhibitory areas was not calculated because the regions are typically shallow and their presence correlates with the peak strength of the excitatory area (Valentine and Eggermont 2004).

CORTICAL AREA BOUNDARIES. The following properties were used in the assessment of cortical area boundaries: reversal of the CF gradient in the tonotopic map and along the electrode array, minimum latency values, the shape of the STRF, and the peak value of the cross-correlation coefficient for recordings straddling boundaries. For delineating the border between AI and AAF, we first of all used the sign and/or reversal of the gradient of CF along the electrode array with distance in the anterior direction (Noreña and Eggermont 2005). The general shorter minimum latency in AAF compared with AI, and particularly the much higher frequency-tuning curve bandwidth at 20 dB above threshold in AAF (Eggermont 1998) were important as well. For the distinction between AI and PAF or potentially EPI (intermediate part of the posterior ectosylvian gyrus), we used mainly latency, which was 20 ms larger in non-AI areas. In addition, the sudden drop in peak cross-correlation coefficient across area boundaries under spontaneous firing conditions (Eggermont 2000) was a highly consistent indicator. It was, however, not possible to differentiate between PAF and EPI on any of those parameters, but on anatomical grounds, it is likely that the recordings were likely all from PAF, excluding the banks of the PES.

CROSS-CORRELATION. Cross-correlograms were calculated using custom made programs in Matlab (Eggermont 1992; Tomita and Eggermont 2005). Quantification of neural correlation in those studies was done on the basis of the cross-correlation coefficient

(2A)

which for relatively low firing rates reduces to

(2B)

where R_AB() is the number of coincidences in the bin corresponding to lag time , E is the expected value for coincidences under the assumption of independent spike trains, E = (N_AN_B)/N, with n = T/, where N_A and N_B are the number of spikes in the recording, is the bin size, and T the duration of the recording. |R()| 1. Stationarity estimates of the recordings were based on firing rate (mean and variance) in 100-s-long segments of the 15-min recordings. Functional correlation strength is not independent of the firing properties of the neurons in the pair. Specifically, periodicities in the neuronal firings imposed by cortical oscillations, e.g., in the spindle frequency range, may affect the peak cross-correlation coefficient (Eggermont and Smith 1995). Thus a deconvolution of the cross-covariance by the geometric mean of the auto-covariance functions of the two spike trains was implemented here. To correct for the overall firing rate, burst firing and periodicities in the firing of the neurons, the cross covariance, [R_AB() – E_AB], was deconvolved with the square root of the product of the autocovariance functions [R_AA() – E_A] and [R_BB() – E_B]. Here E_AB and E_A and E_B are the expected values for the cross- respectively auto-correlation functions under the assumption of independence and Poisson processes. This deconvolution was done in the frequency domain, where it becomes a simple division, Fourier transformation back to the time domain resulted in the corrected cross-correlation coefficient function R_C(). This correction procedure is an extension of the method used in Eggermont and Smith (1996) where the deconvolution was done with the auto-covariance of the trigger signal only. However, because most correlations in cortex result from shared input, the more symmetric form indicated above was used. The cross-correlations were all significantly different from zero at a level of 3 SD (P < 0.01).

Because none of the stimuli used in this study repeated in their 15-min duration, a standard shift predictor could not be used to estimate the effect of the stimulus locking on the value of R_C. For the Poisson-distributed clicks, the PSTH predictor was used (Perkel et al. 1967) and is based here on the cross-correlation between the respective first-order Poisson kernels (Epping and Eggermont 1986).

HIERARCHICAL CLUSTERING. The method of hierarchical clustering (Everitt 1978) applies when some clusters are nested within other clusters and the technique operates on a matrix of individual similarities or distances. In our case, we use pair-wise log(R_C) values, R_C being the peak value of the corrected cross-correlogram, or pair-wise STRF weighted overlap values, which serve as similarity measures. The log(R_C) was used because the distribution of log(R_C) values was approximately normal, and thus equal weight was given to high as well as low values. The implementation in MATLAB uses the nearest-neighbor single-linkage method to form clusters, which are visualized in a dendrogram. As inconsistency level a value of 1.0 was used. The goodness of fit of the dendrogram to the original similarities in the matrix is assessed by the cophenetic correlation coefficient, which is the Spearman rank correlation between the heights of the link in the dendrogram at which the two observations are first joined into a cluster and the original similarities of these two observations. The result from hierarchical clustering is generally unacceptable if the cophenetic correlation coefficient is <0.75.

For comparison, we also used hierarchical clustering based on the "ward" linkage method, which uses a minimum variance algorithm. Results were generally the same for multi-electrode clusters, but most of the nonclustered electrodes were clustered using this method. Inspection of those clusters showed that all poorly correlated electrodes were grouped using this method. As a consequence the nearest neighbor clustering was preferred.

STABILITY OF CLUSTERING UNDER DIFFERENT STIMULATION CONDITIONS. This was measured by calculating per electrode the stability of clustering and averaging this over the number of electrodes and the number of stimulus conditions. The cluster stability was calculated by assigning a value of one for each stimulus that resulted in the same largest cluster and then subtracting the number of conditions that resulted in different clusters. Thus if for a six-stimulus case, four stimuli result in the same cluster then the cluster value is 4 – 2 = 2. If the cluster size is equal to half the number of conditions, then the cluster value is 0. If the largest cluster was only obtained for two of six stimuli, then the cluster value is 2 – 4 = –2, etc. This is calculated for each electrode and averaged over the number of electrodes. The same analysis was done for six time segments (equal to 150 s) of the 900-s silence condition, and those results were used in a pair-wise comparison with the one for different stimulus conditions.

COMPARISON OF CLUSTERING BY STRF-OVERLAP AND PEAK R PER PP-STIMULUS. Here we calculated a simple correlation per stimulus condition, assigning a one to each electrode when the clustering was the same based on pair-wise peak correlation and on pair-wise STRF overlap. The value divided by the number of electrodes was tested against the expected value of obtaining that number of clusters under random assignments.

Locally weighted scatter plot smoothing (Lowess) curves were displayed in several figures. All statistical analyses were performed using Statview 5 (SAS Institute).

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
The results shown here were obtained from 30 sets of stationary recordings obtained from AI, PAF, and AAF in 16 adult cats. We selected recording sites with 12 functional electrodes and where at least four stimulus conditions each lasting 15 min gave stationary recordings. This comprised the calculation of 15,308 MSU pair correlations, and 5,712 pair-wise STRF overlaps.

