Neural coding and perceptual detection in the primate somatosensory thalamus

Abstract

The contribution of the sensory thalamus to perception and decision making is not well understood. We addressed this problem by recording single neurons in the ventral posterior lateral (VPL) nucleus of the somatosensory thalamus while trained monkeys judged the presence or absence of a vibrotactile stimulus of variable amplitude applied to the skin of a fingertip. We found that neurons in the VPL nucleus modulated their firing rate as a function of stimulus amplitude, and that such modulations accounted for the monkeys’ overall psychophysical performance. These neural responses did not predict the animals' decision reports in individual trials, however. Moreover, the sensitivity to changes in stimulus amplitude was similar when the monkeys’ performed the detection task and when they were not required to report stimulus detection. These results suggest that the primate somatosensory thalamus likely provides a reliable neural representation of the sensory input to the cerebral cortex, where sensory information is transformed and combined with other cognitive components associated with behavioral performance.

Detection of a sensory stimulus arises from evoked neural activity starting in the sensory receptors (1) and spanning several subcortical relay stations up to the cortex (2). Previous studies have described the neural activity of relay neurons within the sensory thalamus and its association with cortical activity; however, most of these studies were performed in anesthetized animals (3–7). Only a few studies have recorded thalamic neural activity from behaving subjects (8–11). Thus, the relationship between neuronal activity in the sensory thalamus and a subject’s performance is not clear. In the case of the primate somatosensory thalamus, it is not known how neurons in the primate ventral posterior lateral (VPL) nucleus encode tactile stimuli and impact the animal’s psychophysical behavior.

To further investigate this relationship, we recorded the activity of single neurons in the VPL nucleus while trained monkeys reported the presence or absence of a mechanical vibration of variable amplitude applied to the skin of a fingertip (2). This task allowed us to study how the firing activity that encodes features of the stimulus is related to the animal's psychophysical performance and decision making capacity. We found that VPL neurons with either quickly adapting (QA) or slowly adapting (SA) response properties modulate their firing rates as functions of stimulus amplitude. On average, these modulations accounted for the monkeys’ detection performance, in that neural and behavioral sensitivities were statistically the same, although the sensitivity of most neurons was lower than that of the monkeys when relatively short integration windows were used to measure the rate modulations. Moreover, variations in the firing rate of VPL neurons did not predict the monkeys’ perceptual judgments or motor reports. Finally, the sensitivity to changes in stimulus amplitude was similar during task performance and during passive stimulation, when the monkeys were not required to respond to the stimuli. These results are consistent with the idea that primate thalamic somatosensory neurons essentially provide a reliable neural representation of the sensory world to the cerebral cortex, which then processes the sensory information and combines it with other cognitive elements (e.g., attention, memory) to generate behavior that is congruent with current task demands.

Results

Two monkeys (Macaca mulatta) were trained to perform a vibrotactile detection task (2). In each trial, the animal had to report whether the tip of a mechanical stimulator vibrated or not (Fig. 1A). Stimuli were sinusoidal waves with a fixed frequency of 20 Hz, delivered to the glabrous skin of one fingertip of the restrained hand; crucially, they varied in amplitude across trials. Stimulus-present trials were interleaved at random with an equal number of stimulus-absent trials in which no mechanical vibration was delivered (Fig. 1A). Based on the monkeys’ reports, trials were classified into four types: hits and misses in the stimulus-present condition, and correct rejections (CRs) and false alarms (FAs) in the stimulus-absent condition (Fig. 1B). Stimulus detection thresholds were calculated from the decision reports (Fig. 1C). We recorded 74 VPL (Fig. 1D) neurons with small cutaneous receptive fields confined to the glabrous skin of one fingertip.

Fig. 1.

