Journal of Neurophysiology

Error message

Notice: PHP Error: Undefined index: custom_texts in highwire_highwire_corrections_content_type_render() (line 33 of /opt/sites/jnl-jn/drupal-highwire/releases/20151124215058/modules/highwire/plugins/content_types/

Tuning to Sound Frequency in Auditory Field Potentials

Christoph Kayser, Christopher I. Petkov, Nikos K. Logothetis


Neurons in auditory cortex are selective for the frequency content of acoustical stimuli. Classically, this response selectivity is studied at the single-neuron level. However, current research often employs functional imaging techniques to investigate the organization of auditory cortex. The signals underlying the imaging data arise from neural mass action and reflect the properties of populations of neurons. For example, the signal used for functional magnetic resonance imaging (fMRI-BOLD) was shown to correlate with the oscillatory activity quantified by local field potentials (LFPs). This raises the questions of how the frequency selectivity in neuronal population signals compares with the tuning of spiking responses. To address this, we quantified tuning properties of auditory-evoked potentials (AEP), different frequency bands of the LFP, analog multi-unit (AMUA), and spike-sorted single- and multiunit activity in auditory cortex. The AMUA showed a close correspondence in frequency tuning to the spike-sorted activity. In contrast, for the LFP, we found a clear dissociation of high- and low-frequency bands: there was a gradual increase of tuning-curve similarity, tuning specificity, and information about the stimulus with increasing LFP frequency. Although properties of the high-frequency LFP matched those of spiking activity, the lower-frequency bands differed considerably as did the AEP. These results demonstrate that electrophysiological population responses exhibit varying degrees of frequency tuning and suggest that those functional imaging methods that are related to high-frequency oscillatory activity should well reflect the neuronal processing of sound frequency.


The responses of many neurons in early sensory cortices change systematically when a particular feature of their sensory input is manipulated. Neurons in primary visual cortex, for example, are selective to the orientation or spatial frequency of a visual pattern (Hubel and Wiesel 1962), whereas neurons in primary auditory cortex are modulated by a sound's frequency spectrum (Merzenich and Brugge 1973). Studying individual neurons’ feature selectivity has revealed a spatial organization in which neighboring neurons have similar preferences and collectively form topographic representations of the respective feature; e.g., a tonotopic map of the cochlear partition in the case of the auditory system (Kaas et al. 1999; Rauschecker et al. 1997). This spatial organization aids the study of this feature selectivity using techniques with lower spatial resolution—such as noninvasive functional imaging methods. With regard to the auditory system, several studies demonstrated that functional imaging can reproduce sound frequency maps in auditory cortex (Formisano et al. 2003; Petkov et al. 2006; Talavage et al. 2004; Wessinger et al. 1997).

In trying to integrate information from complementary methodologies, a basic but important question is the similarity between feature maps obtained. The signals underlying functional imaging arise from populations of neurons and are a result of neural mass action (Freeman 1975; Logothetis and Wandell 2004). Recent work suggests that the functional magnetic resonance imaging blood-oxygen-level-dependent (fMRI-BOLD) signal reflects the spatially summed somato-dendritic potentials and correlates well with electrophysiological population signals such as the local field potential (LFP) (Lauritzen 2001; Logothetis et al. 2001): especially the high-frequency oscillatory activity characterized by the LFP seems to correspond to what is measured by the BOLD signal. To the extent that particular LFP frequency bands may reflect distinct aspects of local processing, different imaging methods may also reveal diverse aspects of information processing. The driving, stimulus-related properties of the LFPs, including their response tuning, thus require careful examination.

