| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Departments of 1Biomedical and 2Electrical Engineering, Vanderbilt University, Nashville, Tennessee
Submitted 26 May 2004; accepted in final form 27 July 2004
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
; Hebb 1949). The spatiotemporal relationships of spike trains provide a robust and flexible system of classification that provides enhancement through associations (Hebb 1949) and depends on contextual changes, as well as short- and long-term response history (Hayek 1952).
Milner (1974) and von der Malsburg (1981) developed this theory to provide a framework for many perceptual phenomena that remain unexplained. Their elaborations are termed correlation theory or temporal binding theory, where perceptually related features are linked through correlated firing among subpopulations of cells. The original basis of the cell assembly theory was that relationships were formed between cells basedon anatomical connections (Hayek 1952; Hebb 1949). However, acknowledging the dynamic and adaptive nature of the brain, Hayek (1952) proposed that the formation of cell assemblies might not require actual anatomical changes in synaptic connections. It instead could result from short-term enhancement of synaptic effectiveness (a form of plasticity) generated by changes in the temporal structure of spike trains (von der Malsburg 1981). Short-term plasticity with respect to spike train structure has been demonstrated in a variety of conditions (Dobrunz and Stevens 1999; Lisman 1997; Snider et al. 1998; Tsodyks and Markram 1997; Usrey et al. 1998; Varela et al. 1997).
The spike train structure that induces short-term synaptic plasticity has typically been broken down into 2 classes of patterns, bursts and oscillations. Bursts have been operationally defined from an extracellular perspective as 2 or more spikes with intervals 
8 ms (Bair et al. 1994; Cattaneo et al. 1981a, b; DeBusk et al. 1997; Mandl 1993). Intracellular studies have typically shown bursts to be intrinsic, resulting from Na+ currents (Brumberg et al. 2000; Franceschetti et al. 1995) or Ca2+ currents (Traub and Miles 1991) that lead to nonlinear changes in firing based on changes in the membrane potential (Agmon and Connors 1989; Gray and McCormick 1996; Traub and Miles 1991). Oscillations occur across longer intervals, reflecting periodic firing usually in the gamma range (>30 Hz), depending on the location in the brain and/or the behavioral state or sensory input (Traub et al. 1999). The biophysical basis and perceptual significance of oscillations remains unclear (for review see Traub et al. 1999).
Here we explore the relationships between spike train structure and the orientation-dependent synchronization of cell assemblies from microelectrode array recordings (Samonds et al. 2004) to explore the behavior of cortical oscillation and synchronization. Our results suggest that at least some component of gamma oscillation arises from extrinsic sources such as feedback, although a bursting refractory period is also a likely contributor to the oscillation. The strength of oscillation was only moderately correlated with the strength of synchronization; however, oscillation was associated with maintenance of synchronization. The overall strength of synchronization was more strongly related to synchrony of the onset transients (defined by relative response latencies). We propose that cortical synchrony originates from coherent spatiotemporal stimulus structure, whereas bursts and oscillations maintain and modulate synchrony by preserving its coordination across cell populations.
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Details of experimental protocol and procedures are described in detail elsewhere (Samonds et al. 2003, 2004). Experimental procedures were performed under the guidelines of the American Physiological Society and Vanderbilt University's Animal Care and Use Committee. Three cats were anesthetized with propofol and N2O and paralyzed with pancuronium bromide. Correct refraction was provided by a combination of contact and spectacle lenses, and correspondence of the areae centrales was carefully checked with a reversible ophthalmoscope to ensure registry of binocular stimulation. A 5 x 5 multielectrode array (Bionics, Salt Lake City, UT) was pneumatically inserted to a fixed depth of 0.6 mm (exact recording depth varies because of brain curvature). Twenty-two, 25, and 23 single-unit recordings were simultaneously obtained from area centralis of area 17 from the 3 cats.
The isolation and signal-to-noise ratio (SNR) of single-unit recordings from the arrays were very good because of the low electrode impedances (about 250 k
). The average peak excursion of an action potential was about ±100 µV, ranging from ±50 to 200 µV. Nearly all channels typically had peak noise less than ±40 µV and most were much lower than ±30 µV. The SNR ranged from 1.5 to 20 (in one case as high as 40), in agreement with previous reports on the Bionics multielectrode array capabilities in cat visual cortex (Nordhausen et al. 1996). No data described in the present article were derived from spike separation from multiunit recordings on a single channel. Bionics spike-sorting software, however, was used to remove noise and artifact. During spike-sorting analysis, we were especially careful to consider waveform shrinkage attributed to burst firing (Harris et al. 2000; Snider and Bonds 1998).