Detection task, psychophysical performance, and recording area. (A) Trials began when the stimulator probe indented the skin of one fingertip of the monkey’s restrained right hand (PD). Then the monkey placed its free, left hand on an immovable key (KD, key down). After a variable prestimulus period (1.5–3 s), in one-half of the trials a vibratory stimulus of 20 Hz (variable amplitude, 1–34 μm) was delivered to the glabrous skin of one fingertip. In the other half of the trials, no stimulus was delivered. Ten repetitions per stimulus amplitude were presented, counterbalanced with the same number of stimulus-absent trials, all interleaved at random. After stimulus presentation, the monkey waited for 2 or 3 s until the probe was lifted (PU). This was the cue signal to remove its free hand from the key (KU, key up) and indicate whether the stimulus was present or absent by pressing one of two response buttons. The time interval between KU and button press (PB) is shown in all figures as movement time (MT). (B) The detection task elicited four behavioral responses: hits or misses during stimulus-present trials and CRs or FAs during stimulus-absent trials. (C) Psychometric function showing the probability that the monkey reports the presence of the stimulus, as a function of stimulus amplitude. Each point represents an average over 58 sessions. The line is a Boltzmann fit to the data points. Its amplitude modulation (ΔA = 0.87) is the difference in probability between the maximum and minimum amplitudes. The detection threshold (t_p = 8.0 μm) and slope of the curve (s_p = 0.062 μm⁻¹) at a probability of 0.5 are indicated. (D) The recording site at the VPL nucleus of the somatosensory thalamus.

VPL Responses During Vibrotactile Detection.

Neurons were classified as QA (n = 64) or SA (n = 9) (Fig. 2) depending on their firing rate adaptation to the gentle skin indentation of the probe tip at the beginning of each trial. QA neurons demonstrated an abrupt increment in firing rate after the probe touched the skin [probe down (PD)] but returned to their spontaneous rate in less than 200 ms (Fig. 2A), whereas SA neurons maintained the increased firing rate from PD until the probe was lifted off from the skin [probe up (PU)] (Fig. 2C). During stimulus presentation, both QA and SA neurons modulated their firing rates as a function of stimulus amplitude (Fig. 2 A and C). Both QA and SA neurons had significantly higher spontaneous firing rates compared with primary somatosensory cortex (S1) neurons (median ± SEM, 17.5 ± 1.1 Hz, n = 74 vs. 10 ± 1.0 Hz, n = 59; P = 0.0002) (2). In addition, the response latency was significantly shorter in QA and SA neurons than in S1 neurons (median ± SEM, 19.5 ± 2.0 ms, n = 74 vs. 23.0 ± 0.7 ms, n = 59; P = 0.0003) (12). In all of the analyses described below, no notable differences between the QA and SA populations were found.

Fig. 2.

Activity of two VPL neurons during the detection task. (A) Raster plot of a QA neuron. Each row is a trial, and each black tick represents the time at which an action potential occurred. Trials are arranged according to stimulus amplitude, shown on the left. Blue and red marks indicate behavioral responses: hits and misses, respectively, for stimulus-present trials and CRs and FAs, respectively, for stimulus-absent trials. The upper line indicates relevant task events (PD and PU) and the monkey’s MT. The gray box depicts the stimulation period. (B) Firing rate (blue) and power at 20 Hz (green) as functions of stimulus amplitude (mean ± SD; n = 10 trials). The blue and green lines correspond to linear regression fits. The slopes for each fit are indicated. Slopes were significantly different from 0 exclusively during the stimulation period (P ≤ 0.05). (C) Raster plot of an SA neuron, with the same conventions as in B. (D) Slopes were significantly different from 0 during the stimulus period.

We used linear regression analysis to characterize the association between stimulus amplitude and firing rate (Fig. 2 B and D,blue lines) (Methods). We found that 92% of the VPL neurons had significant positive slopes (mean ± SD, 1.15 ± 0.75 Hz/μm; n = 68), and so their firing rates increased monotonically with stimulus amplitude. In addition to the firing rate modulation, most neurons showed spike trains that were phase-locked to the stimulus sinusoidal wave. To measure the strength of spike train periodicity as a function of stimulus amplitude, we calculated the power of the evoked spike trains at 20 Hz and applied a linear regression analysis to it. In 91% of the neurons, the power increased as a function of stimulus amplitude (Fig. 2 B and D, green lines). All neurons with periodic responses had a positive slope (1.20 × 10⁻⁴ ± 0.1 × 10⁻⁴ AU/μm).

VPL Responses Correlate with Psychophysical Detection Performance.