Pair-wise cross-correlation based clustering

AN ILLUSTRATIVE EXAMPLE. We illustrate the basic method with data from a 15-min recording of spontaneous activity using two eight-electrode arrays in AI. The neighbor-electrode distances in an array are 0.5 mm and 15 of the 16 electrodes recorded spiking activity. Figure 1A shows a subset of electrode MSU-pair correlograms between electrode 1 and electrodes 3–5, electrodes 2 and 3–5, and electrode 3 with 4 and 5. A three-point smoothing is applied to eliminate spurious peaks here and in all correlograms to follow. This resulted in the autocorrelation (3–3) to be 0.6 in value rather than 1.0 because it is essentially only one bin wide. Each panel shows the correlograms scaled on their own extremes. The time base is from –0.3 to 0.3 s. The red traces show the standard correlograms calculated according to Eq. 2B; large oscillations with a frequency of 7 Hz, reflecting anesthesia-induced spindling activity are evident. The blue traces represent the corrected correlograms (R_C, see METHODS). Most of the oscillatory activity has disappeared and the peak values have been reduced. The peak R_C values are indicated in the right-hand corner of each panel. Figure 1B shows the full matrix of the logarithm of R_C, the subset shown in Fig. 1A corresponds to a region in the top-left corner (columns 1–3 with rows 3–5). Column numbers and row numbers correspond to electrode numbers as indicated. The dark row and column for electrode 9 indicate that this electrode was not recording spontaneous activity; the value of the peak correlation for this electrode was arbitrarily set at 0.01 (only for graphical purposes) to allow use of the full color scale for representing the matrix entries. In cluster calculations, we used a value of R_C = 0.001 for pairs with nonresponding electrodes. The color bar to the right indicates the range of log(R_C) values. Based on this pair-wise R_C matrix, a hierarchical clustering procedure (Fig. 1C) resulted in four clusters, one comprising activity on five electrodes (1–5), one comprising three electrodes (6–8), and two comprising two electrodes (13, 15 and 14, 16), and there were three electrodes that did not cluster with other electrodes (10, 11, 12). Electrode 9, which did not show spontaneous activity, also formed a separate cluster. Clusters 13–15 and 14–16 are very close to being clustered in one group. The multi-electrode clusters are indicated on the dendrogram with colored dots at their final branching points. In Fig. 1D, the electrode positions are indicated on the surface of the auditory cortex by the centers of the colored circles, and the clusters are indicated by the same colors as used in the dendrogram. Black circles reflect nonclustered electrodes. Clustered electrodes are grouped in close proximity and do not cross electrode-array boundaries. The size of the individual clusters does not extend much beyond 1 mm (inter-electrode distance is 0.5 mm). Both electrode arrays were in AI, based on the criteria outlined in the Methods section.

View larger version (63K): [in this window] [in a new window]   FIG. 1. Clustering procedure. The basis for the clustering procedure is the pair-wise cross-correlations (A) shown here for a few pairs in their traditional form (red curves) and the corrected form (blue curves). Each panel shows the correlograms scaled on their own extremes. The time base is from –0.3 to 0.3 s. The full pair-wise correlation matrix is shown for the spikes recorded on16 electrodes in B where the base 10 logarithm of the corrected correlation is shown (color bar indicates these values). Based on the pair-wise correlation matrix a hierarchical clustering procedure performed on the 16 electrodes (C) results in 4 multi-electrode clusters indicated at the branching points with a colored dot. These colors are used to identify the positions of the electrodes belonging to the same cluster on the surface of the auditory cortex (D). Electrodes are numbered on the cortical surface. All electrodes are in primary auditory cortex.

The average pair-wise R_C within a cluster is significantly larger than that between clusters, not surprising because this difference is the basis of cluster formation. Figure 2 shows the distribution of the logarithm of the average R_C values within clusters (Fig. 2A), the average R_C value between clusters (Fig. 2B), and the geometric mean of the within-cluster R_C values that participate in the between-cluster calculations (Fig. 2C). Best-fitting normal distributions are shown; the distributions appear to be approximately normally distributed. Figure 2D shows the scatter plot of between-cluster R_C values and the geometric mean of the corresponding within-cluster R_C values. All between-cluster R_C values are well below those for within cluster values (see Table 1).

View larger version (38K): [in this window] [in a new window]   FIG. 2. Within- and between-cluster pair-wise correlation strengths. A: distribution of the logarithm of the average within-cluster RC values. B: for average between-cluster RC values. C: distribution of the logarithm of the geometric mean of the average within-cluster RC values that participate in the between-cluster pairs. D: scatterplot of between-cluster RC values against the geometric mean of the corresponding within-cluster RC values.

View larger version (40K): [in this window] [in a new window]   FIG. 3. Flower plot of the stability of clustering for consecutive 150-s segments. Petal color indicates clusters. The numbers in the center of each flower indicates electrode number. White petals indicate that the electrode was not recording spike activity. Green petal color indicates nonclustered activity. Color changes for individual electrodes indicate signs of clustering instability.

View larger version (41K): [in this window] [in a new window]   FIG. 4. Flower plot of the stability of clustering under various stimulus conditions (indicated on the petals of the electrode 10 representation). Details same as in Fig. 3. The 240/s multi-tone stimulus induces large changes in the clustering.

Cluster stability was compared between stimulus effects and those of sectioning the 900 s of recording during silence into six subsets. A paired comparison showed that clusters were significantly more stable for different 150-s sections within a 900-s period than under different stimulus conditions (P < 0.01).

Table 1 shows the group means for within- and between-cluster R_C values per stimulus condition and also combined across all conditions (bottom row). Whereas the within-cluster values are all significantly larger (P < 0.0001) than the between-cluster values for any given stimulus, there were no significant differences in within- and between-cluster R_C values for different stimuli.

The fraction of electrodes that participated in the same cluster for all pairs of different stimulus conditions is shown in a similarity matrix form (Table 2). The highest fraction of common-clustered electrodes (0.747) is between two multi-tone stimuli with aggregate rates of 120/s and 240/s, whereas the lowest fraction (0.433) is found between the 20/s multi-tone stimulus and Poisson-distributed clicks. Multi-dimensional scaling on the dissimilarity matrix (unity matrix minus the similarity matrix) shows (Fig. 5A) that the clusters for silence, and the multi-tone stimuli with rates of 120/s and 240/s are relatively close, whereas those for the broadband, i.e., Poisson and low-pass amplitude-modulated noise, stimuli are very different from the other stimuli but also differ considerably among themselves (fraction of electrodes in the same clusters: 0.491). This is also reflected in the changes in cluster size. From silence to multi-tone stimuli, 70% of the cluster sizes did not change significantly and only 13% resulted in larger clusters and 17% is smaller clusters. In contrast, for the comparison of the effect of silence and broadband stimuli, one finds that in 40%, the size did not change, in 40%, became larger, and in 20%, became smaller. Nine simultaneous recordings from PAF and AI showed no significant differences in the fraction of electrodes that participated in the same cluster in PAF or AI. Clusters on the same electrode array were generally contiguous, only in 18/432 clusters did we observe otherwise. Of these 10 were found for the 20/s multi-tone stimulus and 4 for the 120/s multi-tone stimulus. The others were for Poisson-distributed clicks (2), low-pass amplitude modulated noise (1), and silence (1).

View larger version (18K): [in this window] [in a new window]   FIG. 5. A: multi-dimensional scaling of the average difference in the fraction of electrodes that participate in the same cluster for 6 stimulus conditions. For silence and multi-tone stimuli, the differences are relatively small. Large changes occur for Poisson-distributed click stimuli and low-pas amplitude modulated noise (lpamn). B: comparison of the stimulus-corrected peak cross-correlation coefficient (Rpsth) and the peak RC value for Poisson-distributed clicks.