Concerning the auditory system, electrophysiological population responses are only little investigated with regard to sound-frequency tuning, quite in contrast to the large literature studying this property at the single-neuron level. The few existing studies either quantified frequency tuning exclusively in the evoked response (Ohl et al. 2000) or in the lowest frequencies of the LFP (Norena and Eggermont 2002) or did not systematically quantify tuning properties (Brosch et al. 2002). As a result we know relatively little about the relationship of sound frequency tuning at the level of individual neurons and at the level of local population responses. This is in sharp contrast to research in vision, where several systematic studies demonstrated that evoked and local field potentials in V1 and MT exhibit tuning to the prominent features represented in these areas (Bonds 1982; Kayser and Konig 2004; Liu and Newsome 2006; Siegel and Konig 2003). Interestingly, these studies reported somewhat differing results: whereas the V1 studies suggest that feature tuning can be prominent both in lower- and higher-frequency bands of LFPs, the MT study reported a dissociation of lower and higher LFP frequency bands, with only the latter having comparable tuning to the spiking activity.

The goal of the present study was to obtain a systematic account of sound frequency tuning in neuronal spiking activity as well as in local population responses obtained from the same recording sites. Especially, we wondered whether tuning relationship between neural spiking activity and the LFP follows a similar pattern as observed in one of the visual areas.

We recorded from the auditory cortex from two adult rhesus monkeys that were passively listening to the acoustical stimuli in a sound attenuated booth (Illbruck acoustic GmbH). During recording, the animals were awake and not sleeping (as assessed using an infra-red camera and investigation of low-frequency oscillations in the LFP). All procedures were approved by the local authorities (Regierungspräsidium) and were in full compliance with the guidelines of the European Community (EUVD 86/609/EEC). Signals were recorded using commercial microelectrodes (0.8–1.2 MΩ impedance), high-pass filtered >4 Hz and low-pass filtered <10 kHz. Sounds (65 dB SPL) were delivered from two free-field speakers (JBL Professional, Northridge, CA, 70 cm from the ear, 50° to left and right), which have been calibrated to ensure a linear transfer function (condenser microphone, Brüel and Kjær GmbH, Bremen, Germany). Stimuli consisted of band-passed noise (7 center frequencies from 177 to 11,314 Hz, 1-octave steps) and pure tones (15 tones, 125 Hz to 16 kHz, half-octave steps). Both were presented as sequences of eight repeats (50-ms sound duration, 80-ms inter-sound spacing).

From the recorded signals, the following measures of neuronal activity were extracted. The auditory-evoked potential (AEP) was obtained by low-pass filtering the raw data at 150 Hz (3rd-order Butterworth filter) and computing the area under this slow wave during a temporal window of interest. The LFP was obtained by convolving individual trials with Gabor wavelets of different center frequencies (5, 10, 20, 40, 60, 80, 120 Hz and a bandwidth of 0.83). The amplitude of this convolution yields an estimate of the signal power in each frequency band. The analog multiunit activity (AMUA) was obtained by band-pass filtering the raw signal between 500 and 3,000 Hz, taking the absolute value and subsequent low-pass filtering at 300 Hz. To obtain units that are comparable across recording sites, these signals were normalized to units of SDs from baseline (Logothetis et al. 2001). Spike-sorted activity (S/MUA) was extracted using commercial spike-sorting software (Plexon) after high-pass filtering the raw signal at 500 Hz. To contrast the spike-sorted activity versus the other population measures of activity, we grouped single- and multiunit sites into one category, which we abbreviate S/MUA. For the present study, we only analyzed recording sites that responded significantly (P < 0.05) to the stimulus, both in the AEP and the spiking activity: the response amplitude during baseline (−200 to −20 ms before stimulus onset) was compared with the amplitude after stimulus onset (20–200 ms) using a paired t-test. Of a total of 258 sites, 119 were determined as responsive. Based on stereotaxic coordinates, frequency maps constructed for each animal as well as responsiveness for tone versus band-passed stimuli, most of our recording sites were in primary auditory cortex (fields A1 and R, 71 sites) and in the caudal belt (fields CM and CL, 48 sites).

Figure 1A shows examples responses for the different measures of activity, and the right panel displays tuning curves for each measure obtained during a window during the initial transient response (20–200 ms after stimulus onset). This recording site preferred frequencies between 1 and 2 kHz, which was identified as the best frequency (BF) by each activity measure. To quantify the similarity of these tuning curves, we computed the correlation of the S/MUA tuning curve with the tuning curve obtained from each of the other measures. The results (AEP: 0.39, LFP 20, 60 and120 Hz: 0.82, 0.74 and 0.93; AMUA: 0.99) suggest that each activity measure shows tuning to sound frequency but with varying similarity to the tuning curve obtained from the spike-sorted activity.