Stimulation
Drifting sinusoid gratings of 2-s duration (10° aperture on a background with an average mean luminance of 73 cd m–2) were presented (binocular stimulation) at a distance of 57 cm on a 21-in. Sony Trinitron monitor (Tokyo, Japan). A refresh rate of 120 Hz prevented any artifactual stimulus-locked oscillation or synchronization (Snider 1997). Four subgroups (or assemblies) of 6 cells each were selected from the recordings based on similar orientation preferences within a subgroup. This provided 24 single cells and 60 pairs of cells for quantitative analysis. There are more pairs of cells than single cells because each group of 6 cells produces 15 possible pairs. We tested each of the 4 assemblies across a 30° range around their preferred orientation at 2° increments. Contrast was fixed at 50%, spatial frequency was fixed at 0.5 c/°, and temporal frequency was fixed at 2 or 3 Hz. These parameters were appropriate to drive essentially all cells responsive to the grating orientation. We collected responses from an average of 516 ± 33 presentations of 2-s duration of each stimulus for autocorrelation, renewal density, and cross-correlation analysis. The relatively large number of samples was required for the original quantitative examination of the data set (Samonds et al. 2004) and is not necessarily required for the analysis described in the present article. Because the latency analysis required monitoring each individual response, we found 50 samples to be more manageable while still yielding statistically reliable results.
Auto- and cross-correlation
We use methods for quantifying "effective connectivity" (Aertsen et al. 1989; Snider et al. 1998) to produce autocorrelograms and cross-correlograms for single cells and pairs of cells, respectively. The joint poststimulus time histogram (JPSTH) is a 2-dimensional histogram of all joint occurrences between pairs of cells at all possible temporal shifts (for 1-ms bins). For autocorrelation we simply replace y with x for all previous and future descriptions and equations. From this raw or observed JPSTH, the cross-product matrix of the individual PSTHs is subtracted
 | (1) |
). The quantity Dx,y(t1, t2) is the 2-dimensional cross-covariance matrix. The cross-covariance matrix is then normalized for the firing rate using the SD of the predicted PSTH
 | (2) |
Features within autocorrelograms were defined by making quantitative measurements of distinct peaks and valleys: 1) the strength of bursting was quantified by the peak at <8 ms (Bair et al. 1994; Cattaneo et al. 1981a, b; DeBusk et al. 1997; Mandl 1993), 2) the strength of the bursting refractory period (or the first oscillation peak) was quantified by the difference between the valley immediately after the burst peak and the next peak (after the bursting peak), and 3) the strength of oscillation was quantified by the difference between the second valley and the second oscillation peak. The strength of synchrony was quantified in a similar manner by measuring the peak in the CCH (at or near a lag time of 0 ms).
We considered alternative methods of measuring gamma oscillation based on Fourier analysis or fitting Gabor functions to the autocorrelogram (König 1994), but both methods are confounded by the dip of the burst refractory period, which leads to false indications of oscillation or an overestimation of the strength of oscillation (Bair et al. 1994; Young et al. 1992). Measurements based on the first oscillation peak alone would also produce the same problem. Therefore cells were classified as oscillatory when the ACH contained a secondary oscillation peak that was at least twice the magnitude of the random fluctuations in the ACH (0.1–0.3%). Subsequent analysis described below shows that the secondary peak is only weakly related to the refractory period and bursting. Reliance on the second peak of the ACH to quantify oscillation thus minimizes the possible confound with the refractory period. In some cases, third and fourth oscillation peaks were present that were also above the noise level in the ACH. The frequency of oscillation in those cells with a second peak in the ACH meeting the criteria described above was quantified by calculating the discrete Fourier transform (0–500 Hz) of the ACH and noting the peak Fourier energy between 20 and 60 Hz.
Renewal density analysis
Renewal density analysis was used to determine the serial dependency of oscillation (Lebedev and Wise 2000; Mountcastle et al. 1969, 1990). We randomly shuffled the order of the intervals for each response trial and repeated the autocorrelation analysis described above to produce the renewal density histogram (RDH). Any reduction in the oscillation in the RDH signifies serial dependency on the intervals, suggesting an extrinsic source of oscillation (Lebedev and Wise 2000; Mountcastle et al. 1969, 1990). With an external source of oscillation, such as feedforward or feedback, the neuron can still fire spikes within the "quiet interval" of an oscillation. Shuffling the spike intervals will eliminate a portion of the oscillation intervals and in turn reduce the oscillation in the RDH. An intrinsic source of oscillation, such as membrane hyperpolarization, absolutely prevents the neuron from firing during the "quiet interval" of an oscillation. In this case the oscillation interval is preserved after shuffling the order of the spike intervals and a reduction of oscillation is not seen in the RDH. Although a reduction in the oscillation in the RDH suggests extrinsic involvement, it does not exclude contributions to the oscillation by intrinsic factors.