To test whether the responses of VPL neurons accounted for the monkeys’ psychophysical performance, we calculated neurometric detection curves based on a firing rate criterion threshold and compared them with the behavioral curves (2). As a first approach, we used the firing rate modulation during the entire stimulus period (500 ms). The proportion of “yes” responses for the neurometric curves was defined for any amplitude as the proportion of trials in which the neuron's firing rate reached or exceeded a criterion value (2). The criterion value for each neuron was chosen to maximize the total number of correct trials. Fig. 3A shows the resulting neurometric curves (red lines) and the associated psychometric curves (black lines; n = 51). The neurometric and psychometric curves were generally similar. In particular, no difference was seen between the median slopes of the psychometric and neurometric curves (s_p = 0.059 vs. s_n = 0.052; P = 0.13). However, amplitude modulation was greater in the psychometric curves compared with the neurometric curves (ΔA_p = 0.88 vs. ΔA_n = 0.81; P = 0.02), suggesting that the monkeys detected the stimuli slightly more efficiently than single VPL neurons. This is clearly seen when comparing the mean psychometric and neurometric curves (Fig. 3B), which were similar in shape, but with slightly better (higher) amplitude and slope in the mean psychometric curve.

Fig. 3.

Comparison of VPL activity and psychophysical performance during the detection task. All plots show the probability that the monkey's or the neuron's response indicated that a stimulus was present (“yes”) as a function of stimulus amplitude. (A, C, and E) Individual psychometric and neurometric curves. The black lines indicate the monkey’s probability of detection during each recording session. The red lines indicate each neuron’s probability of detection based on either the firing rate (A and E) or periodicity (C) exceeding a criterion threshold. In A and C, firing rate and periodicity were computed over a 500-ms window during the stimulation period. In E, firing rate was computed over a 50-ms window during the stimulation period. (B, D, and F) Mean psychometric (black) and neurometric (red) curves, obtained by averaging the corresponding groups of curves in A (n = 51), C (n = 58), and E (n = 30). ∆A_p and s_p (∆A_n and s_n) indicate the amplitude modulation and slope of the average psychometric (neurometric) curves, respectively.

This finding does not rule out the possible existence of other neural codes that may account for psychophysical performance, however. For example, given the high proportion (91%) of neurons that emitted spikes in phase with the stimulus sinusoidal wave, an ideal observer could detect the stimulus simply by analyzing the periodicity of the evoked spike trains. Thus, for each neuron we calculated a neurometric curve based on the spike train periodicity elicited during the stimulation period (Fig. 3C, red curves) and compared it with the corresponding psychometric curve (Fig. 3C, black curves). In this case, the results were more variable. Some neurons were much better than the monkeys in detecting the stimulus, whereas others were much worse. The median slopes and amplitudes were significantly different in the psychometric curves and the neurometric curves (s_p = 0.054 and ΔA_p = 0.88 vs. s_n = 0.10 and ΔA_n = 0.94; P = 0.001 and 0.005). In addition, the mean psychometric and neurometric curves were similar in shape but differed quantitatively (Fig. 3D); the mean psychometric curve was steeper than the mean neurometric curve (s_p = 0.061 vs. s_n = 0.037), suggesting that on average the monkey was more sensitive than the neurons to changes in stimulus amplitude when detection was based on spike train periodicity, regardless of the few neurons with high detection performance. Notably, the amplitude modulation was smaller in the psychometric curve compared with the neurometric curve (ΔA_p = 0.87 vs. ΔA_n = 0.91); this is because the neuronal FA rate was almost 0 when using periodicity to infer the presence of a stimulus.

Spike train periodicity is strictly related to the type of stimulus used here, and this sensory feature might be redundant or irrelevant for stimulus detection. To take periodicity out of the picture and focus on firing rate as a potential neural code for stimulus detection, we constructed neurometric curves based on firing rate modulations within a window of only 50 ms. This small window, which corresponds to one period of the sinusoidal wave, does not allow access to the periodic structure of the spike train, but could be sufficient to allow a hypothetical observer to track the variations in firing rate elicited by each indentation pulse of the stimulus (Fig. 4 A and B). In addition, the sensitivity to the stimulus amplitude was stronger for the first stimulation pulses (Fig. 4C). Fig. 3E shows the neurometric curves (red traces) based on such a firing rate code, along with their psychometric counterparts (black traces). It can be seen that single neurons were still able to signal stimulus presence quite well in this case, although they generally were less sensitive than the animals. The slopes of the psychometric curves were significantly higher than the slopes of the neurometric curves (s_p = 0.060 vs. s_n = 0.030; P = 0.001). In addition, the psychometric curves had significantly higher amplitude modulations (ΔA_p = 0.89 vs. ΔA_n= 0.63; P = 0.001). Fig. 3F shows the mean psychometric and neurometric curves from the data in Fig. 3E, which confirm these effects. These results suggest that firing rate modulation in a short time window is a plausible candidate for coding the amplitude of the vibrotactile stimulus, and may account for the monkeys’ overall detection performance.