For the Poisson clicks, the stimulus-predictor based on the cross-correlation of the average PSTHs (akin to a time-inverted 1st-order Poisson kernel) was calculated. The mean (SD) peak value of the kernels was 0.039 ± 0.041 spikes/click, with a range of –0.25–0.15 spikes/click. With this predictor, 1,231 cross-correlograms were calculated. The peak correlation coefficient R_C was strongly correlated with the one calculated with the PSTH predictor (r² = 0.965) and the regression line was: log₁₀(R_PSTH) = –0.062 + 0.97 * log₁₀(R_C). The relationship is shown Fig. 5B. In general, the PSTH-corrected R was slightly smaller than the one corrected on basis of the average firing rates, i.e., R_C. In incidental cases, the PSTH-corrected R was larger than R_C. This happened among others when one kernel was initially negative and the other initially positive. The cluster assignments based on R_PSTH and R_C were very similar: In 12 of the 13 recordings, the cluster assignment was exactly the same, in 1 recording the clustering based on R_PSTH split one cluster into 2. As a result, there was only a minute effect in the multivariate scaling plot (Fig. 5A).

CORTICAL AREAS AND CLUSTERING. Table 3 shows the mean R_C values for within- and between-cortical area clusters (across all stimulus conditions because according to ANOVA there was no interaction between area and stimulus type). Not all electrodes could be unambiguously assigned to an auditory cortical area, i.e., in case of nonresponsiveness to tones, so the total number is slightly less than the number of clusters in Table 1. One observes that the within-area, within-cluster R_C values are much higher than the within-area, between-cluster values, and those in turn are higher than the between-area, between-cluster values. The within-cluster R_C values are significantly larger (P < 0.0001) within AI than those within PAF and AAF, which are not significantly different from each other.

View larger version (33K): [in this window] [in a new window]   FIG. 6. Scatterplot of the log of RC as a function of the distance between electrodes. A random sample, comprising 50% of the data, is shown.

STRFS. An example set of STRFs obtained with the 120/s multi-tone stimulus is shown in Fig. 7. The electrode numbers read from left to right and top to bottom as indicated. The positions of the electrodes on the cortical surface are shown in Fig. 8D. Frequency is shown in kilohertz along the vertical axes using a logarithmic scale, and time since tone-pip onset is shown on the horizontal axes from 0 to 100 ms. The frequency-time plots basically show the frequency-dependent PSTH with 81 frequency bins and 50 time bins of 2-ms duration. Dark red colors indicate maximum response; dark blue colors indicate minimal response. The most general pattern consists of a strong excitation followed after 10–20 ms by post activation suppression, sometimes extending for a wider frequency range than the excitation. Best response frequencies range from just below 5 kHz (electrode 1) to just below 20 kHz (electrode 11). Some electrodes show broad sensitivity to frequency (electrodes 3 and 13). Electrodes 8 and 9 have weak responses and their STRFs are ill defined.

View larger version (83K): [in this window] [in a new window]   FIG. 7. Spectrotemporal receptive fields (STRF) for multiple single-unit activity simultaneously recorded on 16 electrodes from primary auditory cortex (AI). The electrodes shown in the top 2 rows were from an array located caudally from the electrodes forming the second array (bottom 2 rows). Consequently the characteristic frequencies are lower (covering about one octave from slightly below 5 kHz to slightly below 10 kHz) for the more caudal electrodes compared with the more rostral electrodes (covering frequencies from 10 kHz and up). Electrodes 8 and 9 show ill-defined STRFs.

View larger version (59K): [in this window] [in a new window]   FIG. 8. Clustering procedure based on STRF overlap. The basis for the clustering procedure is the pair-wise STRF weighted overlap (A) shown here for a few pairs. The full pair-wise overlap matrix is shown for the spikes recorded on16 electrodes in B (color bar indicates these values). C: scattergram of the pair-wise weighted overlap against the pair-wise RC values for this recording. The result of hierarchical clustering based on weighted overlap and on RC values is shown in the positions of the electrodes belonging on the surface of the auditory cortex (D). All electrodes were in AI. The left hand part of the circle fill color indicates the weighted overlap clustering and the right hand side the RC-based clustering.

GROUP DATA. Figure 9A shows the scatter gram of the average within-cluster pair-wise R_C and the average pair-wise weighted STRF overlap, and B shows the same for the average between-cluster values. The average weighted STRF-overlap per cluster and the average pair-wise log(R_C) are significantly correlated with r² = 0.2, those between clusters are correlated with r² = 0.36. Thus the pair-wise R_C between clusters is more dependent on STRF overlap than that within clusters, regardless of the fact that the range of overlap values and log(R_C) values is the same.

View larger version (33K): [in this window] [in a new window]   FIG. 9. Scatterplot of logarithm of the average within-cluster RC values against the average within-cluster STRF overlap (A). The same is shown for the average between-cluster values in B.

Whereas the average number of clusters based on R_C values (3.4 ± 1.0) was the same as that based on weighted overlap (3.4 ± 1.2), the type of clusters differed. We considered the following types: spatially contiguous clusters, noncontiguous clusters and inter area clusters. The average number of contiguous clusters per recording was significantly larger (P < 0.0001) for R_C based clustering (3.0 ± 1.0) compared with weighted overlap based clustering (2.2 ± 1.4). The average number of noncontiguous clusters was significantly lower (P < 0.0001) for R_C based clustering (0.4 ± 0.6) compared with weighted overlap based clustering (0.9 ± 0.9). This was also the case (P = 0.0001) for the average number of inter-area R_C based clusters (0.1 ± 0.3) compared with weighted overlap based clusters (0.4 ± 0.7).

Coincident-firing cluster STRFs versus all-spike cluster STRFs

Figure 10 shows an example of the STRFs in an R_C-based cluster for stimulation with the 20/s multi-tone stimulus ensemble. The cluster contains the activity of 12 electrodes across both electrode arrays positioned within AI (left 3 columns). The percentage of common spikes, defined as those that occur within a 10-ms window across all these 12 electrodes, is only 0.7 but still results in a clearly defined common-spike STRF (top panel in right column) with a CF around 7 kHz. The merged activity of all spikes from the 12 electrodes in the cluster results in an STRF that is shown at the second row of right-most column and the 30% contour line (between peak and mean value) of the common-spike based STRF is superimposed. The common-spike STRF corresponds to the lower half of the all-spike STRF and consequently is not just a diluted version of the all-spike STRF. The signal-to-noise ratio (SNR), defined as 20*log₁₀(peak/mean), of the common-spike STRF is 15.3 dB, whereas that for the merged-activity STRF is 4.3 dB. This suggests that synchronized neural activity can enhance signal-to-noise ratios.

View larger version (46K): [in this window] [in a new window]   FIG. 10. STRFs belonging to 1 12-electrode cluster for 20/s multi-tone stimulation (left 3 columns), the common-spike STRF (top row, right column), and the all-spike STRF (2nd row, right column). The common-spike STRF represents 0.7% of the spikes recorded on the 12 electrodes and shows a clear activity region between 5 and 10 kHz and 15–25 ms after stimulus onset. The contour of this common-spike STRF is shown superimposed on the all-spike STRF and corresponds with the lower half of it.

View larger version (42K): [in this window] [in a new window]   FIG. 11. Comparison of common-spike STRF and all-spike STRFs for clusters of several sizes. The 4 recordings represent different animals. In 3 of the 4 cases, the common- and all-spike STRFs overlap nearly completely. In 1 case (B), the common-spike STRF is again much smaller than the all-spike one. The signal-to-noise ratios, defined as peak value divided by the mean (indicated on top of each panel) are larger for the common-spike STRFs compared with the all-spike STRFs.

View larger version (18K): [in this window] [in a new window]   FIG. 12. Scattergram of the fraction of common spikes against the ratio of the peak values of common-spike and all-spike STRFs (A). Scattergram of the signal-to-noise ratio for the common-spike STRF against the signal-to-noise ratio for the all-spike STRF (B).

