| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1Center for Neuroscience and 2Section of Neurobiology, Physiology and Behavior, University of California, Davis, California 95616
Submitted 3 July 2003; accepted in final form 3 October 2003
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
). However, much of the motion experienced by organisms under normal conditions consists of more complex, spatially nonuniform motion, or "optic flow" (Gibson 1950).
Response selectivities for such complex patterns of motion first become evident in the medial superior temporal area (MST), a relatively high-level area in this cortical pathway. Neurons in MST show selective tuning for patterns such as expansion, contraction, rotation, or combinations thereof (Duffy and Wurtz 1991a,b; Saito et al. 1986; Tanaka and Saito 1989; Tanaka et al. 1986). Thus it is natural to consider MST as a likely substrate for the perception of optic flow, including heading stimuli— optic flow patterns that simulate trajectories experienced through self-navigation in the world (Bradley et al. 1996; Duffy and Wurtz 1995; Froehler 2002; Page and Duffy 1999; Paolini et al. 2000). Additionally, perturbing activity in MST with electrical microstimulation during the performance of a complex heading perception task (Britten and Van Wezel 2002; Britten and Van Wezel 1998) or a linear direction discrimination task (Celebrini and Newsome 1995) influences perceptual judgments. Together, these studies suggest that area MST is indeed involved in optic flow perception but do not elucidate the relationship between the activity of individual neurons and behavioral ability. To examine this issue, we recorded the responses of single neurons in MST of awake behaving monkeys while the monkeys performed a psychophysical discrimination of specified locally and globally opposite optic flow patterns that were optimized for each neuron's stimulus preferences. We were thus able to simultaneously measure both behavioral and neuronal sensitivity to optic flow patterns under conditions where the neuronal signal was most likely to be relevant to the perceptual judgment. These measurements allowed us to address directly whether the sensitivity of MST neurons for optic flow patterns was sufficient to support perceptual judgments of complex motion. In addition, this approach allowed us to investigate the possibility of correlation between neuronal response and behavioral choice on a trial-by-trial basis. Preliminary results of this study have previously appeared as abstracts (Heuer and Britten 1999, 2000).
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
Two adult rhesus macaques (Macaca mulatta) were used in this study: one male and one female. Prior to recording, each monkey was implanted with a stainless steel head restraint post and a scleral search coil (Judge et al. 1980) to monitor eye position. After each monkey completed training on the psychophysical task, a recording chamber was surgically placed over occipital cortex. For recording sessions, a stainless steel transdural guide tube was inserted at known locations within a plastic coordinate grid (Crist et al. 1988). A tungsten microelectrode (FHC) was introduced through the guide tube and advanced using a stepping motor microdrive (National Aperture).
We used both physiological and anatomical landmarks to localize area MST on the anterior bank of the superior temporal sulcus (STS). These landmarks included recording depth from the dural surface, gray matter/white matter transitions, and receptive field (RF) size (Tanaka et al. 1986; Van Essen et al. 1981). On a few penetrations near the floor of the STS, the transition from area MT to MST was ambiguous; for these penetrations, only cells with large RFs that clearly included the fovea or were entirely ipsilateral were included. Generally, these cells were found after an abrupt discontinuity in retinotopy; they account for a small proportion (approximately 5%) of our sample and did not noticeably differ in their results.
After we located MST, we isolated and recorded single-unit activity using standard extracellular techniques. Electrode signals were amplified and filtered: single units were isolated with a window-discriminator (Bak Electronics) and their action potentials converted to TTL pulses. We used the public domain software package REX (Hays et al. 1982) to record the time of stimulus events and action potentials with 1-ms resolution.
Stimuli
All stimuli were presented on a CRT monitor (Mitsubishi Diamond Pro 21TX) that subtended 80° horizontally by 60° vertically at a viewing distance of 28 cm from the monkey. Pixel resolution was 1280 x 1024, with a vertical refresh rate of 72 Hz, corresponding to a frame interval of 13.9 ms. The stimuli were generated using custom software on a Pentium computer with a video card (ATI Mach 64) running in 8-bit grayscale mode. For the primary experiments described here, the background luminance was 10 cd/M2, and the foreground was 60 cd/M2, resulting in a Michelson contrast of 70%. The monitor was regularly calibrated to establish a linear luminance profile. Dots were presented within circular apertures, at a constant average density of 0.12 dots/deg2. Dots were randomly generated, and each stimulus was generated with a new seed.