Fig. 4.

Population response of VPL neurons to a sinusoidal stimulus. In all panels, 0 indicates stimulus onset. (A) Representation of a periodic sinusoidal wave with an amplitude of 34 μm applied to the skin of a fingertip. (B) Mean spike density function (SDF) for the population of VPL neurons (n = 54) in response to the stimulus shown in A. The SDF indicates the presence of more spikes within the first pulses of the stimulus compared with the last pulses. The SDF was calculated using a 2-ms-wide Gaussian filter. (C) Slopes (mean ± SD; n = 63) calculated at different times after onset of stimulus. Slopes indicate neuronal sensitivity to changes in stimulus amplitude, calculated by fitting a linear regression to the firing rate as a function of the stimulus amplitude (0–34 μm). Firing rates were calculated within a 50-ms window at different times after the onset of stimulus. Only significant slopes were used in the analysis. The sensitivity to changes in stimulus amplitude was significantly lower after the fourth sinusoidal pulse compared with the first sinusoidal wave (*P = 0.001).

VPL Responses Do Not Covary with Perceptual Detection Reports.

We sought to quantify whether the activity of VPL neurons covaried with the perceptual “yes” or “no” judgments that monkeys made on a trial-by-trial basis. For this, we compared the mean normalized activity (z-scores) during hits and misses for near-threshold stimuli, as well as the normalized activity during CRs and FAs in the stimulus-absent condition. We found no significant differences between hits and misses (P = 0.21, permutation test) (Fig. 5A, Left) nor between CRs and FAs (P = 0.77, permutation test) (Fig. 5A, Right). This result indicates that within trials with identical stimulation conditions, fluctuations in the responses of individual VPL neurons did not covary with fluctuations in the monkey’s behavior. To further quantify this finding, we calculated a choice probability index, which estimates the probability with which the behavioral outcomes can be predicted from the neuronal responses (Fig. 5B). This analysis also revealed no significant differences (α = 0.01, permutation test) between hits and misses (Fig. 5B, Left) nor between CRs and FAs (Fig. 5B, Right). In addition, no significant differences (α = 0.01) in the maximum firing rate within each whole trial were found between FAs and CRs. Moreover, no significant covariations were found for individual single neurons.

Fig. 5.

Correlation between VPL activity and the monkeys’ perceptual judgments. (A) Comparison of normalized neuronal responses for hits (n = 372, black line) and misses (n = 376, red line) during near-threshold stimulation (Left) and for CRs (n = 873, black line) and FAs (n = 873, red line) during stimulus-absent trials (Right). Upper lines and gray boxes indicate relevant task events (stimulus onset and offset, PU, and MT). (B) Choice probability as a function of time. The black solid line represents the choice probability index between hits and misses for near-threshold trials (Left) and between FAs and CRs for stimulus-absent trials (Right). The dashed gray lines represent mean choice probability values obtained by resampling the values from the original distributions 1,200 times, thereby shuffling the hit/miss and the CR/FA labels. Mean choice probability values were 0.4766 ± 0.14 for the stimulus-present trials and 0.5 ± 0.0004 for the stimulus-absent trials. No significant values were found (α = 0.01).

Our monkeys also performed a visual instruction task that allowed us to dissociate the perceptual judgment and the motor command associated with its report (Methods). Choice probability indices for this task were calculated based on trials with the same stimulus amplitudes used during the detection task, but unrelated to the correct motor responses. In one half of the trials (chosen at random), the monkeys were cued to push the right button, whereas in the other half they were cued to push the left button. As expected, choice probability values were not significantly different between these two response conditions (0.5; α = 0.01), indicating that VPL activity did not predict the animals’ motor choices in this case.

Context Dependence of VPL Neuronal Activity.

