Sensory stimuli under natural conditions often consist of a temporally irregular sequence of events, contrasting with the periodic sequences commonly used as stimuli in the laboratory. These experiments compared the responses of neurons in rat barrel cortex with trains of whisker movements with different frequencies; each train possessed either a periodic or an irregular, “noisy” temporal structure. Periodic stimulus trains were composed of a sequence of 21 whisker deflections separated by 20 equal interdeflection intervals (IDIs). Noisy trains were matched for mean IDI but included intervals shorter and longer than the mean IDI. Cortical responses were equivalent for periodic and noisy stimuli for frequencies up to 10 Hz. Above 10 Hz, temporal noise led to a larger response magnitude, and this effect was amplified as deflection frequency increased. Noise also caused a sharpening of the temporal precision of response to the individual deflections of the stimulus train. Cortical neurons thus appear to be “tuned” to respond in a different way to stimuli characterized by temporal unpredictability. As a consequence, perceptual judgments that depend on somatosensory cortical firing rate may be affected by the presence of temporal noise.
In natural settings, sensory stimuli often consist of sequences of discrete events distributed across a short time span. In audition, examples are the sound of an animal call—a series of discrete vocalizations in a birdsong (Woolley et al. 2006)—or the temporal envelope of human speech (Shannon et al. 1995); in touch, an example is a vibration—a sequence of movements transmitted to the skin mechanoreceptors. The present study aims to characterize the response of neurons in the somatosensory cortex of rats to sequences of repeated events (a train of whisker deflections) of 2 types: first, a regular temporal structure and, second, an irregular, unpredictable temporal structure. To date, the effect of frequency on barrel cortex response has been examined only for periodic whisker stimuli (Simons 1978; Khatri et al. 2004; Kleinfeld et al. 2006; Melzer, Champney, et al. 2006; Melzer, Sachdev, et al. 2006). Because the vibrations induced by movement of a whisker along a textured surface are approximated by a sequence of kinetic events occurring at unpredictable times—a “noisy” motion profile (Arabzadeh et al. 2005), it is important to know whether the presence of noise affects the basic features of neuronal coding. Here, we examine how the firing parameters of rat somatosensory cortical neurons differ in response to noisy, unpredictable stimuli compared with periodic, predictable stimuli. Our experiments indicate that barrel cortex neuronal response is altered by temporal noise—in particular, firing rate is elevated and the temporal precision of response (how closely spikes are aligned to the time of whisker deflection) is heightened. In general, the firing rate of cortical neurons is believed to contribute to sensory judgments. For example, in primates, it is correlated with the percept of vibration frequency (Salinas et al. 2000; Luna et al. 2005). Therefore, we discuss the present results in the context of how firing rate changes induced by the presence of noise might alter the percept of the underlying stimulus.
Subjects, Surgery, and Data Acquisition
All experiments were conducted in accordance with National Institutes of Health, international, and institutional standards for the care and use of animals in research. A summary of the experimental preparation is given below; for a detailed description, see Arabzadeh et al. (2003). Eight adult male Wistar rats, weighing 250–350 gm, were used. Anesthesia was induced by urethane (1.5 g/kg, intraperitoneally). A 10 × 10 grid of 1.5-mm-long electrodes with 400 μm tip-to-tip spacing (Cyberkinetics Inc., Foxborough, MA) was inserted into the whisker region of left somatosensory cortex, identified by vascular landmarks and stereotaxic coordinates (Rousche et al. 1999; Arabzadeh et al. 2003). The depths of electrode penetrations were 700–1000 μm in all experiments. Signals from all 100 electrodes were simultaneously amplified and filtered (gain 5000; bandpass 250–7500 Hz; Cyberkinetics Inc.). For each electrode, a threshold was set to a value of 2.5–3.5 times the root mean square voltage. A digital signal processor (Cyberkinetics Inc.) extracted 1 ms of voltage (0.33 ms before threshold crossing, 0.67 ms after) and saved the waveform at 30 000 samples per second. The waveforms produced by a multiunit neuronal cluster at each channel were selected using spike-sorting programs, and their time stamps were saved for further analysis. Earlier work showed that under these conditions, each electrode captures the activity of a cluster of about 3–5 neurons (Petersen and Diamond 2000). Single units were selected for Figure 7.