The random dot patterns employed consisted of a subset of optic flow patterns termed "spiral space" by Graziano et al. (1994). Spiral space can be formalized as a pair of orthogonal axes, with radial motion on one axis and rotary motion on the other (Fig. 1A). Intermediate directions are spirals— combinations of rotation and radial motion, such as an expanding counter-clockwise pattern. To vary the strength of these stimuli, we used a technique previously applied to linear direction stimuli (Britten et al. 1992; Celebrini and Newsome 1994). This technique varies the stimulus strength by changing the percentage of dots that are replotted in a manner consistent with the specified direction pattern. We varied this percentage, and thus the motion strength, by setting the probability that each dot would be repositioned in this manner. We refer to the proportion of dots that carry the specified motion pattern as the percent coherence. A stimulus that is fully coherent has all the dots replotted with the appropriate spatial and temporal offsets to produce an apparently smooth motion pattern. The opposite extreme, the 0% coherence stimulus, contains only dots that are randomly repositioned, forming "white" motion noise (see Fig. 1B for schematic examples). The speed of the stimulus dots increased linearly from the center of the stimulus to the edge.
|
Once a neuron was isolated, we determined the RF location and size using handheld moving bar stimuli or computer-generated moving dot patches. Many neurons were only weakly responsive to the handheld moving bar stimuli; for these, we used dot patterns of varying size to estimate the receptive field boundaries. Some neurons showed sensitivity to eye position within the orbit, as previously reported (Squatrito and Maioli 1996, 1997); for these cells, we attempted to maximize response due to both eye position and stimulus extent. We positioned the stimuli within the RF for maximal response. When a cell's RF exhibited a "hot spot" of greater excitability, we placed stimuli over the hot spot and may not have filled the entire RF. For the remaining cells, we placed the stimuli to maximize areal extent within the RF without extending past the RF edges. Once we determined stimulus size and location, we undertook parametric optimization of the stimulus. We initially approximated both preferred spiral space direction and preferred speed qualitatively.
Following the initial estimates of location and speed, we obtained quantitative direction tuning data. We presented eight directions of spiral space motion, equally spaced at 45° intervals, and eight directions of linear motion. The direction tuning stimuli were presented as highly coherent (75%) dot patterns; 100% coherent patterns were not used to minimize dot density artifacts. Generally, trials consisted of three stimuli of 500-ms duration, with 500-ms intervals between stimuli. For some cells, trials consisting of two stimuli of 1-s duration were used; there was no significant difference between the two measurements for cells where data were collected in both ways. The different directions of linear and spiral space were presented in a random sequence until responses were collected for a minimum of five repetitions of each motion pattern. We then fit these data with Gaussian functions to obtain a measure of the neuron's preferred direction. Such fits have previously been shown to describe accurately the direction tuning of MST cells to this family of stimuli (Graziano et al. 1994).
Psychophysical discrimination task
The primary data described here were collected while the monkeys performed a psychophysical discrimination between opposite spiral space directions. For each session, the directions for the behavioral discrimination were set to be the neuron's preferred stimulus and its opposite (or "null"). These directions were determined from the Gaussian fit to the tuning data (see Fig. 4A for example). The direction of the stimulus was randomly varied from trial to trial between the opposed alternatives; the monkey's task was to accurately identify which pattern was presented. The strength of the motion signal was also varied systematically using log-spaced coherence levels chosen to span both neuronal and behavioral thresholds.
|
|
All descriptive fits of our data were performed using a maximum-likelihood, iterative fitter (Chandler 1965), and nested likelihood ratio analysis was used to determine the comparative quality of the resulting fits. We used this approach when the equations fit were variants of the same function with differing numbers of free parameters (Hoel et al. 1971). We transformed the likelihoods obtained from each fit by
 |
is distributed approximately as 
2, where the degrees of freedom are the additional number of free parameters in the expanded function. If exceeded the critical value for the degrees of freedom, we concluded that the expanded version provided a significantly better account of the data. Analysis of psychophysical data
Behavioral data were analyzed as a function of motion coherence (%). For each level of motion strength, we calculated the proportion of correct choices, collapsed across the two directions of motion to form a psychometric function. We fit these data with a modified Quick (1974) function
 |
indicates the monkey's threshold, or the motion strength level at which the monkey reliably gets 82.5% correct. 
indicates the slope of the function, or how sharply the monkey's performance improves as a function of coherence. This analysis method has been described previously in greater detail in the context of a linear direction discrimination task (Britten et al. 1992; Celebrini and Newsome 1994). Analysis of physiological data
The primary experiment compared neuronal sensitivity to perceptual sensitivity for spiral space motion patterns. To provide this direct comparison, we analyzed our physiological data to produce "neurometric functions" analogous to the psychometric functions used to describe the behavioral data. This method has been used in a variety of contexts (Bradley et al. 1987; Britten et al. 1992; Celebrini and Newsome 1994; DeAngelis and Uka 2003). Effectively, this method allows a calculation of the discriminability of direction based solely on neuronal firing rate.