FIG. 1.

Example data from 1 recording site. Left: response time course for each activity measure analyzed (AEP, auditory evoked potential; LFP, local-field potentials; AMUA, analog multi-unit activity; S/MUA, spike sorted single- or multi-unit activity). Shown is the mean response across all sound frequencies. For the LFP, each gray level displays a different LFP frequency band; otherwise black lines indicate the mean and gray lines indicate the 99% confidence interval (bootstrap estimate). The S/MUA has units of spike/s, the other measures have units of z scores from baseline. The gray box indicates the time window, for which the tuning curves are shown on the right. Right: mean ± SE of the response for each sound frequency band of the broad-band noise stimulus. For the S/MUA, the dashed line indicates the baseline response before stimulus presentation (for the other measures the baseline is 0).

The population data (Fig. 2) confirm that this example was typical. Figure 2A shows the BF difference from each activity measure to BF obtained from the S/MUA. The average difference for the AEP was 2.0 octaves, which was significantly larger than for the AMUA (0.83 octaves, P < 10−4). For the LFP, there was a systematic trend: the BF- difference decreased with increasing LFP frequency. To quantify this dissociation between low and high LFP frequencies, we directly compared these (paired t-test, 5–40 vs. 60–120 Hz). The significant difference (P < 10−4) demonstrates that the high-frequency LFP better represents the tuning properties of the spiking activity than does the low-frequency LFP.

FIG. 2.

Population results. A: difference of the best frequency (BF) obtained from each activity measure to the BF obtained from the S/MUA. B: correlation coefficient between the tuning curves obtained from S/MUA and the other activity measures. C: mutual information between response and stimulus identity. In each panel, circles indicate the mean value (across all sites) and vertical bars indicate its 99% confidence interval (bootstrap estimate). In addition, small gray circles and boxes indicate the mean values separately for recording sites in the core and belt regions. Stars on the bottom indicate the significance of a sign test between the respective measure and the AMUA or S/MUA (as indicated by the arrow): * P < 0.05, ** P < 0.01, *** at least P < 0.001, all Bonferoni corrected for multiple comparisons.

Complementing the BF analysis, we computed the similarity (correlation) between the tuning curves obtained from the different activity measures (Fig. 2B). For the AEP, the correlation with the S/MUA tuning curve was significantly lower than for the AMUA (P < 10−7). For the LFP, there was a systematic trend of increasing correlation with increasing LFP frequency again resulting in a significant difference between the high- and low-frequency LFP (P < 10−8). The magnitude of this correlation depends on the number of measured points and the number of averages contributing to a smooth estimate of the curve. To deal with these limitations, we also compared tuning curves after fitting each individual curve with a Gaussian model (see Liu and Newsome 2006 for a similar approach). As expected, the resulting correlations were indeed higher (mean correlation across all activity measures 0.57 compared with 0.39). But again, the low-frequency LFP showed a significantly lower correlation with the spiking activity compared with the high-frequency LFP (P < 0.01).

Finally, we computed the mutual information between response and stimulus identity (Borst and Theunissen 1999; Panzeri and Treves 1996) (Fig. 2C). With seven stimuli, the maximal possible information is 2.8 bit. However, as most tuning curves have two flanks (cf. Fig. 1), the number of actually distinguishable frequencies and the mutual information are usually smaller. Both measures of spiking activity provided similar information (S/MUA: 0.57 bit, AMUA: 0.55 bit, P > 0.05), which was significantly higher than the information in the AEP (0.47 bit, P < 10−3). In addition, the information conveyed by the LFP was significantly lower compared with the spiking activity (at least P < 0.05, see Fig. 2C), and again the low and high LFP components differed significantly (P < 10−6).

As the preceding analysis focused only on the initial transient response, we verified that our findings do not depend on this choice of temporal window: the same analysis was repeated using one window immediately following the first (+200 to +380 ms) and one window during the sustained response (+600 to +780 ms). Analysis of these confirmed all of the observed findings (Fig. 2D demonstrates this for the similarity of tuning curves).