Trains of Periodic and Noisy Whisker Deflections—Standard Protocol
Using thin glass micropipettes, a 5-rung “ladder” was constructed and attached to a piezoelectric wafer (Morgan Matroc, Bedford, OH). With a micromanipulator, the ladder was positioned adjacent to the right side of the animal's snout so that the 5 rungs lay just below the corresponding 5 rows of whiskers (A–E), each whisker shaft resting lightly on the ladder about 5 mm from the skin. Whiskers were cut 1 cm from the base. One trial consisted of a train of 21 deflections, each deflection having the form of a truncated Gaussian of 3 ms duration (Supplementary material Fig. 1). This gave 20 interdeflection intervals (IDIs) per trial. Deflection frequency for that trial was 20 divided by the time between the 1st and the 21st deflections. A pause of 15 s was inserted between consecutive trials to allow recovery from the preceding stimulus train (Chung et al. 2002).
For 5 rats, the mean vibration frequency of the piezoelectric wafer on each trial assumed 1 of the 6 values—1, 2, 4, 10, 20, or 40 Hz, either “periodic” or “noisy”. Periodic trains consisted of 21 deflections with the identical IDI equal to 1000 ms/frequency (Fig. 1A). A stimulus train with 30% noise was created by mixing IDI values centered on the periodic IDI but also including intervals 15% and 30% shorter and longer (Fig. 1B). For 10-Hz stimuli, for example, in the periodic condition, all IDIs were 100 ms, whereas in the noisy condition, they were 70, 85, 100, 115, and 130 ms. One noisy train included 4 repetitions of each of the 5 possible IDIs, so that the total duration of the frequency-matched periodic and noisy trains was equal. Every trial had a unique sequence of IDIs (“mixed” noise condition), allowing each IDI to be presented within varying contexts. Because the temporal jitter in deflection time was scaled by frequency, the absolute amount of temporal jitter decreased as frequency increased. For each frequency, there were 120 trials (60 periodic and 60 noisy). Noisy and periodic trials with different frequencies were presented in a pseudorandom order.
Noisy Whisker Stimuli—Additional Protocols
To study the frequency dependence of neuronal response in greater detail, in 2 rats, we applied periodic and noisy trials with mean frequencies from 10 to 40 Hz in steps of 5 Hz. The noisy trials contained 30% noise in the mixed condition. For each frequency, 120 trials (60 periodic and 60 noisy) were applied.
To examine the effect of the quantity of noise on neuronal response, in 2 rats, we applied 25-Hz periodic and noisy stimulus trains with 3 noise levels (10%, 30%, and 50%). In these experiments, 60 periodic and 60 mixed noisy trials were presented for each noise level.
Finally, to study the temporal alignment of spike trains across trials, in 3 rats, we applied 25-Hz periodic and noisy stimuli (60 periodic and 60 noisy trials) where the noisy stimuli were “repeated.” This means that 1 sequence of noisy IDIs was selected at random and then presented on each trial. Noise level was 50%.
Supplementary material (Table 1) shows the details of stimulus presentation for each subject.
Receptive Field Mapping
After delivery of the stimulus set, whiskers were stimulated individually by a piezoelectric wafer with 60 pulses (Supplementary material Fig. 1) at 1 Hz. The resulting data were used to determine for each electrode whether the recorded neuronal cluster had a statistically significant response to whisker movement and, if so, to identify the principal whisker and the response onset latency (Armstrong-James et al. 1994). Clusters with statistically significant responses and a clear principal whisker were selected for analysis of responses to the full stimulus set. Response onset latencies were in the range of 5–9 ms, implying that recording sites were located in a cortical layer with direct thalamic input (Petersen and Diamond 2000), as expected from inspection of the electrode array insertion depth during the experiment.