Functional cluster properties

If individual clusters from the same multi-electrode recording show quite different cluster STRFs, one is inclined to consider the clusters functionally different, as they will respond preferentially to different stimuli. This is not always that clear. For example, Fig. 13 shows a set of common- and all-spike STRFs for four clusters from the same recording that feature only slight differences across clusters and those are largely in CF and frequency bandwidth. This tonotopic gradient with very similar excitatory STRFs shapes was the most common finding when the electrode arrays were both in AI. The common-spike STRFs overlap substantially with the all-spike ones (30% contour lines for common-STRF shown on the all-spike STRF), as also exemplified by the large percentage of common spikes (indicated above each panel).

View larger version (40K): [in this window] [in a new window]   FIG. 13. Comparison of the STRFs for 4 clusters of activity recorded simultaneously in response to the 120/s multi-tone stimulus. The clusters from left to right comprise activity of 3, 4, 2, and 2 electrodes. One observes that there is clearly overlap of the STRFs for different clusters but that the center frequencies of the STRF are different. The common-spike STRFs (top row) overlap with the all-spike STRFs (bottom row).

View larger version (79K): [in this window] [in a new window]   FIG. 14. Comparison of the STRFs for 4 clusters of activity recorded simultaneously in response to the 20/s multi-tone stimulus (top 2 rows) and the 240/s multi-tone stimulus (bottom 2 rows) from the same recording sites. Cluster sizes were the same for clusters in columns 2 and 4 and showed 1-electrode difference for clusters in columns 1 and 3. Here, for both stimulus conditions, the STRFs are very different for the 4 clusters.

View larger version (43K): [in this window] [in a new window]   FIG. 15. An example of a 12-electrode cluster (STRFs shown in the 3 left columns) without a common-spike STRF (shown in the top row, right column).

View larger version (24K): [in this window] [in a new window]   FIG. 16. Scattergrams of the logarithm of the fraction of common-spikes against the number of electrodes participating in the cluster (A) and against the spatial zize of the cluster (B). The data points are shown separately for conditions with a clear common-spike STRF () and absence of a common-spike STRF ().

This was further analyzed for five randomly selected multi-electrode recordings. For each set of recordings, the fraction of common spikes was calculated for all possible pairs of electrodes and then the mean value thereof was used to predict the expected values for all possible three, four, five, six, etc electrode combinations. In all cases, the prediction on basis of that mean values was slightly larger that the actual mean values for the various electrode combinations.

Based on these calculations, one cannot reject the assumption that the decrease in the fraction of common spikes with the number of electrodes in the cluster is due to some independence in firing to the stimulus. This also suggests that most pair-wise correlations contain significant contributions of nonstimulus-related spikes (but likely not driven by brain rhythms in the >1 Hz range because those effects are corrected for in the calculation of R_C).

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
We have shown that neural activity recorded from the auditory cortex can be clustered on basis of pair-wise cross-correlation. The resulting neuron clusters are local and typically do not cross cortical-area boundaries and are even confined to a particular electrode array in the majority of cases. As a consequence, the spatial size of the clusters rarely exceeds a few mm and on average is 1 mm. Cluster size was significantly larger in PAF compared with AI, whereas the fraction of common spikes (within a 10-ms window) across electrodes participating in a cluster was significantly higher in AI compared with PAF. The difference in the fraction of common spikes could result from the longer latencies and hence larger SD in first spike firing of PAF neurons (Phillips and Orman 1984). The average cluster size may reflect the decrease in R_C with distance (Fig. 6). However, within AI this decrease is much slower with distance than in PAF, yet cluster size is on average larger in PAF. Thus other factors play a role in the local changes in R_C. For instance the difference between the average within-cluster R_C values is much larger in AI than in PAF (Table 2) so that the cluster boundaries may be less sharp in PAF.

Large R_C-based clusters comprising activity from both multi-electrode arrays were observed only 11 times (4.5%). These large clusters were never found for silence or low-pass amplitude-modulated noise and were most frequently encountered for the Poisson-distributed click trains and the 20/s multi-tone stimulus. These two stimuli have distinct onsets with sufficiently long silent periods between the stimuli. In general, cluster boundaries changed under different stimulus conditions. Taking the cluster sizes found for spontaneous firing as a reference, Poisson-distributed clicks and low-pass amplitude-modulated noise produce the largest changes in cluster size, whereas multi-tone stimuli have a more modest effect. This suggests that the clusters found under spontaneous firing conditions are largely determined by the same common inputs that govern the STRF properties of the neurons, i.e., the specific thalamic afferents. Topographic mapping of activity in visual cortex using intrinsic optical signals also showed a strong similarity in the locations of patches of synchronous spontaneous firing and evoked firing of the order of 1 mm in size (Tsodyks et al. 1999). This is furthermore reflected in the fact that 70% of the cluster sizes did not change significantly between the silence and multi-tone stimuli condition, whereas only 40% stayed the same in comparison between silence and broadband stimulation. Thus broadband stimulation with low-pass amplitude-modulated noise has very different effects than stimulation with a multi-tone 240/s stimulus, which has a similar amplitude-modulation spectrum and the same long-term carrier spectrum but of course sounds completely different.

Coarse-grained recording and homogeneity in response

Responses for units recorded with the same electrode in AI show strong homogeneity for CF, threshold at CF, 20-dB bandwidth, and minimum latency (Eggermont 1996). The spatial representation in AI also changes slowly for those parameters (Schreiner et al. 2000). The information about a stimulus in AI is most likely contained in this slowly changing spatial representation of neural activity, i.e., in a population code, thus the size and shape of the clustered activity should be important. It is therefore assumed that for mapping those parameters one can use a relatively coarse sampling as done in the present study. There is one caveat, the continuity in mapping is generally observed for near threshold properties (CF, threshold at CF, 20-dB bandwidth), whereas the stimuli used in the present study were all presented at 65 dB p.e. SPL, and thus well above the level of the 20 dB bandwidth of the neurons. For those intensity levels, tone pips activated a large (4 mm along an isofrequency contour as well as perpendicular to it) and patchy area in AI (Phillips et al. 1994), so there is no a priori reason to expect spatially contiguous clusters. Furthermore, our two electrode arrays were mostly positioned in a caudal to rostral alignment (as in Fig. 8) and only infrequently within the direction of the isofrequency sheets. Yet, 89% of the R_C-based clusters on the same electrode array were contiguous. In contrast, and probably relating more to the Phillips et al. (1994) findings, 26% of clusters based on STRF overlap were noncontiguous. This suggests that despite the potentially significant STRF overlap for neurons in different patches, there is probably a weaker correlation between different patches than within a single patch.

Should one correct for stimulus-correlation?