Temporal dynamics of the CCH
Two-dimensional CCHs were produced using sliding-window analysis (Castelo-Branco et al. 1998; Gray et al. 1992; Nase et al. 2003; Neuenschwander et al. 2002). Although simply examining Cx,y(t1, t2) before integrating across the diagonal would reveal the temporal dynamics of the synchrony at a 1-ms resolution, too much noise is present in the matrix for meaningful interpretation of the temporal dynamics of synchrony. Sliding-window cross-correlation analysis simply reflects sliding-window averaging of Cx,y(t1, t2) to reduce the noise. We use a window of 200 ms width at a bin resolution of 50 ms.
Measuring latency
When looking at individual or averaged PST histograms across time, within the first 100 ms cortical cells have a transient peak with a relatively high probability of firing or average firing rate (Ghose and Freeman 1992). In some cases, we found that the transient probability of firing could be as high as 0.8 for 5-ms bins (or an average firing rate 160 sps). The initial peak is followed by a relaxation, after which there can be a secondary peak representing the sustained response of both simple and complex cells. Simple cells have additional peaks in the firing rate caused by the temporal frequency of the drifting sinusoid grating.
To determine latency, we measured the first-spike time within a 100-ms window for single cells (see Fig. 1A) and the difference between first-spike times within a 100-ms window between pairs of cells. Data were included only for cells with responses within the window in 
75% of the trials. We measured first-spike times on only the first 50 data samples for each cell and pair of cells because direct observation of the individual responses was necessary to ensure that maintained activity (see Fig. 1A) was not mistaken for first spikes in response to visual stimulation. Maintained activity was usually apparent only in complex cells. In these cells the transient peak was generally strong, making the first spike (as a response to the stimulus and not maintained activity) easily identifiable. The difference between the average latency determined from the transient peak in the PSTH (using all 516 ± 33 samples) and determined from the first-spike time procedure described above (using the first 50 samples) was not statistically significant (t-test, P > 0.4).
|
|
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
). We first analyzed the single-cell spike trains using autocorrelation and renewal density analysis (seeMETHODS) to quantify the bursting, oscillation, and serial dependency of interval information. We examined the possible underlying mechanisms of the oscillation and bursting refractory period found in the ACH. After quantifying the temporal structure of the spike trains, we inspected the relationship between the structure and the cross-correlation between pairs of cells (i.e., synchrony). We also inspected the temporal dynamics of the synchrony. Finally, after comparing orientation-dependent changes of the structure and synchrony, we tested how latency differences (i.e., synchronization of the input) and orientation-dependent changes in latency were related to the overall strength of cortical synchronization. Autocorrelation and renewal density analysis
Figure 3, A, B, and C represent the 3 basic qualitative autocorrelation histogram (ACH) patterns (black lines) we find for the 24 cells we examined. The ACH has been normalized for the firing rate, as well as for temporal modulations in the firing rate (Aertsen et al. 1989; see METHODS for details). The same procedure was previously used to produce cross-correlation histograms (CCH) (Samonds et al. 2003; Snider et al. 1998). The ACH is plotted as the percentage of the maximum possible ACH value (i.e., if all spikes were correlated) with the aforementioned normalization procedure.
|
; Cattaneo et al. 1981a, b; DeBusk et al. 1997; Mandl 1993). In 5 out of 24 cells, the ACH showed little or no modulation beyond the bursting peak (e.g., Fig. 3A). The ACH was either relatively flat or had a slight decay. Note also that the 1-ms autocorrelation value (the value is actually below zero in all cases; not shown) represents the refractory period of the spiking mechanism of the cell, which limits its maximum firing rate.
The remaining 19 out of 24 ACHs showed a clear valley immediately after the bursting peak (Fig. 3, B and C), which corresponds to the bursting refractory period centered at about 8 ms found in interspike interval (ISI) histograms (Bair et al. 1994; DeBusk et al. 1997; Snider et al. 1998; Young et al. 1992). The term "refractory period " in this case is used to describe an absence of intervals over an intermediate range reflecting the short "silent" period following a burst. The average center of the refractory period was 10.5 ± 2.7 ms, ranging from 8 to 17 ms. Fifteen out of the 24 ACHs (or 63% of the 24 single cells) also had a clear secondary peak (Fig. 3C) signifying oscillatory firing, which fell within the gamma range (30–60 Hz).
We determined the frequency of the oscillation by measuring the peak between 20 and 60 Hz of the Fourier transform of the ACH (Fig. 3D). The secondary peak in Fig. 3C leads to a clear peak at 43 Hz in Fig. 3D. The average frequency for the 15 cells that exhibited gamma oscillation was 49.8 ± 4.5 Hz (range 43–54 Hz). The SD of the frequency of oscillation is significantly smaller (mean = 2.2 Hz) when we examine each 6-cell assembly independently (F-test, P < 0.06).