To determine how the coding of stimulus trains depends on the temporal structure within the train, we measured barrel cortical responses to sequences of whisker deflections, where each sequence was either periodic or noisy. Stimulus trains consisted of 21 deflections (Supplementary material Fig. 1), giving 20 IDIs. Deflection frequency was 20 divided by the total duration between the 1st and the 21st deflection times. For a periodic train of whisker deflections, all IDIs were of equal length (Fig. 1A). For a noisy train of matched frequency, the IDI length varied within the train, but the total duration of the train was equal (Fig. 1B). In the “mixed noise” condition, each of the 60 noisy trains at a given frequency contained a unique sequence of IDIs. In the repeated noise condition, each noisy trial at a given frequency contained the identical sequence of IDIs. All results except those shown in Figure 4 were obtained with mixed noise (Supplementary material, Table 1).
Frequency Dependence of Neuronal Response and the Effect of Noise
The response magnitude of barrel cortex neurons to periodic and stimuli containing 30% noise is given in Figure 2A (percent noise refers to the maximum temporal interval that was added to and subtracted from the mean IDI; see Methods). At each recording site, firing rate is the multiunit cluster's spike count (number of spikes occurring between t0 and t21 + 50 ms, where t0 and t21 are the times of the 1st and the 21st deflections) divided by elapsed time between t0 and t21 + 50 ms, normalized to give a value of 1.0 for 1-Hz periodic stimulation. Firing rates were averaged across 76 recording sites in 5 rats. For both periodic and noisy stimuli, the mean firing rate increased as whisker deflection frequency increased (Pearson correlation coefficient, r = 0.911 for periodic stimuli and r = 0.946 for noisy stimuli). At frequencies of 1, 2, 4, and 10 Hz, responses to periodic and noisy stimuli were equivalent for the matched frequencies (paired t-test, P > 0.4 at all frequencies). However, at 20 and 40 Hz, the average firing rate was significantly greater for noisy than for periodic stimuli (P < 0.001 for both 20 and 40 Hz).
We further investigated the response difference to periodic and noisy stimuli by presenting trains with frequency steps of 5 Hz ranging from 10 to 40 Hz. At each recording site, response was calculated as the firing rate normalized to a value of 1.0 for 10-Hz periodic stimulation. Then, the normalized firing rates from 49 recording sites in 2 rats were averaged. The result was a gradual increase in the differential response to noisy and periodic stimuli as frequency increased (Fig. 2B). The difference in response to noisy versus periodic trains became statistically significant at 20 Hz and remained so for 25, 30, 35, and 40 Hz (paired t-test, P < 0.001 for all frequencies). We conclude that, for the selected parameters of temporal noise, neuronal sensitivity to the presence of noise began to emerge at 10 Hz, when the average IDI was 100 ms, and grew as stimulus frequency increased.
We also measured the number of spikes evoked in a 20-ms window after whisker deflection (Fig. 2c). When calculated as spikes/deflection, response magnitude decreased as stimulus frequency increased. As before, responses to noisy and periodic stimulus trains were equivalent at 1, 2, and 4 Hz. At 10 Hz, there was the first appearance of a facilitated response to noisy stimuli, but the difference was not statistically significant (paired t-test, P = 0.103). There were significantly larger responses for noisy than for periodic stimuli at 20 and 40 Hz (P < 0.005 for both cases). Analysis of phasic firing rate (0–20 ms postdeflection window) gave the same result for the stimulus set with 10–40 Hz in 5-Hz steps (Fig. 2D). Responses to noisy and periodic stimulus trains were equivalent at 10 Hz. At 15 Hz, there was the first appearance of a facilitated response to noisy stimuli, but the difference was not statistically significant (paired t-test, P = 0.218). Then, there were significantly larger responses for noisy than for periodic stimuli at 20, 25, 30, 35, and 40 Hz (P < 0.005 for all frequencies).
If stimulus frequency is encoded by firing rate in primary somatosensory cortex (for rats, see Arabzadeh et al. 2003, 2004; Melzer, Sachdev, et al. 2006; for primates, see Romo et al. 1998; Salinas et al. 2000; Luna et al. 2005), then the relationship between stimulus frequency difference and firing rate difference is the basis upon which frequencies could be discriminated. To define this relationship, we fit lines to the histogram values of Figure 2B, as plotted in the inset. For periodic stimulus trains, the slope of the line, in units of change in normalized firing rate per 1 Hz, was 0.031; the slope was steeper for noisy stimulus trains: 0.039.