To generate the neurometric function for each cell, we calculated the neuron's responses to individual stimulus presentations. Unless otherwise specified, all neuronal responses were analyzed as the total spike count for the duration of the visual stimulus, shifted 50 ms to account for neuronal latency. We required 
10 stimulus repetitions for each condition (range, 10-45; mean, 17). For each coherence (motion strength), we compiled two distributions of these responses: one for stimuli in the preferred spiral space direction and a separate distribution for anti-preferred (null) stimuli. To compare these distributions, we assumed the existence of a neuron with equal and opposite tuning in spiral space [the "anti-neuron" (Britten et al. 1992)]. The response distributions were then assigned to neuron and anti-neuron according to preferred direction, allowing us to evaluate the discrimination performance for each stimulus level using receiver operator characteristic (ROC) analysis (Green and Swets 1966). We were then able to fit these ROC values with the same equation as our psychophysical data (Eq. 2) to obtain a neurometric function that was directly comparable to the monkey's simultaneously measured psychometric function.
One might wonder about the choice of an opponent formulation for the neurometric function, given that there is little evidence for direct opponency between neurons tuned for opposite directions of optic flow. We believe that this choice is sensible, because the opponency in the model has more to do with the readout of the representation. In our task, two opposite directions are being compared. Thus a sensible decision rule would be to compare the weight of evidence in favor of each alternative to form a decision (Gold and Shadlen 2002). While the representation of complex motion is distinctly biased in favor of expansion (Graziano et al. 1994), this does not effect the outcome in any fundamental way. Pools representing each alternative direction may differ in size, which would affect the performance of a population-based analysis. However, for the opponent analysis of single neuron signals, the presence of all directions in the population representation is sufficient.
We measured the neurometric functions of 112 neurons in two monkeys. Seven of these cells were discarded from analysis due to poor behavioral performance during the psychophysical task; these were cases where the monkey's psychometric function was nonmonotonic or did not reach 95% correct at the highest coherence tested. Due to the poor behavioral performance, the comparison between neuronal and psychophysical sensitivity was invalid for these seven cases.
Of the remaining 105 cases, an additional 15 cells were excluded because their neurometric functions were not significantly better fit by the Quick function than by a constant value (nested log-likelihood test; P < 0.05, df = 3). These cells were ones that responded significantly above baseline to spiral space stimuli during the direction tuning series but were only weakly directional even at the high coherence used for the tuning measurements. This weak directionality did not produce a significant neurometric function across motion strengths.
Calculation of choice probability
To examine the relationship between neuronal activity and behavioral choice on a trial-by-trial basis, we computed a "choice probability" for each cell.
For each cell, we created two distributions of neuronal response sorted by the monkey's choice. Only conditions where the monkey made at least three choices in each direction were included. To avoid stimulus-dependent effects due to different firing rate means for each condition, we Z-transformed all spike counts before combining the data across conditions. Before combining the data in this way, we verified that there was no systematic variation in choice probability as a function of motion strength (data not shown). This allowed us to calculate a single pair of distributions of Z-transformed firing rates for preferred and null choices for each cell. These distributions were used to calculate the choice probability by ROC analysis.
To test the significance of observed choice probability values, we used a bootstrap method. For the bootstrap test, we randomly reassigned each trial's Z-score to the preferred or null distribution. This dissociated the response from the behavioral choice without distorting the relative proportion of preferred and null choices. We then calculated the choice probability for the permuted distributions; we repeated this manipulation for 2,500 iterations, allowing us to estimate the probability the observed value would occur by chance.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
Perceptually, the monkeys were sensitive to the spiral space stimuli used in these experiments, although thresholds differed significantly between the two monkeys. Monkey H had thresholds that were significantly higher than those of monkey F; average thresholds for the two monkeys were 14.5 and 5.9% coherence, respectively. Thresholds were measured for two human observers (1 naïve), who were instructed to use the full global motion pattern to perform the task, as opposed to local motion features. For central stimuli, thresholds for the two observers were 4.5 and 6.7%, respectively. Thus one of our monkeys performed similarly as did human observers on our task, while the other did not.