We also wanted to determine whether the oscillation we found in the ACH was an intrinsic property of the cell membrane (e.g., Gray and McCormick 1996; Nowak et al. 2003) or an extrinsic property resulting from, say, GABAergic interneuronal networks (Cunningham et al. 2003; Traub et al. 1999, 2003). We compared the renewal density histogram (RDH; gray line in Fig. 3C) to the ACH (the original autocorrelation histogram described above; black lines) to assess the influence of extrinsic factors on oscillation. If the oscillation in a spike train is extrinsically driven, the ACH will depend on the serial order of the ISIs (Lebedev and Wise 2000; Mountcastle et al. 1969, 1990; see METHODS for details). The RDH is calculated in the same manner as the ACH described above except the intervals for each response are randomly shuffled. A reduction in the periodicity of the ACH after random shuffling of the ISIs (i.e., the RDH) reflects contributions from external influences (Lebedev and Wise 2000; Mountcastle et al. 1969, 1990). This is because intrinsic (membrane-based) oscillation, such as hyperpolarization, will likely result in absolute silence between spike clusters, whereas with an external source of oscillation, spikes can still occur between the oscillation intervals. A reduction of periodicity in the RDH does not, however, rule out additional contributions from intrinsic properties.
We found no reduction in the RDH in only 3 out of 19 cells (of these 3 cases, 2 had gamma oscillation). In the other 16 RDHs, the periodicity was clearly reduced with both a decrease in the depth of the refractory period valley and decreases in the 2 oscillatory peaks. Overall, the ACHs tended to "flatten" out. Figure 1C shows an example of the reduction (gray lines) in the periodicity of the RDH. The reduction is also clear in the Fourier analysis with the 43-Hz peaks becoming less prominent (Fig. 3D).
The reduction in the RDH does not appear to arise simply because the responses contain both short-interval (<8 ms) and long-interval (>15 ms) structures. All 3 cells that showed no reduction in the RDH also demonstrated both bursts (short-interval structure) and a bursting refractory period (long-interval structure). Additionally, 2 of these cells exhibited gamma oscillation. We therefore believe that the reduction in the RDH observed in the other 16 cells represents influences beyond what might be expected from the combined presence of short- and long-interval structures and suggests some tangible external influence on the gamma oscillation. However, those 2 cells showing gamma oscillation but no RDH reduction demonstrate that intrinsic properties of the cells can also play a considerable role in producing the gamma oscillation.
To quantify the serial dependency of oscillation across our population of cells, we quantified the strength of the oscillation in the ACH by: 1) the first oscillation peak (difference between the burst refractory period valley and the first oscillation peak) and 2) the second oscillation peak (the difference between the second valley and the second oscillation peak). Figure 4 shows the average values we measure for the 2 oscillation peaks observed in the ACH (black) and the shuffled RDH (gray). The first and second oscillation peaks are reduced by 31 and 34%, respectively, from the ACH to the RDH, presumably reflecting significant contributions to the oscillation from extrinsic factors (t-test, P < 0.02 and P < 0.007, respectively).
|
The bursting refractory period can suggest the presence of a first oscillation peak in the ACH even though it may be absent, thus leading to a mistaken identification of gamma oscillation (Bair et al. 1994; Young et al. 1992). Conversely, measurements of the bursting refractory period can be confounded by the oscillation. For example, our measurement of the first oscillation peak in the ACH and RDH (above) might suggest the bursting refractory period is produced by extrinsic factors. However, most intracellular studies clearly find an intrinsic afterhyperpolarization (AHP) after bursts that would produce the refractory period (Agmon and Connors 1989; Brumberg et al. 2000; Connors and Gutnick 1990; Frenceschetti et al. 1995; Gray and McCormick 1996; Nowak et al. 2003; Silva-Barrat et al. 1992; Traub and Miles 1991).
If the bursting refractory period results from an AHP after a burst, the probability of firing within this refractory period nears zero. The cell does not necessarily have to fire a spike immediately after the AHP. There will be more instances of ISIs corresponding to the AHP (and the bursting refractory period) when the cell does fire immediately after the AHP, which would be more likely with higher average firing rates. Therefore if the bursting refractory period reflects an intrinsic property such as an AHP, there should be correlation between the strength of the bursting refractory period and the firing rate (and the highest possible firing rate would in effect be limited by this AHP and refractory period). Figure 5A demonstrates there is a relatively strong relationship between the refractory period measurement and the firing rate (R2 = 0.48), suggesting that an AHP is the source of the refractory period.