The data shown in Figure 2B suggest that stimulus pairs separated by just 5 Hz would be readily discriminable by firing rate provided that both stimuli were periodic or both were noisy. For example, for 25- and 30-Hz periodic stimuli, firing rates in normalized units were 1.40 and 1.57, respectively, and the difference between mean responses was significant (paired t-test, P < 0.05). However, discriminations might be confused by the introduction of noise into the lower frequency train: when the 25-Hz vibration was noisy, the normalized firing rate rose to 1.60 (vs. 1.57 for 30-Hz periodic stimuli) and no frequency discrimination would be possible from firing rate (P = 0.526). By analyzing the mutual information between firing rate and vibration frequency, we were able to further characterize the effect of temporal noise on stimulus discriminability (Supplementary material Fig. 2).
Any property of neuronal response that varies systematically with the stimulus is a potential stimulus code. Accordingly, phasic firing rate (spikes per whisker deflection) could be a neuronal representation of frequency—this quantity decreased systematically as stimulus frequency increased (Fig. 2C,D). The frequency dependence of phasic firing rate differed for periodic and noisy stimuli, as plotted in the inset to Figure 2D: an observed phasic firing rate would be decoded as 2 different stimulus frequencies depending on whether the train was periodic or noisy. For example, a normalized phasic firing rate of 0.74 would be decoded as 25 Hz for a periodic train of deflections but as 30 Hz for a noisy train. Thus, the presence of noise could confound frequency discrimination whether the discrimination mechanism was based on the whole-stimulus firing rate or on the phasic firing rate to each deflection in the stimulus.
Graded Effects of Noise
Is the effect of temporal noise on neuronal responses graded or else all-or-none? In the second case, neuronal responses would be altered if the quantity of noise surpassed some threshold, but further increases in noise would not cause further changes in response. To test this, responses to a 25-Hz periodic train were compared with responses to trains with quantities of jitter set to 10%, 30%, or 50%. The data from 44 cortical neuronal clusters in 2 rats were similar and were averaged to generate 1 plot (Fig. 3). For the illustration, firing rate was normalized to give the value of 1.0 for the periodic condition. Increasing the noise level increased the neuronal response magnitude (Pearson correlation coefficient, r = 0.97). This result excludes the possibility that cortical networks switched to some qualitatively different state when the quantity of noise exceeded a threshold level. Rather, larger quantities of temporal jitter led to progressively greater responses.
Temporal Precision of Neuronal Response
We now focus on the effect of noise on the precision of response to individual deflections. This can be addressed by the repeated noise condition, where the identical sequence of IDIs was presented in each of 60 trials. The data came from a total of 70 recording sites in 3 rats. Figure 4A shows raster plots collected at 2 electrodes during the first 400 ms of the 25-Hz periodic trains (Rat 6). After about 200 ms, responses tended to become poorly aligned to whisker deflection times (red ticks along time axis). The progressive loss of timing precision is also evident in the peristimulus time histograms (PSTHs) covering the same time span as the raster plots (Fig. 4B: red, electrode 26; blue, electrode 46).
Temporal response precision for the same cortical neurons was finer during noisy stimulation. Figure 4C shows the raster plots collected during 25-Hz trains containing 50% noise. Neurons did not fire for all deflections, but deflections that evoked a response did so in a reliable manner. Maintenance of response precision—compared with the periodic condition—is evident in the PSTHs (Fig. 4D). Neuronal responses were well aligned to whisker deflection times (red ticks) throughout much of the stimulus train (compare with Fig. 4B). The improved temporal fidelity during noisy stimuli held for all the neuronal clusters in the 3 rats that received repeated noise, as estimated by summating responses across all electrodes to form a population PSTH (Fig. 4E,F).
For the 2 neuronal clusters presented in Figure 4A–D, we also used the full stimulus train to generate PSTHs aligned on whisker deflection time (Fig. 4G,H); these extended from −10 to +30 ms (deflection time = 0 ms). For both neuronal clusters, the response to noisy stimuli (black PSTH labeled N) was stronger than the response to periodic stimuli (dark gray PSTH labeled P). The difference in PSTH (light gray, N-P) can be taken to represent the “extra” spikes, per deflection, caused by the presence of noise.