The threshold difference between the two monkeys could not be accounted for by differing stimulus parameter distributions for each monkey. Parameters for each experimental session were matched to the stimulus preferences of the neuron under study and therefore could have differed between the two monkeys. The primary stimulus parameters that varied across experiments were the preferred direction, speed, size, and eccentricity of the stimuli. In our analysis, we considered whether these parameters varied between the experiments conducted on each monkey and whether they influenced threshold measurements. While some stimulus parameters were related to psychophysically measured thresholds, none of these relationships was able to explain the observed threshold differences.
Neither stimulus direction nor speed was capable of explaining the monkey threshold differences. There was no correlation between preferred direction (as determined for the cell under study) and the monkey's threshold for that session (linear - circular correlation; Mardia and Jupp 2000; P > 0.05). In contrast, faster stimulus speeds systematically were correlated with higher behavioral thresholds (r = 0.332; P = 0.0015), although this correlation was not significant for each monkey's data considered individually (r = 0.191 and 0.236; P > 0.1). This correlation between speed and psychophysical performance did not explain the differences between the two monkeys' thresholds because speeds were similar for both monkeys (t = -1.451, P > 0.1).
Stimulus size and eccentricity showed a more complicated relationship with threshold; this relationship differed between the monkeys (Fig. 3). While the average stimulus size (diameter) differed between the two monkeys (t = -2.104, P = 0.038), this parameter only affected the behavioral thresholds of monkey F. Monkey F showed a significant correlation between stimulus diameter and threshold (r = -0.327; P = 0.0149), but monkey H did not (r = 0.320; P = 0.0605). On the other hand, eccentricity was sampled similarly between the two monkeys, and again only one monkey (H) showed a significant relationship between eccentricity and threshold (r = 0.367, P = 0.0301). Across the data from both monkeys, stimulus eccentricity had a significant effect (r = 0.309, P = 0.003) on thresholds, but the stimulus size did not. Thus individual stimulus parameters did not systematically appear to affect the thresholds in a way that would explain the observed substantial threshold difference. Individual thresholds are known to vary widely on many tasks, of course, both for monkeys and humans. As we will describe later, we found that neuronal thresholds did not differ between the animals as much as did their perceptual thresholds.
|
Neuronal response properties
The neurons in our sample were heterogeneous, yet typical of MST in their response properties for spiral space stimuli. In Fig. 4A, we show an example cell's responses to eight directions of spiral space along with the axis used for the behavioral discrimination. The Gaussian fit to these responses was used to calculate a standard directionality index (DI) for each cell; the distribution of these DIs is shown in Fig. 4B. The filled bars depict the 15/105 cases that were excluded from further analyses due to flat neurometric functions. As seen in Fig. 4C, the distribution of preferred spiral space directions peaked near 90° (expansion). This anisotropy is consistent with previous reports (Geesaman and Andersen 1996; Graziano et al. 1994; Saito et al. 1986; Tanaka and Saito 1989; Tanaka et al. 1989).
Comparison of perceptual and neuronal sensitivity
Neuronal responses for preferred and null direction spiral space patterns diverged as a function of motion strength (Fig. 5), suggesting that individual neurons could be quite sensitive to these motion patterns. However, our primary question was whether the sensitivity of neurons in MST is sufficient to underlie perceptual sensitivity. To address this, we used ROC analysis to calculate a "neurometric" function for each cell as described in M[SCAP]ETHODS. This allowed a direct comparison of behavioral and neuronal sensitivity using the same metric for both.
|
|
|
Both monkeys showed a wide range of threshold ratios, indicating that the relationship between neuronal and perceptual sensitivity is diverse in our population of MST cells. Observed ratios for monkey F ranged between 0.14 and 13.95, while those for monkey H were between 0.16 and 4.2. The maximum possible ratio for monkey H was lower due to the higher psychophysical thresholds, described earlier. For monkey F, 22/55 (40%) cells were as sensitive as or more sensitive than the simultaneously measured psychophysical performance; for monkey H, the percentage was much higher (24/35; 69%). Thus we see that while the threshold ratios were quantitatively different for the two monkeys, there was nearly complete overlap as well. As we will show below, the only factor contributing to this difference in threshold ratios is the lower perceptual thresholds for monkey F.