|
; Nowak et al. 2003). The results reported above suggest that oscillation appears to have at least some component of extrinsic input. By examining the relationship between bursts and oscillation, we hoped to determine the extent to which oscillation arises from a burst AHP (intrinsic) or from more coordinated feedback (extrinsic). In addition, we also examine the relationship between the refractory period and bursts to increase assurance that the refractory period does indeed arise from bursts. Figure 5B presents a scatter plot of the refractory period and oscillation measures against bursting. The correlation between bursts and the refractory period is moderate (R2 = 0.27), whereas there is almost no correlation between the bursts and oscillation (R2 = 0.06). We thus propose that there is a refractory period likely resulting from an AHP (Fig. 5A) that is a property of bursts (Fig. 5B). However, we do not see a strong link between oscillation and the presumably intrinsic bursting and AHP (Fig. 5B).
We emphasize that the extracellular recordings from our sample of cells do not reveal the underlying biophysical processes that generate temporal structure in single-cell cortical responses. We are merely attempting to quantify the bursting, bursting refractory period, and gamma oscillations in the context of the net output of the cell so that we can test their statistical relationships with the synchronization between groups of cells. Our analysis of extrinsic versus intrinsic sources is simply to provide some reference between our data and the intracellular results reported on these phenomena.
Temporal structure and synchronization
Eckhorn et al. (1988) and Gray et al. (1989) proposed that gamma oscillations drive synchronization across larger cortical distances (>2 mm) when the cells share similar orientation preferences and are coherently stimulated with a single bar of light. The phase locking of the oscillatory activity was present for shorter distances as well. The dimensions of our microelectrode array (5 x 5) yielded cell pairs as close as 400 µm (electrode spacing) and as far as 2 mm, with both overlapping and nonoverlapping receptive fields (Samonds et al. 2004). In this section, we examine the relationship between the temporal structure (with an emphasis on the gamma oscillation) and synchronization. We again use the normalized JPSTH analysis to produce a CCH. Again, correlation is quantified as the percentage of the maximum possible correlation.
Figure 6 is an example of 3 cells that are synchronized when driven by the same orientation. We show only 2 of the CCHs to provide examples of 2 cases of very strong synchronization that have different characteristics. Figure 6D is the CCH between one cell with very weak temporal structure (Fig. 6A; cell 8) and another cell with very strong gamma oscillation (Fig. 6B; cell 13). In spite of the lack of oscillation in one of the 2 cells, there is still very strong synchronization with a lag time of 0 ms, but no oscillation is apparent in the CCH (Fig. 6D). For these 2 cells, the receptive fields were partially overlapping (data not shown), suggesting some common input. In our second example, both cells (Fig. 6, B and C; cells 13 and 11, respectively) have strong gamma oscillation, which is then reflected in their CCH (also 0-ms lag time; Fig. 6E). The receptive fields of these 2 cells were also partially overlapping.
|
|
|
|
proposed that horizontal connections between orientation columns would lead to the orientation-dependent synchrony across longer distances. However, the phase difference of the synchronization is generally 0 ms, making it unlikely that synchronization relies on direct serial synaptic connections between the analyzed cell pairs, although the absence of a shift in the CCH is not an unambiguous indicator of synaptic pathways (Aertsen et al. 1989). Forty out of the 52 peaks in the CCHs were centered at 0 ms, implying no role for synaptic delay, and only 8 out of the 52 peaks differed by more than 2 ms (ranging from 3 to 4 ms), with an average lag time of the CCH peak of 0.8 ± 1.2 ms. Orientation-dependent changes in structure
Varying the orientation of a visual stimulus leads to changes in the number of spikes in a burst (burst length). For a constant firing rate, nonoptimal orientations reduce burst length (DeBusk et al. 1997). This variation in burst length can also be viewed as bursts being tuned for orientation with greater selectivity than that of the total spike count (Cattaneo et al. 1981a). Because longer bursts yield a higher probability of postsynaptic spiking (Snider et al. 1998), correlation is also more selective for orientation than the firing rate. Oscillation has been found to be tuned for orientation as well (Friedman-Hill et al. 2000; Frien and Eckhorn 2000; Frien et al. 2000).