The findings of Figure 4 demonstrate that the greater number of spikes that occurred during noisy trains of whisker deflection was not due to a general upregulation of neuronal firing, distributed broadly in time, but to a strengthened response temporally linked to the whisker deflection.
To characterize the synchrony of firing between neuronal clusters during periodic and noisy stimulus trains, we used cross-correlation analysis. The same neurons illustrated in Figure 4 were analyzed again, but now under the 25 Hz, 50% mixed noise (not repeated noise) condition. For each pair of neuronal clusters, a cross-correlation histogram (CCH) with 2-ms bins extending 20 bins before and after the central bin (total, 41 bins) was constructed during 25-Hz periodic and noisy stimulus trains. To allow comparison across neurons with different firing rates, CCHs were normalized such that their scaling corresponded to the correlation coefficient rather than the number of coincident events (Palm et al. 1988; Erchova and Diamond 2004). The cross-correlation index was then defined as the average coefficient value extending from 2 bins before to 2 bins after the central bin (total 5 bins of 2 ms each). This is equivalent to denoting spike pairs from the 2 clusters as “synchronous” whenever they were emitted within 5 ms of each other. Finally, the difference between the cross-correlation index during the periodic stimulus train, IP, and the same index during the noisy stimulus train, IN, was expressed as the correlation difference index, ID(eq. 1):
Time Course of Recovery from Preceding Deflections
Neurons in barrel cortex are known to exhibit a reduced response to whisker deflection n due to the suppressive effect of the preceding deflection, n − 1 (Simons 1985; Erchova et al. 2003; Boloori and Stanley 2006; Webber and Stanley 2006). The amount of suppression as a function of preceding IDI (the time between deflections n − 1 and n) is a fundamental measure of how neurons integrate sequences of inputs. Is the amount of suppression affected by noise? For periodic stimuli, we obtained each point in the IDI–response curve by computing the mean response across a stimulus train at one frequency, f, where IDI = 1/f. Within noisy stimulus trains, IDIs vary around their mean, so we could compute the full IDI–response curve from the multiple IDI values contained in the trains at a single mean frequency. The stimulus set comprised periodic trains at 15, 20, 25, 30, 35, and 40 Hz (giving IDIs of 67, 50, 40, 33, 29, and 25 ms, respectively) and a 25-Hz train containing 50% noise (IDIs of 60, 50, 40, 30, and 20 ms). Response was taken as the mean number of spikes within 20 ms after whisker deflection.
Data averaged across 70 recording sites in 3 rats are illustrated in Figure 6: the key result is that the entire IDI–response curve shifted upward during 25-Hz noisy stimuli (black plot, periodic; red plot, noisy). Thus, the response of cortical neurons to a whisker deflection preceded by any given IDI was larger in the context of a temporally noisy stimulus train than in the context of a periodic stimulus train. We also plotted the response curve for noisy trains when the second-order IDI sequence was locally periodic, that is, when 2 successive IDIs were equal (blue plot). Because the curve exactly overlies the red one, the “memory” of neurons for the presence of noise must be longer than a period of 2 preceding IDIs.
Although in our experiments pairs of deflections were concatenated within a stimulus train, whereas in previous work each pair of deflections was isolated by a long interval from neighboring pairs, Figure 6 confirms the well known observation that the effect of a preceding whisker deflection does not decay linearly: the release from suppression caused by an increase in IDI by some fixed amount can be larger than the augmentation of suppression caused by a decrease in IDI by the same fixed amount (Simons 1985; Erchova et al. 2003; Boloori and Stanley 2006; Webber and Stanley 2006). This implies that the heightened response to noisy stimuli, as seen in Figures 2 and 3, could have occurred even if the IDI–response curves for periodic stimuli and noisy stimuli overlay each other, as follows. A noisy stimulus with mean frequency f consists of a train of IDIs varying around the value 1/f. The nonlinearity in the IDI–response curve for periodic stimuli means that the response “gain” for IDI values >1/f would be greater than the response “loss” for IDI values <1/f, as shown schematically in the inset of Figure 6 (the upward pointing arrows representing response gain are larger than the downward pointing arrows representing response loss). However, such asymmetry can offer only a partial explanation for the enhanced response to noisy stimuli—the noise-induced upward shift in the IDI–response curve signifies that the principal cause of increased mean firing rate was an increase in responsiveness due to the presence of noise.