Unlike the thresholds, the slopes of neurometric and psychometric functions were consistently quite similar. When we compared neurometric and psychometric slopes by measuring the ratio of the parameters of the two functions, the mean slope ratio was 0.98 across the population. The distributions of slope ratios are shown in Fig. 7B. There was a tendency for the slope ratio to be lower for monkey H. However, this tendency did not produce a significant difference between monkeys (t = 1.639, P > 0.05); mean slope ratios were 1.12 and 0.86 for monkeys F and H, respectively.
For both monkeys, approximately one-quarter of all cases (15/55 for monkey F; 9/35 for monkey H) showed a significant improvement when fit with two functions differing only in their slopes. Each monkey had a substantial overlap between these cells and the group that were better fit when the threshold was free (Fig. 7C). For the cases that fell into the overlap category, we could evaluate which parameter better captured the data by comparing the residual errors. Because these fits had the same number of free parameters, the fit that better captured the data had a smaller residual error. For the vast majority of these cases, allowing to vary captured the data better. Thus even for the neurometric functions that differed from the psychometric functions in both parameters, the larger difference was in the thresholds, indicating that the primary discrepancies between neurometric and psychometric functions reflect differences in sensitivity.
The difference in threshold ratios between the two monkeys could primarily be attributed to the difference in psychophysical performance. Neuronal thresholds for both monkeys varied, and the mean neuronal sensitivity did not differ between the two monkeys (t = -1.775; P > 0.05). Therefore we can conclude that the two monkeys differed far more in their performance than could be accounted for by differences in the sensitivities of their MST cells.
Although the ratios between neuronal and psychophysical threshold were diverse, there was a significant relationship between the two measures of sensitivity (P = 0.006; Spearman rank-order correlation). When the data from each monkey was considered individually, the relationship was only significant for one monkey, monkey H (P = 0.0011). The relationship between neuronal and perceptual performance was steep (Fig. 8), consistent with the fact that there was much more variance in neuronal than perceptual thresholds.
|
Temporal dynamics of neuronal responses
Our primary analysis comparing neuronal and behavioral discrimination used the spike counts from the entire stimulus duration; however, this long integration window could obscure subtleties in the neuronal signal and might affect our main results. The neuronal signal of direction could either accumulate slowly or appear rapidly near the beginning of the trial. As seen in the single cell example (Fig. 5), spike rates diverged early in the trial, suggesting the latter model. We examined the dynamics of neuronal performance as a function of coherence using higher time resolution. Because neuronal thresholds did not differ between the two monkeys, this analysis was applied to the entire sample of 90 cells. For each cell, we divided the trial into 50-ms time bins and calculated the ROC value in each time window for each motion strength level.
The neuronal signals appeared quickly; for all coherence levels, the ROC value arose early (within the 1st 200 ms of the trial) and was relatively flat throughout the trial, as seen for our sample of 90 cells in Fig. 9. At the highest coherences, the ROC value initially rose more steeply but plateaued at a slightly lower level for the remainder of the stimulus duration. Not unexpectedly, the maximal value achieved for the highest coherence was <1.0; the smaller time windows reduced the number of spikes available for the calculation, and the range of different absolute sensitivities further reduced the average sensitivity. The temporal dynamics observed suggest that sensitive information was available early in the trial.
|
n) for each cell as a function of integration time (Fig. 10A). Note that n sharply declined over the first 400-500 ms of the trial and then leveled out to a gradual improvement in performance as additional integration time was allowed. This flattening occurred throughout the population and was not driven solely by the less sensitive cells. To quantify this threshold asymptote as a function of time, we calculated, for each cell, the ratio between thresholds early in the integration period (400 ms) and late (the full 1-s duration). The distribution of these ratios is shown in Fig. 10B; the mean ratio is 1.26. This shows that, on average, the neuronal threshold only improved by approximately 25% with an additional 600 ms of integration time. Some cells showed a slight increase in threshold with the longer integration times, probably due to more transient responses.
|
n approached a plateau around 400 ms of integration time. The drop in ROC value at later time points was due to the difference in firing rate between preferred and null responses decreasing over the course of the trial, as shown for one motion strength level (16% coherence; Fig. 11). Figure 11A shows the average firing rates for preferred and null directions; all rates were normalized to the maximal response for the preferred direction for the individual neuron, which could occur in any time bin. Figure 11B shows the decline in the difference between the null and preferred directions and revealed that the difference did plateau in the later stages of the trial.