Calculating a shift predictor or joint-PSTH predictor is one way to extract the effects of stimulation on the effective neural connectivity (Espinoza and Gerstein 1988). However, as we have previously argued (Tomita and Eggermont 2005), it is unlikely that the nervous system performs a correction for stimulus-induced correlation as estimated by the various predictors. It is the actual spike coincidences that are affecting the potential for firing in a target neuron not the stimulus-corrected ones. Thus our raw correlations may effectively estimate those coincident firings between neurons that could play a role in neural population coding of sound. The effect of removing common periodicities in spike firing and common bursting activity form the cross-correlation can be justified because these contributions typically do not occur in awake animals. Nevertheless, our data show that in the case of the Poisson-distributed clicks, the effects of a stimulus correction based on the first-order Poisson kernels on clustering are minimal and do not affect our conclusions. In case of the multi-frequency stimuli, one expects a larger reduction of the peak cross-correlation from corrections for stimulus locked contributions. This is in first approximation shown by the clustering on basis of STRF overlap. The STRFs are a linear predictor of stimulus effects on the firing rate (or vice versa), and the overlap of STRFs of two recordings is a linear estimator of the fraction of spikes they have in common. Thus the STRF-overlap would correspond to the area under the PSTH-product predictor. Given that the STRF overlap explains 38% of the variance in the pair-wise R_C, and that R_C-based clusters for multi-frequency stimuli are similar, but not identical, to those based on spontaneous activity, one has to conclude that clustering based on stimulus-corrected correlation coefficients would also be similar.

Basis for clustering

Cross-correlation measures allow the derivation of the functional ordering of neurons exemplified in the pair-wise peak-correlation matrix. Functional organizations that show topographic order, such as the tonotopic organization of the auditory cortex, are in fact only observable by the experimenter, whereas the animal has likely to use correlation measures to access that order (Koenderink 1984a,b). From the similarity of R_C-based and STRF-overlap-based clustering, one could, however, claim that the correlation is largely caused by the topographic order of the common inputs. However, <40% of the variance in R_C was explained by STRF overlap, suggesting dominant other contributions.

The correlation measure used is calculated for multiple single-unit activity based on spike-sorting procedures. How many neurons are recorded from with a single microelectrode? Spike separation usually results in two to three well-resolved units and a remaining spike cluster that cannot be resolved. Typically the firing rates of the sorted groups are of the same order of magnitude and can each reach 10/s (Eggermont et al. 1993; Legendy and Salcman 1985). However, optical recordings with single-neuron resolution in vivo from motor cortex in urethan-anesthetized rats and follow-up patch-clamp recordings suggest that individual cortical neurons have a low probability of firing (Kerr et al. 2005). Spontaneous activity in layer II/III of motor cortex during up states was found to be <0.1/s, and only 10% of neurons were simultaneously active during an up state. DeWeese et al. (2003) found that patch-clamped neurons in auditory cortex (ketamine anesthesia) fired either 0 or 1 spike. These findings may suggest a different selection bias for standard extra cellular recordings and patch-clamp recordings (see also Margrie et al. 2002), they may indicate a profound cortical layer difference in firing rate or suggest that the number of units contributing to a well-sorted cluster in extra cellular recordings is grossly underestimated. In awake paralyzed cats, 40% of neurons in AI had spontaneous firing rates <1 spike/s, but no significant layer difference in response to tones was found (Abeles and Goldstein 1970). If the number of contributing units is underestimated, even well-separated single-unit recordings must then be comprised of activity from 10 not-well-synchronized units (otherwise the spikes would be superimposed and likely sorted as 1 unit).

Clustering is based on the distribution of R_C values across electrode pairs. A potential cause for changing correlation strengths under different stimulus conditions is the use of unresolved multi-unit activity on pairs of electrodes to probe neural correlations (Bedenbaugh and Gerstein 1997; Gerstein 2000). As these authors note, the peak cross-correlation coefficient between small groups of units is generally larger than the cross-correlation between a pair of single units from each group. However, the measured inter-electrode correlation is reduced in turn by a nonlinear function of the cross-correlations between the single units recorded on each electrode. This becomes important if the number of units recorded from under different stimulus conditions would change, and this would in turn change the measured cross-correlation strength between the two unit clusters. From our spike-separation procedure, which resulted in reasonable well-isolated spike-waveforms, we could not detect new units activated by different stimuli. Thus the changes in R_C clustering are likely not the result of sampling different sets of units for different stimuli. Although if one accepts that single units fire <1 spike/s, then even close similarity in spike waveform cannot rule out activation of different units by different stimuli.

What causes changes in clustering?

Following the definition of cluster membership, changes in R_C-based clustering are the result of changes in the distribution of pair-wise cross-correlation. In its most simple form, changes in clusters could result from recording different units for different stimulus conditions and hence obtaining different correlation strengths. As discussed in the preceding text, I consider this not very likely.

Changing anesthesia levels can also cause changes in the peak cross-correlation coefficient (Erchova et al. 2002). With deeper anesthesia levels, the spontaneous firing activity recorded with microelectrode arrays in rat barrel cortex took the shape of periodic bursting with periods of 2–2.5s, and the firings became more synchronized between barrels. The peak correlation coefficient for this spontaneous activity increased by about a factor 4 between light and deep urethan anesthesia (Erchova et al. 2002). It is possible that in the course of our 1- to 1.5-h recording that comprised the four to six different stimulus presentations of 15 min each, the anesthesia level changed sufficiently to affect the cross-correlation strength. In contrast to Erchova et al. (2002), we correct for the effect of firing rate, common periodicities and bursting behavior in the neurons on the cross-correlation, so we expect that slight changes in anesthesia levels will have very little effect on the peak R_C values. This is supported by the fact that both the mean value and the variance of the pair correlation matrix did not change significantly with time.

It is also possible that the difference in the clustering of neurons for a long-lasting period of silence and for the other stimuli is related to a state change in the brain caused by appropriate stimulation (Miller and Schreiner 2000). These authors showed that during stimulation with dynamic-ripple noise the spindle-wave-like periodicity in the autocorrelograms of individual units recorded from the auditory thalamus disappeared. It was argued that the random frequency and temporal changes in such a stimulus, clamped the thalamocortical system in an "aroused" state. The multi-tone stimuli that we use also have this property (we found that during presentation of these stimuli the 7- to 14-Hz spindle rhythm in cortical local field potentials and in spike autocorrelograms can be completely suppressed; T. Britvina and J. J. Eggermont, unpublished observations).

One of the properties attributed to a neural assembly is its ability to change neuron membership as a result of a change in brain state (Abeles et al. 1995). The multi-tone stimuli that we employ have a small but distinct effect on the R_C values and on the clustering compared with that under spontaneous activity. This suggests that our clustering procedure, using cross-correlation corrected for common spindle-induced activity, may still recover brain-state-dependent changes produced by the multi-tone stimuli. This brain-state dependence is another property that the neuron clusters have in common with neural assemblies. However, the most apparent cluster changes occurred for stimulation with Poisson-distributed clicks and low-pass amplitude-modulated noise. It is therefore likely that the different set of thalamocortical afferents activated during the presentation of different stimuli have more effect on the clustering observed than the effects of changing anesthesia levels or changes in brain state.

Do the clusters that we observe relate to neural assemblies?

Organization of neural activity in small, local, clusters may be the result of competitive interactions resulting from the demand to map several stimulus parameters on the same cortical surface (Schreiner 1995). Neural assemblies are generally proposed to explain particular forms and properties of population coding. The point is then whether neural clusters and neural assemblies have a common base. The brain-state dependence in cluster membership mentioned in the preceding text is one property that the neuron clusters have in common with putative neural assemblies. Our clusters appear to be largely confined by cortical area boundaries and are affected by stimulus type. Only three R_C-based clusters (4.6% of the number of simultaneous PAF and AI recordings) were found that crossed area boundaries potentially because of a drop in pair-wise correlation strength across area boundaries as noticed before (Eggermont 2000). This drop in cross-correlation strength may reflect the paucity of bifurcating axons that innervate cortical cells in different cortical areas (shown for AAF and AI) (Lee et al. 2004a), thereby only allowing a modest common input from thalamic neurons to cortical neurons. This forms another argument for our assumption that the clustering reflects anatomical connectivity. In contrast, STRF overlap-based clustering resulted in 15% of the number of simultaneous PAF and AI recordings as inter-area clusters, which was significantly higher (P < 0.005) than for R_C-based clusters. Assuming that overlap in STRF implies overlap in thalamocortical afferent input, but not necessarily provided by bifurcating axons, this suggests that purely stimulus-based correlations do not play a large role in the R_C-based clustering compared with connectivity-based correlations.

