Synchronized neuronal firing in cortex has been implicated in feature binding, attentional selection, and other cognitive processes. This study addressed the question whether different cortical fields are distinct by rules according to which neurons engage in synchronous firing. To this end, we simultaneously recorded the multiunit firing at several sites within the primary and the caudomedial auditory cortical field of anesthetized macaque monkeys, determined their responses to pure tones, and calculated the cross-correlation function of the spontaneous firing of pairs of units. In the primary field, the likelihood of synchronous firing of pairs of units increased with the similarity of their frequency tuning and their response latencies. In the caudomedial field, by contrast, the likelihood of synchronization was highest when pairs of units had an octave and other harmonic relationships and when units had different response latencies. The differences in synchrony of the two fields were not paralleled by differences in distributions of best frequency, bandwidth of tuning curves, and response latency. Our findings suggest that neuronal synchrony in different cortical fields may underlie the establishment of specific relationships between the sound features that are represented by the firing of the neurons and which follow the Gestalt laws of similarity in the primary field and good continuation in the caudomedial field.
Correlated or synchronous firing of cortical neurons on different time scales has been hypothesized to play an important role in sensory processing and to contribute to feature binding (Eckhorn, 1999; Singer & Gray, 1995; von der Malsburg, 1983), attentional selection (Roy, Steinmetz, Hsiao, Johnson, & Niebur, 2007; Womelsdorf & Fries, 2007), sensory–motor coordination (Roelfsema, Engel, König, & Singer, 1997), and plasticity of feature maps (Kilgard, Vazquez, Engineer, & Pandya, 2007; Valentine, Teskey, & Eggermont, 2004; Bao, Chan, & Merzenich, 2001). Synchronous firing has also been proposed to provide a general brain mechanism for neuronal communication that is flexibly adapted to current cognitive demands (Fries, 2005).
The degree to which cortical neurons synchronize has been found to be related to the similarity of their receptive field properties (Eggermont, 1992, 1994, 2007; Brosch, Budinger, & Scheich, 2002; Brosch & Schreiner, 1999; Brosch, Bauer, & Eckhorn, 1995; Nelson, Salin, Munk, Arzi, & Bullier, 1992; Engel, König, Gray, & Singer, 1990). For example, in auditory cortex, the height of the cross-correlation function between pairs of neurons increases and the width narrows with increasing receptive field similarity. This included best frequency (BF), threshold as well as binaural and temporal response properties.
In auditory cortex, correlated firing has mostly been studied within the primary auditory field AI. We wondered, however, whether different auditory cortical fields may differ in the degree of correlated firing. This is potentially interesting because it is still an open question why there are multiple cortical fields (Barlow, 1986). The prevailing view is that cortical fields differ in the feature sensitivity of their neurons (Lomber & McMillan, 2010; Qi, Preuss, & Kaas, 2007; Rauschecker & Tian, 2000; Hubel & Wiesel, 1977). For auditory cortex, however, the degree to which fields vary in their feature sensitivities has not yet been resolved and depends on the relative hierarchical level within auditory cortex. In cats, distributions of receptive field properties of neurons in three auditory cortical fields were similar and largely overlapping (Eggermont, 1998). Likewise, in primates, several groups have found largely similar receptive field and response properties in auditory core and belt fields (Oshurkova, Scheich, & Brosch, 2008; Kajikawa, de La Mothe, Blumell, & Hackett, 2005; Recanzone, Guard, & Phan, 2000; Bieser & Müller-Preuss, 1996). By contrast, other groups have reported diminished selectivity for pure tones and an emerging preference for stimuli of broader spectral bandwidth between core and belt areas (Kusmierek & Rauschecker, 2009; Recanzone et al., 2000; Rauschecker, Tian, & Hauser, 1995) as well as different tuning ranges in different fields (Camalier, D'Angelo, Sterbing-D'Angelo, de la Mothe, & Hackett, 2012; Scott, Malone, & Semple, 2011; Bendor & Wang, 2008). Response strength and feature sensitivity of neurons change to some degree when animals are engaged in behavioral tasks (Fritz, David, & Shamma, 2012; Scheich & Brosch, 2012), compatible with the cognitive demand view (Fries, 2005).