In addition, we compared the results between the auditory core and belt. The correlation of tuning curves (paired t-test P = 0.2) and the BF difference to the spiking activity (P = 0.045) showed only a small difference between these areas (Fig. 2, A and B). Comparing the difference of information between the spiking activity and the other measures, however, revealed a significant difference: in the belt, population responses conveyed more information about the stimulus identity (P < 0.01). The latter result might be due to our choice of stimulus, band-passed noise, as it is known that neurons in the belt respond to this stimulus better than neurons in the core.

In contrast to previous studies (Brosch et al. 2002; Norena and Eggermont 2002; Ohl et al. 2000), the present work provides a systematic comparison of sound frequency tuning in several measures of neuronal activity. In contrast to what might be expected from these reports, our results demonstrate that low-frequency oscillations and evoked potentials reflect the tuning properties of the neuronal spiking activity to a lesser degree than the high-frequency LFP. Although this result contrasts somewhat previous findings in the primary visual cortex (Bonds 1982; Kayser and Konig 2004; Siegel and Konig 2003), they fit well with findings from visual area MT (Liu and Newsome 2006), suggesting that the higher frequencies of the LFP in general have a closer relationship to neuronal spiking activity. It should be noted that our results are not a trivial consequence of the higher LFP bands directly reflecting the spiking activity on the same electrode. At least for frequencies ≤120 Hz, the LFP can be stimulus selective even in the absence of clear spiking activity (see supplemental figure1

There are several reasons why the tuning of the low-frequency LFP might be less specific. Previous studies suggested that the cortical tissue might act as a capacitative filter, which allows lower frequencies to travel long distances while attenuating high frequencies (Bedard et al. 2006; Ranck 1963). This would “blur” the low-frequency signals and reduce their selectivity (Liu and Newsome 2006). However, there is good evidence that this is not the case; in fact, the impedance spectrum of cortex is flat (Logothetis, unpublished data). As an alternative, it could be that the neuronal sources generating low- and high-frequency rhythms themselves have a different size. Although high-frequency signals are likely to originate from localized neuronal clusters (below the scale of a cortical column), low-frequency oscillations might be driven by divergent ascending input from the thalamus and brain stem (including neuromodulatory projections). Such difference in the generators is a likely explanation for both, the varying spatial correlation of low- and high-frequency LFPs (Eckhorn et al. 1988) and their differences in sound frequency tuning.

In addition, previous studies that carefully quantified the tuning properties of neurons in auditory cortex found a considerable heterogeneity even between neurons recorded on the same electrode (Recanzone et al. 2000): although the center frequencies closely matched, other properties like response thresholds and tuning width differed. Pooling neurons with such varying properties is likely to result in population tuning curves that do not perfectly correlate with those of single neurons. In combination with the varying extent of spatial pooling, this likely explains the pattern of correlations between different response measures observed in the present study.

Our findings have several implications for auditory activation patterns acquired in functional imaging studies. Work on the neuronal basis of fMRI-BOLD imaging demonstrated a good match between the LFP and the imaging signal (Logothetis and Wandell 2004). In particular, high-frequency oscillatory activity, i.e., frequency bands between 40 and 130 Hz (Logothetis et al. 2001) and 50–90 Hz (Niessing et al. 2005), showed a good correspondence to the BOLD signal. Interestingly, in the present study these high frequencies of the LFP showed stronger sound frequency tuning than the low-frequency components. Together with the finding that the BOLD signal also correlates with spiking activity (although sometimes to a weaker degree than the LFP), these results suggest that frequency maps of auditory cortex obtained using fMRI should closely correspond to the sound selectivity of the underlying neuronal populations (Formisano et al. 2003; Petkov et al. 2006; Talavage et al. 2004; Wessinger et al. 1997).


This work was supported by the Max-Planck Society, the German Research Foundation, and the Alexander von Humboldt Foundation.


  • 1 The online version of this article contains supplemental data.).

  • 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.


View Abstract