Does Noise Act through Inhibitory Networks?
Networks of inhibitory neurons can be entrained by repetitive inputs (Llinas et al. 1991; Fellous et al. 2001). We hypothesize that during a periodic stimulus, their entrainment might be particularly pronounced and they will, as a consequence, suppress the rest of the cortical population; during a noisy stimulus, inhibitory networks may suppress their targets less effectively. Neurons in barrel cortex are categorized as regular-spiking units (RSUs), putative excitatory neurons, and fast-spiking units (FSUs), putative inhibitory neurons (Simons 1978). Because less than 15% of neurons in barrel cortex are positive for γ-aminobutyric acid (Beaulieu 1993), the results presented until now mainly reflect the behavior of excitatory neurons. In order to test whether noise acts differently on distinct functional classes of neurons, we isolated 5 FSUs and 23 RSUs from 70 neuronal clusters recorded in 3 rats. Criteria for selection of FSUs were a spike duration <0.65 ms (Fig. 7A, left) and a spontaneous activity rate (measured in the 15-s interval between trains) >4 spikes per second; RSU criteria were a spike duration >0.75 ms (Fig. 7A, right) and a spontaneous activity rate <3 spikes per second. These criteria are consistent with previous studies (Simons 1978; Armstrong-James et al. 1993; Bruno and Simons 2002).
As expected, the RSUs showed larger responses during 25-Hz trains containing 50% noise than during 25-Hz periodic trains (4.7 vs. 3.4 spikes per second, respectively, averaged across the entire stimulus train; paired t-test, P < 0.01). But the FSUs showed the opposite effect of noise—they gave larger responses during 25-Hz periodic trains than during 25-Hz trains containing 50% noise (13.5 vs. 11.7 spikes per second, respectively, averaged across the entire stimulus train; paired t-test, P < 0.01). Further evidence for a direct interaction between inhibitory and excitatory neurons comes from the time course of their responses (Fig. 7B). The FSUs showed a preferential response to periodic stimuli beginning 160 ms after stimulus onset, precisely when the RSUs began to show a preferential response to noisy stimuli. In light of the concurrent but opposite effects of noise in putative interneurons, our interpretation is that their reduced response to noisy inputs releases the remaining cortical population from suppression.
Firing Rate as a Function of Stimulus Frequency
When a train of stimuli is presented to the whiskers, neuronal firing rate measured over the whole-stimulus train increases as stimulus frequency increases (Ahissar et al. 2001; Arabzadeh et al. 2003; Khatri et al. 2004; Kleinfeld et al. 2006; Melzer, Champney, et al. 2006; Melzer, Sachdev, et al. 2006), as we found in Figure 2A,B for both periodic and noisy stimuli. Firing rate measured per individual deflection decreases as stimulus frequency increases (Simons 1978; Castro-Alamancos 2004; Khatri et al. 2004; Melzer, Sachdev, et al. 2006), as we found in Figure 2C,D for both periodic and noisy stimuli. These observations have been replicated using a variety of electrophysiological methods including local field potential, multiunit, and single-unit recordings (Chung et al. 2002; Castro-Alamancos 2004; Andermann et al. 2004; Melzer, Champney, et al. 2006; Melzer, Sachdev, et al. 2006). An exception to the monotonic relationship between deflection frequency and firing rate was pointed out by Garabedian et al. (2003): firing rate peaked at stimulus frequencies of 5–10 Hz. However, that analysis measured “steady state” responses from 1–5 s after the onset of the stimulus train, excluding what they call the “dynamic period” of response. We included this dynamic initial response in our analysis because the timescale of rat (Carvell and Simons 1990) and primate (Luna et al. 2005) tactile discriminations can be on the order of a few hundred milliseconds up to a second.