Figure 10 is an example of the tuning of temporal structure for orientation. At first glance, the ACH features (Fig. 10, A and B) and CCH peaks (Fig. 10C) do not appear to be highly selective for orientation. However, we are examining only a 30° range of orientation and both the ACH and CCH are normalized for changes in the firing rate. Therefore any observed change with respect to orientation means that the feature is more selective for orientation than is the firing rate. Previous analysis on these data found the synchrony to have a half-height bandwidth on average 31% narrower than the average firing rate (Samonds et al. 2004). Figure 10 thus presents a representative example confirming that bursts, oscillation, and correlation are all more spatially selective than the average firing rate (Cattaneo et al. 1981a; DeBusk et al. 1997; Frien and Eckhorn 2000; Frien et al. 2000; Samonds and Bonds 2004; Samonds et al. 2003, 2004; Snider et al. 1998).
|
; Friedman-Hill et al. 2000), although it has been much more elusive in isolated single-unit studies (Bair et al. 1994; Bringuier et al. 1997; de Oliveira et al. 1997; Molotchnikoff et al. 1996; Nowak et al. 1999; Tovee and Rolls 1992; Young et al. 1992). In our experience, oscillation can be very difficult to detect in single units because it can be very selective for orientation. For example, in the cell of Fig. 10A, oscillation only occurred over a 6° range of orientation and was relatively weak. Testing of orientation at 2° increments thus increased our chances of observing small amounts of gamma oscillation in single units.
Figure 10 also shows that the correlation (synchrony) appears to be more selective to orientation than bursts. Snider et al. (1998) found that the correlation varied even beyond what changes in the burst length would predict and proposed that coincident inputs might also play a role in the synaptic efficacy (see also Samonds and Bonds 2004; Samonds et al. 2003). The narrow selectivity of oscillation (Fig. 10A) is an alternative explanation of the narrow selectivity of correlation, but the relatively moderate relationship between oscillation and synchrony (Fig. 9) suggests that other factors influence orientation-dependent changes in synchronization. In the next section, we explore how stimulus-induced timing relationships between responses among an assembly of cells are related to cortical synchronization.
Coincidence detection and orientation-dependent phase shifts
In a previous study (Samonds and Bonds 2004), we showed that orientation information is contained in the spike timing (in the form of temporal shifts in the response) and proposed that it resulted simply from the relationship between a visual stimulus and the spatiotemporal properties of the receptive field (particularly off-centered relationships; see Fig. 11, A and B). Because response latency is correlated between cells across trials (Fries et al. 2001), we speculated that stimulus-dependent relative time shifts would be reflected in the synchrony among cell assemblies and could be decoded by means of coincident detection. Figure 11, C–F (left vs. right) provides a clear example of an orientation-dependent timing shift in the response. When a 174° drifting grating is presented, the response latencies for the 2 cells are the same (Fig. 11C, white arrow). When the grating is rotated by 10° (to 184°), the response latencies now differ, with the response for cell 10 (gray) arriving about 50 ms later (Fig. 11D, black arrow). Figure 11, E and F confirm the relationship between latency and synchrony (Fig. 11, A and B) by showing that the synchrony is nearly halved when the latencies are out of phase. The example in Fig. 11, C–F represents simple cell responses, with very strong modulation that is coupled to the drift rate. However, information on orientation resulting from response latency shifts is also seen in complex cell responses, with little or no rate modulation (Samonds and Bonds 2004).
|
75% of the trials had a spike within a 100-ms window. Because responses characteristically have a strong onset transient peak in the PST (<100 ms) followed by another strong, but broader, peak (about 200–300 ms) at the start of the sustained response, we sometimes used the secondary peak (using a 200-ms window) if it was more reliable (i.e., meeting our 75% criteria). For relative latency, spikes from both cells had to be within the 100- or 200-ms window for 75% of the trials. The average transient response latency (n = 12 cells) was 53.5 ± 18.2 ms and the average sustained response latency (n = 18 cells) was 239.0 ± 88.1 ms. The average SD of either the transient or sustained latency (n = 18 cells; from n = 50 stimulus trials) was 29.5 ± 16.8 ms. The average difference between latencies (n = 31 cell pairs) was 42.3 ± 42.2 ms with an average SD (n = 31 cell pairs; from n = 50 stimulus trials) of 48.5 ± 24.4 ms. The difference in latency ranges from 0 to 120 ms, whereas the SD ranges from 11 to 91 ms.
Figure 12 has scatter plots showing the relationship between synchronization and either the difference between latency (
latency; Fig. 12A) or the SD of the difference between latency (SD latency; Fig. 12B). Both the difference in latency (R2 = 0.50) and the SD of the difference (R2 = 0.53) show strong inversely logarithmic correlation with the cross-correlation peak. This demonstrates that the chance of 2 cells synchronizing is much greater when their response latencies are nearly the same (and reliably so across stimulus trials). Because the synchronization is present from the first moments of the response, we conclude that it is induced by the simultaneity of the input signals and is not developed post hoc throughout the response.
|
75% of the trials biases our analysis to those pairs of cells that have stronger responses, where the response latency is likely to be more reliable. Figure 12B might be thus interpreted as simply demonstrating that the strength of synchrony depends on the strength of the response. However, we find almost no correlation between the strength of the synchrony and the average firing rate (Fig. 13).