We therefore hypothesized that different auditory cortical fields may also differ with respect to which neurons with specific feature sensitivities can engage in correlated firing. In more general terms, we tested the hypothesis that different auditory cortical fields carry out different cognitive operations on the stimulus features that are represented by the firing of neurons and that these operations reflect different grouping principles that have been identified by Gestalt psychologists (Bregman, 1990). Several studies suggest that synchronization patterns may indeed differ between fields. In cats, stimulus-driven but not spontaneous correlation was slightly weaker in AI than in the anterior and in the secondary auditory field (Eggermont, 2000), and the peak correlation coefficient decreased more slowly for unit pairs in AI than in the posterior auditory field (Eggermont, 2006). In rats, correlation of discharges was higher for pairs of neurons in AI than in the surrounding belt (Bao et al., 2001). After repetitively pairing a tone with electrical microstimulation, the correlation strengths of neurons increased strongly within the ventroposterior field only but not in AI.
Here, we studied correlated firing in two auditory cortical fields that are at different hierarchical levels of the auditory cortex but that have largely overlapping receptive field properties, namely AI and the caudomedial auditory field CM. We analyzed the relationships of response and receptive field properties for pairs of units whose firing was correlated. These correlations were measured during epochs of “spontaneous” firing, that is, without auditory stimulation because, in this way, the “true” interactions between neurons beyond an increased and coordinated firing caused by common sensory stimulation could be assessed; thus, the cross-correlation method was used as a quasi-anatomical method to reveal connectivity (Ts'o, Gilbert, & Wiesel, 1986). We did not aim to distinguish the origin of correlated firing, that is, whether it is caused by rate covariations because of, for example, nonstationarities or whether it reflects true spike synchronization (Staude, Rotter, & Grün, 2008). Our approach is similar to how global resting state networks are obtained by evaluating the coherence in spontaneous fluctuations of the BOLD fMRI signal (e.g., Hutchison & Everling, 2012). It only differs by the spatial and temporal scales that are under consideration. Our approach complements and extends results obtained by measuring local field potentials (Lakatos et al., 2013; Fukushima, Saunders, Leopold, Mishkin, & Averbeck, 2012; Ohl, Scheich, & Freeman, 2000). For example, a recent study found that the spatio-temporal activity profiles of epidural electrocorticograms resemble the tonotopic arrangement of several auditory cortical fields on the supratemporal plane (Fukushima et al., 2012). Our study extends this type of research because neuronal firing also allows to investigate response properties other than the BF, which, because of the large number of neuronal elements whose signals are reflected in field potentials (Buzsáki, Anastassiou, & Koch, 2012), cannot be resolved with this population measure (e.g., Kayser, Petkov, & Logothetis, 2007).
Experiments were approved by the authority for animal care and ethics of the federal state of Saxony Anhalt (No. 43.2-42502/2-253 IfN) and conformed with the rules for animal experimentation of the European Communities Council Directive (86/609/EEC). Methods have been detailed elsewhere (Brosch et al., 2002). During the experiment, anesthesia of the seven long-tailed monkeys of either sex (Macaca fascicularis) was maintained by a mixture of ketamine and xylazine or ketamine and diazepam. Neuronal signals were registered from the upper layers of the primary auditory field AI and of the caudomedial field CM of the left hemisphere using a recording system in which seven microelectrodes were arranged in a row with an interelectrode spacing of 330 μm (Thomas Recording, Giessen, Germany). On each electrode, we accepted signal deflections as action potentials, if the amplitude of a signal deflection was more than three times above the background and if its duration was between 50 and 500 μs (DataWave, Chicago, IL). The cortical area from which such a “unit” was recorded was established by analyzing the spatial distribution of the BFs (see below; Figure 1B). AI showed a characteristic orderly low-to-high progression of BFs in the rostrocaudal direction. Its caudal-most high-frequency representation was bordered by the field CM, in which the BF progression was reversed (high to low) but less orderly (Hackett, 2002). In three animals, the areal delineation corresponded to anatomical subdivisions of the auditory cortex using cytoarchitectural, fiber-architectural, and chemoarchitectural characteristics of the auditory fields (Figure 1D). Sites of <330 μm from the estimated border were excluded from further analyses.
Acoustic stimuli were produced digitally (Tucker-Davis Technologies, Gainesville, FL) and presented free field from a loudspeaker in a sound-attenuated chamber (IAC, Niederkrüchten, Germany). Response properties of units were measured by presenting 40 pure tones of different frequencies (within a range of two to eight octaves, depending on the frequency tuning of the recording sites) 10 times each. These tones were presented at a rate of 1 per 1475 msec (1178 msec in one monkey). They had an intensity of 40- to 60-dB SPL (constant in each recording) and a duration of 100 msec, including 5-msec rise/fall time.