Anatomical connectivity may also underlie the synchrony for sensitivity to visual stimulus orientation in spatially segregated patches. For example, gamma-oscillation synchrony in the visual cortex crosses area boundaries and is driven by common sensitivity to stimulus aspects such as orientation sensitivity has also been related to neural assembly activity. This inter-area synchrony was frequently observed (König et al. 1995) and could reflect a more widespread common input between different cortical visual areas compared with the auditory ones. It should be mentioned that gamma oscillations are conspicuously absent in spike firing from auditory cortex (Horikawa et al. 1994). Comparison of the topological organization of the auditory and visual systems (Scannell et al. 1995) shows that striate cortex is interconnected with eight other visual cortical areas and AI with five other auditory cortical areas, albeit that Winer (1992) lists interconnections of AI with seven areas. This suggests no a priori differences in connectivity. Whereas in the visual system intercortical connections largely reflect similar orientation columns and ocular dominance columns (Schmidt et al. 1997), in the auditory system, this connectivity may reflect besides frequency tuning only binaural interaction type (Winer 1992). The fact that our stimuli were presented in the contralateral field may have restricted the probability to observe inter-cortical area clusters. Another reason might be that PAF and AI are not directly connected (Rouiller et al. 1991) and any clustering would thus be based solely on shared thalamo-cortical afferents.

R_C and weighted STRF overlap values in AI were larger than for PAF and AAF, suggesting that the divergent thalamocortical fiber activity is either more synchronized in AI or that the afferent fibers make stronger synaptic connections with the pyramidal cells than is the case in PAF and AAF. The longer response latencies and the temporally less precise STRFs in PAF indicate that the lower input synchrony in PAF plays a large role in its lower R_C values. In cat AAF, latency is typically shorter than in AI (Eggermont 1998) and STRFs are well defined, but within-cluster R_C values are even smaller than in PAF. Therefore the different origin of the thalamo- and cortico-cortical afferents and their connection strengths in the three areas may be at the base of the differing R_C values.

The STRF determines part of the functional operation of a cluster in the auditory cortex. The STRFs can be considered as spectrotemporal filters that take partial derivatives of the sound, i.e., in time (modulation envelope) and frequency (Fishbach et al. 2001; Koenderink and van Doorn 1990) and thus act as modulation selective filters (Chi et al. 2005). One can ask what is more important for describing the functional properties of a cluster, its common-spike STRF or the one calculated from all spikes from the units taking part in the cluster. The common-spike STRF is a subset of the all-spike STRF. In some cases, however, properties from the common-spike STRF largely mimic those from the all-spike STRF (Fig. 11, A, C, and D).

One of the things a cluster accomplishes is that it extracts common-spike STRFs with higher signal-to-noise ratio than that of population STRFs that result from merging all spikes emitted by the cells participating in the cluster. This suggests that an all-spike STRF is preferable over a common-spike STRF as far as stimulus encoding is concerned. This, however, has to be tested.

Limitations of this clustering approach

The clustering was deliberately studied for steady-state stimuli that each lasted 15 min and allowed stationary recordings and good statistics for the neuronal firing and synchrony. In reality, the brain has to deal with activity that changes meaning and context over time scales of the order of seconds (Aertsen et al. 1994; Ahissar et al. 1992; Eggermont 1994). Thus time-dependent evaluation of correlations and dynamical clustering need to be considered. The statistically less robust results of using short-duration spike trains will likely result in less stable clusters, which either reflect the noisiness of the data or the actual transient nature of neural assemblies. However, the gravitational clustering method in its newest adaptation (Lindsey and Gerstein 2006) has potential for this purpose.

Is pair correlation sufficient or even meaningful as a basis for clustering? It is for linear systems. It is also sufficient for a nonlinear system with only feed-forward connections, but it is not in the case of nonlinear systems with feedback connections (Johannesma et al. 1986). Because the thalamo-cortical system is obviously nonlinear and has strong feedback connections, one potentially needs to extent this analysis by incorporating nonlinear relations, e.g., by using mutual information as a basis for clustering.

Summary

Neuron clusters in auditory cortex formed on basis of the neural correlation matrix, reflecting the firing synchrony across cortical areas of several mm², expand and contract in response to different stimuli. The most striking result was that the cluster positions and size were very similar for spontaneous activity and multi-tone stimuli evoked activity. Clusters rarely comprised neurons from more than one cortical area, as follows from the decrease in average correlations strength across area boundaries. Clusters are therefore not synonymous with the view of neural assemblies, advanced for visual cortex, that assumes synchrony across several areas. Several operations on the coincident output of the clusters, as reflected in the common-spike and all-spike STRFs, may serve different modes of stimulus coding.

	GRANTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
This work was supported by the Alberta Heritage Foundation for Medical Research, by the Natural Sciences and Engineering Research Council, and by the Campbell McLaurin Chair of Hearing Deficiencies.

	ACKNOWLEDGMENTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
G. Shaw provided programming assistance. A. Noreña, M. Tomita, and N. Aizawa helped with the data collection. A. Noreña and B. Gourevitch commented on an earlier version of this manuscript.

	FOOTNOTES

 
The costs of publication of this article were defrayed in part by the payment of page charges. The article must therefore be hereby marked "advertisement" in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

Address for reprint requests and other correspondence: J. J. Eggermont, Dept. of Psychology, 2500 University Drive N.W., University of Calgary, Calgary, Alberta T2N 1N4, Canada (E-mail: eggermon{at}ucalgary.ca)

	REFERENCES

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION GRANTS ACKNOWLEDGMENTS REFERENCES

 
Abeles M, Bergman H, Gat I, Meilijson I, Seidemann E, Tishby N, and Vaadia E. Cortical activity flips among quasi-stationary states. Proc Natl Acad Sci USA 92: 8616–8620, 1995.[Abstract/Free Full Text]

Abeles M and Goldstein MH Jr. Functional architecture in cat primary auditory cortex: columnar organization and organization according to depth. J Neurophysiol 33: 172–187, 1970.[Free Full Text]

Aertsen A, Erb M, and Palm G. Dynamics of functional coupling in the cerebral cortex: an attempt at a model-based interpretation. Physica D 75: 103–128, 1994.