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
). When driven by a drifting grating of a particular orientation, distinct groups of cells are strongly synchronized, increasing the chances that their aggregate signal integrates at later stages of visual processing. Even small changes in orientation reduce the synchronization, yielding in effect an assembly filter (Samonds et al. 2004). By controlling the effective integration at later stages of visual processing through synchronization, the cortical network limits or filters what collective information is processed. Beyond the creation of this cooperative information, the assembly filter also provides a dynamic mechanism to filter for structure across the visual field because cell assemblies contain both overlapping and nonoverlapping receptive fields. We considered 3 possible sources, each known to be orientation dependent, for this synchronization: 1) gamma oscillations (Eckhorn et al. 1988; Gray et al. 1989), 2) bursting (Snider et al. 1998), and 3) stimulus-dependent timing (Samonds and Bonds 2004).
Before we could measure the statistical relationships between these properties and synchrony, we needed to quantify their respective strengths. Our approach was to analyze properties (bursting, burst refractory period, oscillation) of the autocorrelation analysis (Aertsen et al. 1989) that were suggested by intracellular studies. We then measured the correlation between each property and the synchrony. We found only a moderate relationship between bursts or oscillation and synchronization, suggesting that they are only peripherally associated with synchronization. We found a stronger inverse logarithmic relationship between the cell pairs' response latency differences (as well as the reliability of those differences) with synchronization. This result suggests that those cell pairs that are synchronized must be synchronized for their first spikes and that synchrony is not gradually acquired. Interestingly, we do find a possible functional role for oscillation. In most CCHs that do not exhibit any oscillation, synchrony decays over intervals on the order of seconds. All of those CCHs that exhibit oscillation maintain their synchrony, suggesting that oscillation provides a "framework" that prevents the decay.
Supratheshold activation?
Eckhorn et al. (1988) and Gray et al. (1989) demonstrated that cortical cells with similar orientation preferences would synchronize when activated by a single stimulus that spanned across both receptive fields. They proposed that the synchronization supported the temporal binding theory, in which single cells representing individual object features synchronize when those cells are perceptually related (e.g., respond to part of the same object). They also suggested that intracortical oscillation provided a substrate for synchronization of cells across longer cortical distances. Theoretical models have supported the synchronization of oscillators (Ernst et al. 1995, 1998; Grossberg 1976, 1980; Rietboeck et al. 1987; Tsodyks et al. 1993; von der Malsburg and Schneider 1986).
Our results, however, suggest instead that synchrony is a transient phenomenon that occurs for particular groups of cells when they are coherently activated by stimulus structure. Although our results do not prove whether oscillation maintains synchrony or synchrony maintains oscillation, we suggest the transient nature of the synchrony implies the former. Synchronous inputs could trigger the oscillation by suprathreshold potentials but whether oscillation is sustained by feedback circuitry (Cunningham et al. 2003; Traub et al. 2003) or intrinsic sources (Gray and McCormick 1996) also remains unclear. Stimulus-dependent gamma oscillation is found in the membrane potential of hyperpolarized cells (Jagadeesh et al. 1992), but because hyperpolarization of one cell does not disable the network, that does not rule out a role for feedback in the oscillation. Bursting also requires a suprathreshold potential (Agmon and Connors 1989; Brumberg et al. 2000; Connors and Gutnick 1990; Frenceschetti et al. 1995; Gray and McCormick 1996; Nowak et al. 2003; Silva-Barrat et al. 1992; Traub and Miles 1991), suggesting that synchronous inputs encourage both bursts and oscillation. Bursts and GABAergic networks might even be viewed as coincident detectors (Galarreta and Hestrin 2001; Harris et al. 2001). This implies that bursts and oscillation do not actually directly amplify the stimulus information and produce synchrony, but in essence reflect and preserve the salient information from the synchronous inputs by making the relatively weak cortical synaptic connections more effective.
Cortical gamma oscillation has also been suggested as arising from afferent LGN input (Ghose and Freeman 1992, 1997). However, there appear to be 2 forms of oscillation: 1) 60- to 120-Hz oscillation that originates in the retina and is passed on to the LGN and cortex and 2) 30- to 60-Hz oscillation that originates in the cortex and propagates back to the LGN (Castelo-Branco et al. 1998; Neuenschwander et al. 2002). The 60- to 120-Hz feedforward oscillation was transient and decayed (along the same time frame as our nonoscillatory synchrony), whereas the 30- to 60-Hz intracortical oscillation was sustained.