For each of the 40 tones, we counted the number of discharges of a unit that occurred during the presentation of the tones. A response to a tone was considered significant if the number of discharges was >4 SD above the spontaneous rate (corresponding to p < .001 under the assumption of Poisson-distributed interspike intervals). BF of a unit was defined as the frequency with the maximal response magnitude. BF ratio of two units was defined as the absolute value of the logarithm to the basis 2 of the quotient of the two BFs. For a finer estimate of BFs, we also plotted the response to each of the 40 tones on a high-resolution frequency scale (0.00513 octaves) and subsequently convolved the tuning curve with a Gaussian kernel (σ = 0.27 octaves; red line in Figure 1A). Bandwidth of a tuning curve was defined as the distance in octaves between the highest and lowest frequency that yielded a significant response. Overlap of tuning curves of two units was defined as the frequency range eliciting a response in both units divided by the frequency range eliciting a response in either unit. Response latency was measured on poststimulus time histograms computed for each of the 40 stimuli, with a bin size of 1 msec, by finding the first bin in which the number of spikes was >4 SD above the spontaneous rate.
The cross-correlation function of the spontaneous firing of pairs of units was calculated by counting the number of spike coincidences of a pair for various time shifts between the two spike trains. Bin size was set to 1 msec. For noise reduction, cross-correlograms were smoothed by adding a quarter of each of the two adjacent bin counts to half of the count at a given bin. Subsequently, the correlogram was normalized by dividing each of its bins by the square root of the product of the number of discharges in both spike trains. The spontaneous activity was obtained from 400 distinct periods of 500 msec before each tone was presented, corresponding to a 200-sec period in which neuronal firing was not affected by the stimulus. Deflections in correlograms were taken to indicate statistically significant (p < .01) neuronal synchrony when they were larger than those expected in the correlograms of two independent spike trains with Poisson-distributed interspike intervals. Correlograms with only one significant bin were rarely encountered in our sample, and most correlograms were centered on the origin of the correlograms. For further analyses, synchronous firing was described by the following characteristics of correlograms (Figure 1C): (1) The correlation strength was defined as the maximal correlation in the normalized correlogram. (2) The temporal displacement of the correlation peak from zero delay was measured from the center of gravity of the timing of significantly correlated spikes. (3) The correlation width was measured at the half height of the correlation peak.
Using a seven-electrode array, we recorded action potentials from multiunit clusters at each of 349 acoustically responsive sites in the primary auditory field AI and 360 sites in the caudomedial field CM of seven anesthetized monkeys. We determined a unit's frequency tuning by measuring its responses to 40 different tones and obtaining its BF (Figure 1A). The response properties of the units covered very similar ranges in AI and CM (Figure 2A–C). In our sample, neither BF (p = .91, t test) nor response latency (p = .15) were significantly different between the two fields, and only the bandwidth of the tuning curves was slightly narrower in CM than in AI (p < .01, t test).
Neuronal synchrony was examined by computing the cross-correlation between the spontaneous discharges of pairs of these units, 926 pairs in AI and 905 pairs in CM. Figure 1C shows a typical correlogram of two units simultaneously recorded in CM with significantly synchronized firing. This correlogram, like most others seen in our recordings from AI and CM, was symmetrically around the origin of the correlogram with a half-height width of 175 msec. This suggested that the neuronal firing of neurons in AI and CM was partially synchronized with a temporal precision of tens of milliseconds.
Correlation strengths were similar in the two fields (p = .95), with medians of 0.076 and 0.078 in AI and CM, respectively (Figure 2D). Correlograms had phase shifts around zero, with no difference between AI and CM (p = .53; Figure 2E), and half-height widths of correlograms were similarly distributed in the range of 6–386 msec in AI and CM (U test, p = .78; Figure 2F), with medians of 178 and 182 msec in the two fields. These characteristics of significant cross-correlograms were obtained from 272 pairs in AI (29.4%) and from 214 pairs in CM (23.7%). Thus, there were significantly more correlated pairs in AI than in CM (χ2 = 4.47, p < .05).
Despite similar response properties in AI and CM, neurons with correlated spontaneous firing in the two fields had different relationships between their response properties. In AI, the proportion of synchronized pairs among the 926 pairs varied significantly with the BF ratio of the units (Figure 3A). The statistical significance was determined with a permutation test, with p < .05. For this test, we were generated 10,000 surrogate data, which each had a random relationship between the BF ratio of a pair and its synchrony. They were generated from the original data by keeping the same values of BF ratios and the same overall percentage of synchronized pairs. Units in AI with the same BF most frequently fired synchronous spikes, and the prevalence of synchronized units, on average, decreased with their BF ratio.