What is the exact shape of the monotonic increase? Previous studies have demonstrated that factors such as the cell type studied (Simons 1978; Khatri et al. 2004), the shape and amplitude of the individual stimulus deflection (Arabzadeh et al. 2003; Khatri et al. 2004), laminar location (Derdikman et al. 2006), location with respect to barrel boundaries (Melzer, Champney, et al. 2006; Melzer, Suchdev, et al. 2006), and overall brain state (Erchova et al. 2002; Castro-Alamancos 2004) influence the relationship between the frequency of the applied stimulus and neuronal firing rate (reviewed in Maravall et al. 2007). In general, the relationship has been found to be either linear (Chung et al. 2002; Arabzadeh et al. 2003) or logarithmic (Arabzadeh et al. 2003; Andermann et al. 2004; Melzer, Champney, et al. 2006; Melzer, Suchdev, et al. 2006). The discrepancy may be resolved by considering deflection amplitude together with frequency—at low amplitude, the relationship between stimulus frequency and firing rate was linear and at high amplitude it was logarithmic due to saturation of the neuronal dynamic range (Arabzadeh et al. 2003). In the present data, the relationship was linear (Fig. 2), indicating that neurons did not reach the limit of their dynamic range.
The Effect of Temporal Noise on Neuronal Response
In our data, the slope of the firing rate increase was steeper for noisy than for periodic trains at frequencies above 10 Hz (Fig. 2B, inset). Firing rate per deflection decreased as stimulus frequency increased, and the slope was less steep for noisy than for periodic trains (Fig. 2D, inset). The effect of noise was graded such that larger quantities of noise led to progressively greater responses (Fig. 3). In addition to the effects on response magnitude, noise enhanced the temporal firing precision of neurons (Figs 4 and 5).
Two observations point to involvement of inhibitory mechanisms in the enhanced response to noise. First is the upward shift in the IDI–response curve (Fig. 6). Because the suppression of response to whisker deflection n is taken to reflect an inhibitory effect arising from the preceding deflection, n − 1 (Simons 1985; Erchova et al. 2003; Boloori and Stanley 2006; Webber and Stanley 2006), any reduced suppression could reflect reduced inhibition. Second is the fact that FSUs—putative inhibitory neurons—were more effectively driven by periodic than by noisy stimulus trains (Fig. 7). What might underlie the FSU “preference” for periodic stimuli? One clue comes from the strong, 10- to 50-Hz intrinsic oscillation of membrane potential of inhibitory neurons in mammalian cortex (Llinas et al. 1991; Fellous et al. 2001). We speculate that periodic inputs resonate with intrinsic oscillations more effectively than do inputs with temporal noise and therefore drive inhibitory neurons more effectively.
Another mechanism that could play a role in the differential response to periodic and noisy stimuli is the presence of frequency-selective synapses in the whisker-to-barrel system. Facilitating and depressing short-term dynamics endow synapses with frequency filtering characteristics (Fortune and Rose 2001; Abbott and Regehr 2004). A population of synapses onto a barrel cortex neuron can have heterogeneous facilitatory and inhibitory properties (Markram et al. 1998). A periodic stimulus would excite only a restricted synaptic population (the synapses whose band-pass matches the frequency of the periodic stimulus), and these would quickly fatigue. A noisy stimulus centered at the same frequency would excite a variety of synapses with different band-pass properties; because each synapse would be excited by only a fraction of stimulus events, they could recover to a more refreshed state between successive activations. Increased quantities of noise would “distribute” activity more widely across synapses, explaining our observation of the graded effects of noise (Fig. 3).
Common Effects of Temporal Noise across Modalities
Ganglion cells in the retina of mouse and salamander fire strongly in response to a deviation from periodicity in the stimulus train (Schwartz et al. 2007). In alert monkeys, neurons in middle temporal cortex emit more reproducible spike trains when the animal views motion stimuli with unpredictable velocity changes as compared with stimuli with constant velocity (Buracas et al. 1998). Moreover, functional magnetic resonance imaging studies in humans reveal reduced blood oxygen level–dependent response in visual cortex in response to periodic presentation of checkerboard as compared with aperiodic presentation (Parkes et al. 2004). Sounds with more natural frequency dynamics cause primary auditory cortex neurons to fire at higher rates and with better timing fidelity than do sounds with unnatural frequency dynamics (Garcia-Lazaro et al. 2006). The present work extends to the tactile modality the fundamental idea that neurons express different firing behavior in response to unpredictable time-varying features.