Finally, we compared the sensitivity of VPL neurons to the stimulus amplitude in a behavioral context in which monkeys actively reported stimulus detection versus a condition in which they passively received the stimuli and no report was required. Fig. 6 shows the slope values obtained from linear fits applied to the firing rate as a function of stimulus amplitude in the two conditions (n = 28). No significant differences were found among the slope distributions obtained during active detection and passive stimulation (mean ± SEM, 1.22 ± 0.16 Hz/μm vs. 1.08 ± 0.14 Hz/μm; P = 0.56). We also searched for differences in the neurons’ response latency with respect to stimulus onset. The latencies of VPL neurons (n = 28) during the standard detection task were not significantly different from those obtained during passive stimulation (mean ± SEM, 19.5 ± 2.4 ms vs. 18.0 ± 3.2 ms; P = 0.45).

Fig. 6.

Responses of VPL neurons during task performance versus passive stimulation. Slopes were obtained from the linear regression fits performed on the firing rate as a function of stimulus amplitude for 28 neurons recorded in two conditions: during active detection and during passive delivery of the stimuli, for which no responses by the monkeys were required. Points are close to the diagonal, indicating similar slopes in the two conditions. No significant differences were found (P = 0.56; n = number of neurons). The red cross indicates the slope values for active detection (mean ± SEM, 1.22 ± 0.16 Hz/μm) and passive stimulation (1.08 ± 0.14 Hz/μm).

Discussion

The sensory thalamus is classically viewed as a relay station of sensory information to the cerebral cortex, but recent studies suggest that it contributes to perception and decision making (9, 13, 14). Other studies have not assigned such roles, however. The present study addressed this problem by using a somatosensory detection task in which monkeys had to report the presence or absence of stimulus. The task allowed us to carefully explore the role of the somatosensory thalamus in perceptual performance and decision making.

Our results show that the activity of VPL neurons is faithfully modulated as a function of stimulus amplitude, such that it matches overall psychophysical performance. However, VPL activity does not predict an animal’s decision in individual trials, suggesting that the somatosensory thalamus is more likely associated with providing information to the somatosensory cortex, where other cognitive components of the detection task are elaborated.

These results point toward a neural code for stimulus amplitude based on firing rate modulations, rather than on spike train modulations in periodicity. Consistent with this interpretation, modulations in the firing rate of VPL neurons within a 50-ms window essentially accounted for detection performance. We chose this window size because it captures modulations in firing rate elicited by single sinusoidal pulses, leaving out periodicity-related effects. However, individual neurons were slightly worse than the monkeys, which raises several questions: (i) What is the relevant integration time scale for VPL neurons? (ii) Why are individual VPL neurons slightly worse detectors than the monkeys? (iii) How do S1 neurons read the firing rate modulations of their VPL inputs to evoke stimulus perception in conjunction with other cortical areas (12)?

If we assume a code based exclusively on firing rate modulation, then a 50-ms integration window seems sufficient for an ideal observer to detect the stimulus. Actually, the window might be shorter, given that the response latency after stimulus onset ranges from 24 to 29 ms for S1 neurons and is ∼68 ms for secondary somatosensory (S2) neurons (12). Thus, after 30 ms, S1 neurons already have tactile information arriving from thalamic neurons, and after 70 ms, information is already in S2 neurons. In addition, thalamic neurons have been shown to adapt to sinusoidal stimuli (7, 15), and indeed, we also observed stronger rate modulations for the first stimulation pulse (Fig. 4 B and C). Thus, we believe that modulations in the firing rate elicited during the first pulse might be sufficient for the subject to detect the stimulus; however, further experiments are needed to confirm whether this is so.

The fact that single VPL neurons were slightly worse detectors than the monkeys, based on the firing rate elicited during the 50-ms time window, might be related to how their activity is read out downstream (12, 16). The neural code might not be exclusively rate-based, but may involve additional contributions, such as spike timing or synchronous activation from multiple VPL neurons. It has been shown that the activities from several convergent and synchronous thalamic inputs onto an S1 neuron are necessary to elicit an action potential (4). De La Rocha et al. (17) reported that the correlation between spike trains increases with the firing rate, suggesting that VPL neuronal synchrony could be modulated by stimulus amplitude. On the other hand, in anesthetized rats, low-velocity sinusoidal stimuli such as those used in the present study do not increase the level of synchrony compared to that in spontaneous activity (4, 18). Moreover, most previous studies investigating the synchrony of somatosensory thalamic neurons used anesthetized animals and ramps, deflections, or sinusoids of high amplitude as stimuli, often lasting more than 1 s (4, 6, 7). Thus, it is not clear how thalamic synchrony affects S1 when near-threshold stimuli of short duration, like those used here, are detected by a behaving subject. Further investigation involving simultaneous recordings of VPL neuronal populations and S1 neurons in awake, behaving animals are needed to answer this question.