For a more fine-grained analysis of this dependence, BFs were determined at a higher resolution from tuning curves by filtering them with a Gaussian kernel (σ = 0.27 octaves; Figure 1A). Subsequently, the fraction of synchronized neuronal pairs with a specific BF ratio was determined by counting the number of significantly correlated pairs with this BF ratio and another 20 pairs with most similar BF ratios (10 pairs with smaller and 10 with larger BF ratios) and dividing this number by 21. With this reanalysis, we found that superimposed on the global decrease of synchronization with BF ratio were several other “preferred” BF ratios, that is, ratios at which the proportion of synchronized pairs assumed a local maximum (Figure 3C), such as the BF ratio close to one octave.
In CM, by contrast, the BFs of synchronized units were typically not the same, and there was hardly any global decrease of synchronization with BF ratio (Figure 3B). Analysis of this relationship at a finer frequency resolution, as for AI, rather suggested that the probability of units to fire synchronously varied with the BF ratio with a periodicity of 1/12 of an octave, that is, at semitone intervals (Figure 3D). The most preferred BF ratio of synchronization was at the octave, that is, at the same BF ratio at which there was also a local maximum in AI.
The predominance of preferred BF ratios close to semitone ratios in CM was unlikely in random data (p < .005). This was revealed with another permutation test for which we also generated surrogate data with a random relationship between the BF ratio of a pair and its synchrony. For each of the 10,000 of surrogate data, we plotted BF ratio versus the fraction of synchronized pairs and counted the number of preferred BF ratios that were in the range of the 12 semitone intervals within an octave, that is, within <0.017 octave. In AI, by contrast, the number of preferred BF ratios close to semitone intervals was not significantly increased over that when there was a chance relationship between BF ratio and synchronization (p = .33), that is, there was no quasi-periodic function. The difference between AI and CM indicates that the relationship between synchrony and BF that is found in CM is not a methodological artifact.
The different importance of BF for correlated neuronal firing in AI and CM was also reflected by different dependencies on the overlap of tuning curves in the two cortical fields. Overlap of tuning curves of two units was defined as the frequency range eliciting a response in both units divided by the frequency range eliciting a response in either unit. In AI, the proportion of correlated units increased as the overlap in the tuning curves increased (Figure 4A, gray curve). In CM, by contrast, the probability of correlation varied only weakly with the overlap of the tuning curves for the units in each synchronized pair (black curve).
Differences between the two cortical fields were further evident in the response latencies. For each pair of units, we calculated the difference of the average latency of the first spike obtained in each unit in response to the 40 tones that were used to determine the frequency tuning and compared this with the synchronization of their spontaneous spikes. In AI, synchronous spontaneous firing was mostly found in units with the same response latency, and the proportion of synchronized units decreased as the difference in response latency increased (Figure 4B, gray curve). In CM, units were most often synchronized when their response latencies differed, the highest proportion of pairs having a latency difference of 15–35 msec (black curve).
This study describes two auditory cortical fields that did not differ by basic neuronal response properties but by which neurons spontaneously engaged in synchronized firing. In AI, neurons were the more likely to fire synchronously, and the more similar were their response properties. The likelihood of synchronization increased with similarity of BF, overlap of tuning curve, and similarity of response latency. In CM, synchronized neurons had other relationships of their response properties. The likelihood of synchronization was increased for multiple BF relationships, most notably at the octave, and for a latency difference of a few ten milliseconds. We speculate that the two cortical fields are also specialized with respect to which neurons can engage in synchronized firing. The different synchronization rules in the two fields may underlie the establishment of specific relationships between the sound features, which are represented by the firing of the neurons and which follow the Gestalt laws of similarity and good continuation.
In AI, synchrony was related to the similarity of response properties. This is consistent with the general picture of neuronal synchrony in cortex that has emerged over the last decades of research. In AI of cats and monkeys, correlation strength increased and correlation width decreased with increasing similarity of frequency, threshold as well as of binaural and temporal response properties (Brosch et al., 2002; Brosch & Schreiner, 1999; Eggermont, 1992). Our findings extend previous findings obtained with high-spatial resolution mapping of epidural field potentials showing that spatial profiles of spontaneous firing (Fukushima et al., 2012) corresponded to the tonotopic map within AI and other auditory cortical fields. Likewise, in the early visual cortex of the same species, synchronized firing was related to receptive field overlap and similarity of orientation sensitivity (Brosch et al., 1995; Nelson et al., 1992; Engel et al., 1990). Thus, if synchronized firing contributes to feature binding (Eckhorn, 1999; Singer & Gray, 1995), our findings corroborate that synchrony in AI can support feature binding according to the grouping principles of proximity and similarity in feature space (Bregman, 1990).