Ahissar E, Vaadia E, Ahissar M, Bergman H, Arieli A, and Abeles M. Dependence of cortical plasticity on correlated activity of single neurons and on behavioral context. Science 257: 1412–1415, 1992.[Abstract/Free Full Text]

Bedenbaugh P and Gerstein GL. Multiunit normalized cross correlation differs from the average single-unit normalized correlation. Neural Comput 9: 1265–1275, 1997.[CrossRef][Web of Science][Medline]

Ben-Shaul Y, Bergman H, Ritov Y, and Abeles M. Trial to trial variability in either stimulus or action causes apparent correlation and synchrony in neuronal activity. J Neurosci Methods 111: 99–110, 2001.[CrossRef][Web of Science][Medline]

Blake DT and Merzenich MM. Changes of AI receptive fields with sound density. J Neurophysiol 88: 3409–3420, 2002.[Abstract/Free Full Text]

Brody CD. Correlations without synchrony. Neural Comput 11: 1537–1551, 1999a.[CrossRef][Web of Science][Medline]

Brody CD. Disambiguating different covariation types. Neural Comput 11: 1527–1535, 1999b.[CrossRef][Web of Science][Medline]

Chi T, Ru P, and Shamma SA. Multiresolution spectrotemporal analysis of complex sounds. J Acoust Soc Am 118: 887–906, 2005.[CrossRef][Web of Science][Medline]

Clarke S, de Ribaupierre F, Rouiller EM, and de Ribaupierre Y. Several neuronal and axonal types form long intrinsic connections in the cat primary auditory cortical field (AI). Anat Embryol 188: 117–138, 1993.[Medline]

Cohen YE and Knudsen EI. Maps versus clusters: different/ representations of auditory space in the midbrain and forebrain. Trends Neurosci 22: 128–135, 1999.[CrossRef][Web of Science][Medline]

DeWeese MR, Wehr M, and Zador AM. Binary spiking in auditorycortex. J Neurosci 23: 7940–7949, 2003.[Abstract/Free Full Text]

Eggermont JJ. Neural interaction in cat primary auditory cortex. Dependence on recording depth, electrode separation and age. J Neurophysiol 68: 1216–1228, 1992.[Abstract/Free Full Text]

Eggermont JJ. Neural interaction in cat primary auditory cortex. II. Effects of sound stimulation. J Neurophysiol 71: 246–270, 1994.[Abstract/Free Full Text]

Eggermont JJ. How homogeneous is cat primary auditory cortex? Evidence from simultaneous single-unit recordings. Audit Neurosci 2: 76–96, 1996.

Eggermont JJ. Representation of spectral and temporal sound features in three cortical fields of the cat. Similarities outweigh differences. J Neurophysiol 80: 2743–2764, 1998.[Abstract/Free Full Text]

Eggermont JJ. Sound induced correlation of neural activity between and within three auditory cortical areas. J Neurophysiol 83: 2708–2722, 2000.[Abstract/Free Full Text]

Eggermont JJ and Smith GM. Separating local from global effects in neural pair correlograms. Neuroreport 6: 2121–2124, 1995.[Web of Science][Medline]

Eggermont JJ and Smith GM. Neural connectivity only accounts for a small part of neural correlation in auditory cortex. Exp Brain Res 110: 379–392, 1996.[Web of Science][Medline]

Eggermont JJ, Smith GM, and Bowman D. Spontaneous burst firing in cat primary auditory cortex: age and depth dependence and its effect on neural interaction measures. J Neurophysiol 69: 1292–313, 1993.[Abstract/Free Full Text]

Epping WJ and Eggermont JJ. Sensitivity of neurons in the auditory midbrain of the grassfrog to temporal characteristics of sound. I. Stimulation with acoustic clicks. Hear Res 24: 37–54, 1986.[CrossRef][Web of Science][Medline]

Erchova IA, Lebedev MA, and Diamond ME. Somatosensory cortical neuronal population activity across states of anesthesia. Eur J Neurosci 15: 744–752, 2002.[CrossRef][Web of Science][Medline]

Espinosa IE and Gerstein GL. Cortical auditory neuron interactions during presentation of 3-tone sequences: effective connectivity. Brain Res 450: 39–50, 1988.[CrossRef][Web of Science][Medline]

Everitt BS. Graphic Techniques for Multivariate Data. London: Heinemann Educational Books, 1978.

Fishbach A, Nelken I, and Yeshurun Y. Auditory edge detection: a neural model for physiological and psychoacoustical responses to amplitude transients. J Neurophysiol 85: 2303–2323, 2001.[Abstract/Free Full Text]

Gerstein GL. Cross-correlation measures of unresolved multi-neuron recordings. J Neurosci Methods 100: 41–51, 2000.[CrossRef][Web of Science][Medline]

Gerstein GL and Kirkland KL. Neural assemblies: technical issues, analysis, and modeling. Neural Networks 14: 589–598, 2001.[CrossRef][Web of Science][Medline]

Harris KD. Neural signatures of cell assembly organization. Nat Rev Neurosci 6: 399–407, 2005.[CrossRef][Web of Science][Medline]

Horikawa J, Tanahashi A, and Suga N. After-discharges in the auditory cortex of the mustached bat: no oscillatory discharges for binding auditory information. Hear Res 76: 45–52, 1994.[CrossRef][Web of Science][Medline]

Johannesma P, Aertsen A, van den Boogaard H, Eggermont J, and Epping W. From synchrony to harmony: ideas on the function of neural assemblies and on the interpretation of neural synchrony. In: Brain Theory, edited by Palm G and Aertsen A. Berlin: Springer-Verlag, 1986, p. 25–47.

Kerr JN, Greenberg D, and Helmchen F. Imaging input and output of neocortical networks in vivo. Proc Natl Acad Sci USA 102: 14063–14068, 2005.[Abstract/Free Full Text]

Kimpo RR, Theunissen FE, and Doupe AJ. Propagation of correlated activity through multiple stages of a neural circuit. J Neurosci 23: 5750–5761, 2003.[Abstract/Free Full Text]

Kistler WM and Gerstner W. Stable propagation of activity pulses in populations of spiking neurons. Neural Comput 14: 987–997, 2002.[CrossRef][Web of Science][Medline]

Koenderink JJ. Simultaneous order in nervous nets from a functional standpoint. Biol Cybern 50: 35–41, 1984a.[CrossRef][Web of Science][Medline]

Koenderink JJ. Geometrical structures determined by the functional order in nervous nets. Biol Cybern 50: 43–50, 1984b.[CrossRef][Web of Science][Medline]

Koenderink JJ and van Doorn AJ. Receptive field families. Biol Cybern 63: 291–297, 1990.[CrossRef][Web of Science]

Konig P, Engel AK, and Singer W. Relation between oscillatory activity and long-range synchronization in cat visual cortex. Proc Natl Acad Sci USA 92: 290–294, 1995.[Abstract/Free Full Text]

Lampl I, Reichova I, and Ferster D. Synchronous membrane potential fluctuations in neurons of the cat visual cortex. Neuron 22: 361–374, 1999.[CrossRef][Web of Science][Medline]

Lee CC, Imaizumi K, Schreiner CE, and Winer JA. Concurrent tonotopic processing streams in auditory cortex. Cereb Cortex 14: 441–451, 2004a.[Abstract/Free Full Text]

Lee CC, Schreiner CE, Imaizumi K, and Winer JA. Tonotopic and heterotopic projection systems in physiologically defined auditory cortex. Neuroscience 128: 871–887, 2004b.[CrossRef][Web of Science][Medline]

Legendy CR and Salcman M. Bursts and recurrences of bursts in the spike trains of spontaneously active striate cortex neurons. J Neurophysiol 53: 926–939, 1985.[Abstract/Free Full Text]

Lindsey BG and Gerstein GL. Two enhancements of the gravity algorithm for multiple spike train analysis. J Neurosci Methods 150: 116–127, 2006.[CrossRef][Web of Science][Medline]

Margrie TW, Brecht M, and Sakmann B. In vivo, low-resistance, whole-cell recordings from neurons in the anesthetized and awake mammalian brain. Pfluegers 444: 491–498, 2002.