It has been suggested that information flow through the sensory thalamus is subject to strong modulatory inputs coming from the cortex, brainstem, reticular thalamic nucleus, and interneurons (19). Thus, we wondered whether VPL neurons could be modulated by cortical activity related to perceptual decisions (13, 14). We found no differences between hits and misses or between CRs and FAs, as in S1 (2, 12). Moreover, the sensitivity to stimulus amplitude was the same during active task performance and when animals were not required to attend the stimuli or use them to obtain a reward. Although it is possible that an effect on VPL activity may be seen in a task in which (unlike our passive control task) the locus of attention is specifically directed (9), the negative result is still notable in light of previous observations that neurons from multiple cortical areas engaged during a tactile discrimination task display large changes in firing between active performance and passive stimulation, and may even stop encoding stimulus-related information altogether (20). This means that the state of the cortical network must be very different in the two conditions. In addition, the fact that cueing the monkey’s motor choice did not affect the activity of VPL neurons is significant, because part of the feedback to the sensory thalamus consists of axons that send branches to subcortical motor centers (21). Thus, the neural activity of VPL neurons with tactile receptive fields is associated mainly with a stable representation of stimulus features, and does not reflect the animal's internal state. This result is similar to findings reported by Wilke et al. (11) in the lateral geniculate nucleus during a visual awareness task.

To conclude, the present study demonstrates that relay neurons in the VPL nucleus modulate their firing rate as a function of stimulus amplitude, and that a neural code based on such rate modulation can account for overall detection performance (i.e., stimulus sensitivity). However, perceptual decisions do not affect the activity of VPL neurons, and their sensitivity to stimulus amplitude is the same regardless of behavioral output and task context. These results favor the hypothesis that, at least under conditions akin to those tested here, the somatosensory thalamus behaves as a relay station of sensory information to cortex that is rather insensitive to additional cognitive components. Such cognitive processes are likely to develop and to be integrated with sensory information in circuits downstream from the somatosensory thalamus (2, 12, 22).

Materials and Methods

Detection Task.

The sensory detection task used here has been described previously (2). Monkeys were handled in accordance with the institutional standards of the National Institutes of Health and the Society for Neuroscience. Protocols were approved by the Institutional Animal Care and Use Committee of the Instituto de Fisiología Celular.

Visual Instruction Task.

Trials proceeded exactly as described for the standard detection task with identical tactile stimuli, except that the correct response was assigned independently of the stimulus and was indicated by illuminating the corresponding button when the probe indented the skin (PD). The light was on during the whole trial until the probe was lifted up from the skin (PU). The monkey was rewarded for pressing the illuminated button. Arm movements were identical to those in the detection task.

Passive Stimulation.

The same set of stimuli was delivered to the fingertip, but detection was restricted by removing the key and the interrupt target switches. The animal remained alert because it was rewarded with drops of liquid at random times, but was no longer required to attend to the stimuli or react to them. Monkeys did not appear to make any attempt to respond with the free hand/arm in this condition.

Recording Sessions and Sites.

Neuronal recordings were obtained with an array of seven independent, movable microelectrodes (2-3 MΩ) (2) inserted into S1 medial to the hand representation in such a way that allowed us to lower the microelectrodes into the VPL nucleus. Recordings were performed contralateral to the stimulated hand. Each recording began with a mapping session to determine the cutaneous representation of the fingertips. Neurons from the VPL had small cutaneous receptive fields with QA or SA properties. Locations of electrode penetrations were confirmed in the VPL with standard histological techniques.

Data Analysis.