In CM, synchronized neurons had specific relationships of their response properties. The BFs of synchronized units were typically not the same, and there was no global decrease of synchronization with BF ratio; rather, there were a number of preferred frequency intervals, including the octave, yielding an increased probability of synchrony. This was also reflected in that the probability of synchronization varied only weakly with the overlap of the tuning curves for the units in each synchronized pair. In addition, CM units were most often synchronized when their response latencies differed, the highest proportion of pairs having a latency difference of a few 10 msec. As will be detailed below, we propose that a function of neuronal synchrony in CM may be to establish syntactic relationships between auditory stimuli rather than to establish relationships between stimuli according to their degree of similarity, as in AI. Thus, if synchronized firing contributes to feature binding (Eckhorn, 1999; Singer & Gray, 1995), this suggests that synchrony in CM can support feature binding according to the grouping principles of good form and good continuation (Bregman, 1990).
What could the results in CM, namely, that the probability of units to fire synchronously was highest for a BF ratio of one octave and varied with the BF ratio with a periodicity of 1/12 of an octave, imply for the cognitive operations that are carried out when animals hear sounds? We speculate that synchronous firing is selective for particular frequency relationships. Am important class of sounds that match this preference are harmonic tone complexes, in which each component has an octave relationship to at least one other component. As synchrony was observed during spontaneous neuronal activity, it is probably generated by common input to or reciprocal connections between neurons with specific BF relationships (Munk, Nowak, Nelson, & Bullier, 1995). Thus, there may be strong synchronization between neurons in CM when they are stimulated with harmonic tones whose components correspond to their BFs. This hypothesis needs to be tested in further experiments with specifically tailored tone complexes. In the current experiment, only pure tones were used, with the purpose of measuring tuning curves.
Although this hypothesis has an immediate biological bearing because many environmental sounds and animal vocalizations have a harmonic structure, the results from CM may also relate to mechanisms that could be useful in humans for appreciating music in general and appreciating the consonance of chords in particular. A chord is composed of two or more harmonic tone complexes. Thus, when all possible frequency relationships in a chord are considered, they can be broken down into those that arise from the frequency intervals within a single tone complex and those that arise from frequency components of different tone complexes. The former have fixed simple ratios of two integers, for example, 2/1, 3/2, 4/3, 5/3, and so forth, and only the latter are variable. Thus, when concurrent tone complexes have semitone intervals, the preferred neuronal coupling in CM may yield maximal neuronal synchronization in this field.
The effect of generating the largest possible number of synchronously firing neurons in CM when concurrent tones have octave or semitone intervals may be enhanced by taking into consideration that tones of a chord are rarely intonated exactly at the same time but with slight onset asynchronies of a few tens of milliseconds. This is a range in which concurrent frequency components are perceptually fused or segregated (Kubovy, 1981; Bregman & Pinker, 1978; Rasch, 1978). This is consistent with what we found in CM, namely, that units were also most often synchronized when their response latencies differed, the highest proportion of pairs having a latency difference of 15–35 msec. Thus, the synchronization of pairs of units with different latencies of their responses to acoustic stimuli is able to identify or compensate for stimulus onset asynchronies frequently present in chords. The association of the frequency and asynchrony mechanisms in CM could be used for selecting frequency relationships between different musical tones and would lead to the neglect of relationships within single musical tones.
In conclusion, we propose that different auditory cortical fields are functionally specialized with respect to which neurons can synchronize with each other. The different rules found in AI and CM suggest that synchronization may be used for the establishment of relationships between the components within and across concurrent tone complexes, that is, to cognitive aspects of Gestalt psychology. Such specializations may add to other specializations of cortical fields and may underlie task-specific processing of sound features in different cortical fields in awake animals and humans (Scheich & Brosch, 2012; Scheich & Ohl, 2010). Our findings that neuronal synchronization is systematically related to neuronal response properties and that different relationships are observed in different cortical fields further support and extend the hypothesis that synchronized firing has a cognitive role in cortex (Womelsdorf & Fries, 2007; Fries, 2005; Eckhorn, 1999; Singer & Gray, 1995).
This study was supported by intramural funding and the Deutsche Forschungsgemeinschaft (SFB TRR 31, SFB 779). The authors thank Drs. Jennifer Altman, Andrew Bell, Xiaoqin Wang, Jos Eggermont, and Peter Heil for valuable discussions about this research.
Reprint requests should be sent to Michael Brosch, Leibniz-Institut für Neurobiologie, Brenneckestraße 6, 39118 Magdeburg, Germany, or via e-mail: email@example.com.