Miller LM and Schreiner CE. Stimulus-based state control in the thalamocortical system. J Neurosci 20: 7011–7016, 2000.[Abstract/Free Full Text]

Mountcastle VB. Perceptual Neuroscience: The Cerebral Cortex. Harvard: Harvard University Press, 1998.

Noreña AJ and Eggermont JJ. Enriched acoustic environment after noise trauma reduces hearing loss and prevents cortical map reorganization. J Neurosci 25: 699–705, 2005.[Abstract/Free Full Text]

Noreña AJ, Tomita M, and Eggermont JJ. Neural changes in cat auditory cortex after a transient pure-tone trauma. J Neurophysiol 90: 2387–2401, 2003.[Abstract/Free Full Text]

Perkel DH, Gerstein GL, and Moore GP. Neuronal spike trains and stochastic point processes. II. Simultaneous spike trains. Biophys J 7: 419–440, 1967.[Web of Science][Medline]

Phillips DP and Orman SS. Responses of single neurons in posterior field of cat auditory cortex to tonal stimulation. J Neurophysiol 51: 147–163, 1984.[Abstract/Free Full Text]

Phillips DP, Semple MN, Calford MB, and Kitzes LM. Level-dependent representation of stimulus frequency in cat primary auditory cortex. Exp Brain Res 102: 210–226, 1994.[Web of Science][Medline]

Reyes AD. Synchrony-dependent propagation of firing rate in iteratively constructed networks in vitro. Nat Neurosci 6: 593–599, 2003.[CrossRef][Web of Science][Medline]

Rosenberg JR, Amjad AM, Breeze P, Brillinger DR, and Halliday DM. The Fourier approach to the identification of functional coupling between neuronal spike trains. Prog Biophys Mol Biol 53: 1–31, 1989.[CrossRef][Web of Science][Medline]

Rouiller EM, Simm GM, Villa AE, de Ribaupierre Y, and de Ribaupierre F. Auditory corticocortical interconnections in the cat: evidence for parallel and hierarchical arrangement of the auditory cortical areas. Exp Brain Res 86: 483–505, 1991.[Web of Science][Medline]

Scannell JW, Blakemore C, and Young MP. Analysis of connectivity in the cat cerebral cortex. J Neurosci 15: 1463–1483, 1995.[Abstract]

Schmidt KE, Kim D-S, Singer W, Bonhoeffer T, and Löwel S. Functional specificity of long-range intrinsic and interhemispheric connections. J Neurosci 17: 5480–5492, 1997.[Abstract/Free Full Text]

Schreiner CE. Order and disorder in auditory cortical maps. Curr Opin Neurobiol 5: 489–96, 1995.[CrossRef][Web of Science][Medline]

Schreiner CE, Read HL, and Sutter ML. Modular organization of frequency integration in primary auditory cortex. Annu Rev Neurosci 23: 501–529, 2000.[CrossRef][Web of Science][Medline]

Singer W and Gray CM. Visual feature integration and the temporal correlation hypothesis. Annu Rev Neurosci 18: 555–586, 1995.[CrossRef][Web of Science][Medline]

Tomita M and Eggermont JJ. Cross-correlation and joint spectro-temporal receptive field properties in auditory cortex. J Neurophysiol 93: 378–392, 2005.[Abstract/Free Full Text]

Ts'o DY, Gilbert CD, and Wiesel TN. Relationships between horizontal interactions and functional architecture in cat striate cortex as revealed by cross-correlation analysis. J Neurosci 6: 1160–1170, 1986.[Abstract]

Tsodyks M, Kenet T, Grinvald A, and Arieli A. Linking spontaneous activity of single cortical neurons and the underlying functional architecture. Science 286: 1943–1946, 1999.[Abstract/Free Full Text]

Usrey WM and Reid RC. Synchronous activity in the visual system. Annu Rev Physiol 61: 435–456, 1999.[CrossRef][Web of Science][Medline]

Valentine PA and Eggermont JJ. Stimulus dependence of spectro-temporal receptive fields in cat primary auditory cortex. Hear Res 196: 119–133, 2004.[CrossRef][Web of Science][Medline]

Wallace MN, Kitzes LM, and Jones EG. Intrinsic inter- and intralaminar connections and their relationship to the tonotopic map in cat primary auditory cortex. Exp Brain Res 86: 527–544, 1991.[Web of Science][Medline]

Winer JA. The functional architecture of the medial geniculate body and the primary auditory cortex. In: Springer Handbook of Auditory Research. The Mammalian Auditory Pathway: Neuroanatomy, edited by Webster DB, Popper AN, and Fay RR. New York: Springer-Verlag, 1992, vol. 1, p. 222–409.

This article has been cited by other articles:

				M. Pienkowski, G. Shaw, and J. J. Eggermont Wiener-Volterra Characterization of Neurons in Primary Auditory Cortex Using Poisson-Distributed Impulse Train Inputs J Neurophysiol, June 1, 2009; 101(6): 3031 - 3041. [Abstract] [Full Text] [PDF]

				B. Gourevitch, A. Norena, G. Shaw, and J. J. Eggermont Spectrotemporal Receptive Fields in Anesthetized Cat Primary Auditory Cortex Are Context Dependent Cereb Cortex, June 1, 2009; 19(6): 1448 - 1461. [Abstract] [Full Text] [PDF]

				A. Ghoshal, P. Pouget, M. Popescu, and F. Ebner Early Bilateral Sensory Deprivation Blocks the Development of Coincident Discharge in Rat Barrel Cortex J. Neurosci., February 25, 2009; 29(8): 2384 - 2392. [Abstract] [Full Text] [PDF]

				A. J. Norena, B. Gourevitch, M. Pienkowski, G. Shaw, and J. J. Eggermont Increasing Spectrotemporal Sound Density Reveals an Octave-Based Organization in Cat Primary Auditory Cortex J. Neurosci., September 3, 2008; 28(36): 8885 - 8896. [Abstract] [Full Text] [PDF]

				J. L. Reed, P. Pouget, H.-X. Qi, Z. Zhou, M. R. Bernard, M. J. Burish, J. Haitas, A. B. Bonds, and J. H. Kaas Widespread spatial integration in primary somatosensory cortex PNAS, July 22, 2008; 105(29): 10233 - 10237. [Abstract] [Full Text] [PDF]

				I. Ozden, H. M. Lee, M. R. Sullivan, and S. S.-H. Wang Identification and Clustering of Event Patterns From In Vivo Multiphoton Optical Recordings of Neuronal Ensembles J Neurophysiol, July 1, 2008; 100(1): 495 - 503. [Abstract] [Full Text] [PDF]

				B. Gourevitch and J. J. Eggermont Evaluating Information Transfer Between Auditory Cortical Neurons J Neurophysiol, March 1, 2007; 97(3): 2533 - 2543. [Abstract] [Full Text] [PDF]

This Article Abstract Full Text (PDF) All Versions of this Article: 96/2/746    most recent 00059.2006v1 Alert me when this article is cited Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via ISI Web of Science (10) Citing Articles via Google Scholar Google Scholar Articles by Eggermont, J. J. Search for Related Content PubMed PubMed Citation Articles by Eggermont, J. J.