The responses of 74 VPL neurons collected while two monkeys performed the detection task were analyzed. All neurons had small cutaneous receptive fields located in the distal segment of one digit (fingertip 2, 3, 4, or 5). Stimuli were delivered to the center of the neuron’s receptive field. A response was considered task-related if the distribution of firing rates during the stimulation period was statistically different from that in a period immediately before trial initiation (α = 0.01, ROC analysis) (20). The relationship between the stimulus amplitude and the neuronal firing rate elicited during the stimulation period (500 ms) was quantified using linear regression (firing rate = slope × stimulus amplitude + basal firing rate). For further analyses, only slopes that were statistically different from 0 were taken into account (α = 0.05). The same analysis was conducted for spike train periodicity as a function of stimulus amplitude, with a mean power at 20 Hz during the stimulation period (20). In each trial, the power spectrum of the spike train evoked during the period of stimulation (512 ms) was calculated using the discrete Fourier transform (23). The following parameters were used: sampling frequency, 10 kHz; resolution, 1.22 Hz; range, 15–65 Hz; n = 2¹³. The DC component was ignored.

The response latency was defined as the first bin in which the firing rate reached a value equivalent to the mean prestimulus firing rate plus 2 SDs. In addition, the next two consecutive bins were required to have a significantly higher firing rate than the first bin (20). To calculate the mean firing rate, a peristimulus time histogram was constructed containing activity from 500 ms before to 500 ms after stimulus onset, using a bin size of 1 ms and a Gaussian filter with a 10-ms span for smoothing.

Neurometric curves based on firing rate were constructed as described previously (2). For each trial, the maximum firing rate was obtained in a 500-ms or 50-ms window that was displaced every 1 ms in the period between 1.5 s before stimulus onset until the probe was lifted (PU); the analogous period was used for stimulus-absent trials. Neurometric curves were constructed as the proportion of trials in which the maximum firing rate reached or surpassed a criterion level (2). For each neuron, this criterion was chosen to maximize the number of hits and CRs trials. Neurometric curves based on periodicity were constructed as the proportion of trials in which the power at 20 Hz reached or surpassed a criterion level (23). For each neuron, the criterion was set as the mean power plus 3 SD at 20 Hz for the stimulus-absent trials. Thus, for any given stimulus amplitude, the proportion of trials with a power value higher than the criterion determined the probability of detecting the stimulus. For stimulus-absent trials, CRs were calculated using the proportion of trials in which the power did not reach the criterion level.

Psychometric and neurometric fits were obtained by fitting a Boltzmann equation. Lower and upper Boltzmann parameters were fixed according to the psychometric and neurometric performance in each session. Only curves with statistically significant goodness of fit were considered. For each curve, the detection threshold was calculated as the stimulus amplitude at which performance reached 50% detection. The amplitude modulation of the curve (ΔA) was the difference in detection probability between the maximum and minimum amplitudes. This value accounted for the variation in monkey/neuron performance within the relevant interval. Slopes for individual and population curves were calculated within the interval in which the probability of detection ranged from 0.4 to 0.6. Psychometric slopes and ΔA values were compared with their neurometric counterparts using the Wilcoxon rank-sum test (α = 0.05).

To compare neural activity between hits and misses at near-threshold stimulus values, or between CR and FA trials, neural responses in all trials were sorted according to the behavioral response elicited by the monkey and normalized. Normalization was achieved by subtracting the mean firing rate in a prestimulus control window and dividing the result by the SD of the rate in that same window. Firing rate was calculated with a 250-ms sliding window displaced every 50 ms, from 1 s before stimulus onset to 1 s after the probe was lifted off from the skin (PU). In addition, in stimulus-absent trials, the maximum firing rates elicited during the FA trials were compared with the maximum rates elicited during the CR trials. Firing rates were calculated along the entire trial, from 500 ms after PD to 100 ms before PU, using different bin windows (100, 250, 500, and 1,000 ms). Significant differences were quantified with a permutation test (n = 1,200 permutations; α = 0.01).

Choice probability indices based on signal detection theory were calculated as in previous studies (2, 24). The normalized firing rate (z-score) distributions were compared for hits versus misses in near-threshold trials (Fig. 5B, Left) and for FAs versus CRs in stimulus-absent trials (Fig. 5B, Right). In the former case, the analysis was restricted to stimulus amplitudes (6–12 μm) eliciting 40–60% of hits; in the latter case, at least three FAs per session were required, and equal numbers of CR trials were chosen at random. To compute the choice probability indexes across time, firing rates were calculated within the same windows and periods used for the normalization. To determine when choice probability indexes deviated from 0.5, a permutation test was applied for each window (n = 1,200 permutations; α = 0.01). A permutation test for means of nonpaired data (n = 1,200 permutations; α = 0.05) was performed to compare the latencies and linear regression slopes of neurons recorded during the active detection task and passive stimulation.