In the whisker sensory system, stimuli marked by irregular temporal sequences occur under a number of conditions. Of particular interest are the vibrations induced by whisker movement along textured surfaces (Arabzadeh et al. 2005). The kinetic signal transmitted to the whisker follicle can be defined as an irregular vibration centered around a frequency that is characteristic for the contacted texture (Arabzadeh et al. 2006; Hipp et al. 2006). The present results suggest that the “noisiness” of such kinetic stimuli helps maintain precise neuronal firing along the whisk cycle. Recently, a study of membrane potential in an awake mouse, which actively whisked an object, revealed that individual object contacts evoked robust sensory responses with a low coefficient of variation (Crochet and Petersen 2006). The fidelity of responses to unpredictable whisker contacts might be comparable with the reliable spiking in response to the noisy, unpredictable stimulation pattern observed in the present study.
Possible Sensory Correlates
Although there is no certainty that our findings will generalize to the awake brain, we can nevertheless speculate how temporal noise in a vibratory stimulus could affect frequency discrimination. We pose the question in relation to experiments in which monkeys were trained to compare the frequencies of 2 flutter vibration stimuli applied to the finger tip (Romo et al. 1998; Salinas et al. 2000; Luna et al. 2005). In the primary somatosensory cortex, spike count per trial was the feature of neuronal activity most closely correlated with the animal's judgment of stimulus frequencies in the range 14–30 Hz (Luna et al. 2005). Spike counts were not altered by the addition of temporal noise to the vibration; thus, response magnitude for a noisy stimulus with a given mean frequency was equal, on average, to response magnitude for a periodic stimulus with this same frequency.
These results would seem at odds with our current finding that, for frequencies of 20 Hz or more, cortical neurons fired at higher rates in response to noisy stimuli. Whereas monkeys could compare the frequencies of a noisy and periodic stimulus train just as accurately as 2 periodic trains (Romo et al. 1998; Salinas et al. 2000; Luna et al. 2005), our data suggest that a firing rate coding scheme would result in a systematic error in frequency estimation whenever one of the stimuli contained temporal noise (Fig. 2). At this point, there is not enough evidence available to reconcile the findings and we can only hypothesize. In the untrained somatosensory cortex, as we have tested here, response differences to noisy and periodic stimuli may occur when differential engagement of inhibitory networks (Figs 6 and 7) leads to varying degrees of adaptation. But in the cortex of monkeys that have received extensive periods of training and testing, inhibition may stay at a consistently high level and excitatory networks may be, as a consequence, in a state of long-term, stable adaptation. This would abolish firing rate differences to the 2 kinds of stimuli, consistent with the physiological and psychophysical evidence (Romo et al. 1998; Salinas et al. 2000; Luna et al. 2005). Consistent with this hypothesis, when awake rats were trained to receive a whisker stimulus, the degree of adaptation increased as the animal learned the task (Castro-Alamancos 2004).
In recent work (Harris et al. 2006), we have found that the judgment of vibrotactile stimulus frequency in naive human subjects is impaired when the temporal structure of the vibrations is noisy. This indicates that temporal noise can affect neuronal representations at some level involving the perceptual readout. Still, much additional work is required to understand the physiological and perceptual effects of temporal noise. Any observation of covariance between neuronal response and percept could shed light on the brain basis of sensation.
Supplementary material can be found at: http://www.cercor.oxfordjournals.org.
Human Frontier Science Program (RG0043/2004-C); European Community (IST-2000-28127); the Ministry of Universities and Research (2006-050482); and the Region of Friuli Venezia Giulia. Human Frontier Science Program Postdoctoral Fellowship (E.A.).
We are grateful to Erik Zorzin for instrumentation. Miguel Maravall Rodriguez, John Nicholls, Moritz von Heimendahl, and Rasmus Petersen participated in helpful discussions. Conflict of Interest: None declared.