The temporal binding theory
One of the more prominent challenges of cognition is the integration of features into gestalt entities. How does the brain know that the features belong to the same object? Conceptual problems arise when binding is considered as strictly intracortical (i.e., Eckhorn et al. 1988; Gray et al. 1989). Intracortical binding through oscillation essentially solves the problem using circular logic. The cortex binds object features to identify the object, but to bind the features the cortex needs to know in advance that they are part of the same object. This loop requires a homunculus to extract the advanced knowledge. From a mathematical perspective, communication channels cannot create information; information can only be preserved or lost (Shannon 1948). Therefore if we consider synchrony to be informative to the perceptual process, some representation of that information must also be available in the retina.
Given these constraints, we would suggest that cortical synchrony starts with the visual stimulus and the retinal input. Spatial and temporal correlations in the visual scene cause synchronous activation of populations of retinal cells. This leads to matching cortical latencies, triggering synchronization. Transient synchrony occurs with either dynamic or novel stimuli (Kruse and Eckhorn 1996) propagating from retina to LGN to cortex (Castelo-Branco et al. 1998; Neuenschwander et al. 2002). Kruse and Eckhorn (1996) proposed that transient synchrony and oscillatory synchronization were distinct, but we suggest instead that sustained steady-state stimulation creates oscillation that maintains the transient form of synchrony.
Oscillation occurs more often for drifting versus stationary gratings (Engel et al. 1990) and might therefore be a mechanism that is relevant only for motion processing. However, with stimulus flicker or alert monkeys (allowing residual eye movements), stationary gratings also induce gamma oscillation (Eckhorn et al. 1988; Friedman-Hill et al. 2000). When testing both stationary and drifting gratings, cells can oscillate for both forms of stimulation, only stationary stimuli, or only drifting stimuli (Friedman-Hill et al. 2000). These different populations might form subclasses that serve different types of analyses and project to different locations. If oscillation and synchrony constituted a mechanism to propagate information reliably, one might expect cells that oscillate only for stationary stimuli to project to object-recognition regions (e.g., V4, IT, etc.) and cells that oscillate only for drifting stimuli to project to motion-processing areas (e.g., V2, MT, etc.).
The successive layers of the pathway from retina through extrastriate cortex could, as nonlinear networks, progressively extract higher-order spatial and temporal properties of visual stimuli. At the level of the striate cortex, the correlations become progressively harder to detect as stimulus properties (e.g., Purpura et al. 1994) become more abstract and global. This might explain why it is typically more difficult to find both synchronization and the appropriate form of stimulation to identify the representation of visual information at higher levels of the visual system (Usrey and Reid 1999).
Therefore we prefer to think of synchrony as a mechanism for reliable signal transmission that extracts higher-order correlations as a gestalt rather than as an active binding mechanism that represents a secondary code to link a system of simple feature extraction. Eckhorn et al. (1988) proposed the hypothesis that global feature extraction (i.e., detection of complex structures) was simply a higher-order form of feature extraction that arises in cortex. We believe that "global properties" (extending beyond the boundaries of single receptive fields) might simply be embodied in the higher-order stimulus correlations revealed in the synchronous firing in the cortex. Higher-order feature extraction could be interpreted as the binding of lower-order local features, but the importance in the distinction between an active intracortical binding mechanism versus higher-order filtering is that the latter process will not run into the circular logic contradiction described above and might not pass every possible test of feature binding (e.g., Shadlen and Movshon 1999).
In conclusion, we have described a simple and straightforward system that allows the selective synchronization of particular subgroups of cells by the cortical network. The synchronized assemblies can extract salient visual information such as coherent visual structure. The scheme is supported by our data showing a strong relationship between the overall strength of cortical synchrony and the initial strength of the cortical synchrony (see also Fries et al. 2001). The transient nature of the synchronization supports both models (Opara and Wörgötter 1996) and psychophysical data (Hancock and Phillips 2004; Lee and Blake 1999; Usher and Donnelly 1998) that suggest a fast cortical response latency-dependent form of binding that does not rely on top-down mechanisms. Our results also suggest that gamma oscillations maintain the transient form of synchrony so that the brain can retain and use this information for higher cognitive (i.e., top-down) processing.
|
GRANTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
ACKNOWLEDGMENTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
FOOTNOTES |
|---|
Address for reprint requests and other correspondence: A. B. Bonds, Department of Electrical Engineering, Vanderbilt University, 255 Featheringill Hall, 400 24th Ave. South, Nashville, TN 37212 (E-mail: ab{at}vuse.vanderbilt.edu)
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Agmon A and Connors BW. Repetitive burst-firing neurons in the deep layers of mouse somatosensory cortex. Neurosci Lett 99: 137–141, 1989.[CrossRef][ISI][Medline]
Bair W, Koch C, Newsome W, and Britten K. Power spectrum analysis of bursting cells in area MT in the behaving monkey. J Neurosci 14: 2870–2892, 1994.