|
There were modest, but systematic, departures of the neuronal thresholds from the simulation, indicating that the spiking statistics departed from purely Poisson predictions in two ways (Fig. 12A). First, the observed thresholds dropped more rapidly in the early portion of the trial than did the simulated thresholds. Second, the observed thresholds remained higher at the end of the trial than the simulation predicted. To quantify this departure from Poisson statistics, we calculated the average Fano factor (variance: mean ratio) across all cells and all conditions, as a function of time within the trial. Time windows early in the trial had more reliable spiking statistics (Fano factors < 1), while late time epochs showed more variable firing (Fig. 12B); the average Fano factors were significantly correlated with time (r = 0.831; P < 0.0001). This was consistent with the departures of the data from the simulation and indicates that responses early in the trial may be more reliably informative about the stimulus.
|
In many tasks for which neuronal sensitivity and perceptual sensitivity are quite similar, a signature of neuronal involvement in the task is found in the "choice probability" (Britten et al. 1996; Celebrini and Newsome 1994; Dodd et al. 2001; Parker et al. 2002). This metric captures the amount of trial-by-trial correlation between the response of a sensory neuron and the subsequent choice of the monkey. Given the correlation between neuronal and perceptual thresholds in this study, it seemed likely that a choice probability would exist in our data as well.
We explored the choice probability for our population of MST cells and were surprised to find little relationship between choice and neuronal activity on a trial-by-trial basis. We computed a single choice probability for each neuron by combining across all stimulus conditions where the monkey made at least three decisions in each direction (for details, see METHODS). Choice probability values are bounded between 0 and 1, with 0.5 representing chance (no correlation). Choice probabilities over 0.5 reveal that the neuron's response is higher for trials when monkey chooses the target representing its preferred stimulus than for null choice trials with identical stimuli.
In contrast to previous studies, we did not find a systematic correlation between firing rate and behavioral decision. For our sample of 90 neurons, the average choice probability was 0.513 (not significant, P > 0.05, 2-tailed t-test; Fig. 13). We used a permutation test to determine significance for the individual choice probabilities. The choice probability was significant for 20% (18/90) of cells; for one-half of these, the effect was backward, indicating an anti-correlation.
|
Our observed choice probabilities were not related to neuronal response properties. There was no trend for more sensitive neurons to exhibit higher choice probabilities, as previously seen (e.g., Fig. 8, Celebrini and Newsome 1994). Similarly, there was no correlation between choice probability and threshold ratio. Intuitively, it would seem likely that cells with sensitivities most similar to the behavioral sensitivity would have a higher choice probability, but we did not observe this relationship in our data. We also examined the relationships between choice probability and dynamic range, preferred direction, and direction selectivity of the neuron under study. None of these response properties showed a significant correlation with choice probability. This suggests that the lack of significant choice probability in our data set was not due to the inclusion of insensitive or irrelevant neuronal signals.
These choice probabilities were calculated with the spike counts from the full stimulus duration, which could obscure subtleties in the time course of neuronal firing. Therefore we also investigated choice probabilities as a function of time, using the same 50-ms bins as we did for the sensitivity analysis. Previous investigations that reported significant choice probabilities in the population found that the signal accumulated early and remained constant for the duration of the trial (e.g., Britten et al. 1996). However, given that much of the observed sensitivity accumulated within the first 500 ms of the trial, it was possible that choice probabilities would be higher during the early portion of the trial, reflecting the most potent neuronal signal. Alternatively, significant choice probabilities could occur late in the trial, perhaps reflecting the upcoming decision. Neither of these possibilities was seen clearly in the data (data not shown); choice probabilities were noisy but generally flat across the trial duration. This suggests that the lack of significant choice probabilities was not a result of the time window used for analysis; choice-related activity was not consistently seen in any part of the trial.
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
Complex motion analysis in MST
The neuronal sensitivities we observed support the idea that MST can provide sufficient information for perceptual analysis of optic flow. However, this does not necessarily imply that MST is specialized for optic flow processing, nor that the processing optic flow must occur only within MST. Instead, we believe that the wide range of relative sensitivities we obtained is likely to be a common observation at the higher levels of cortical hierarchies and may reflect the multitude of sensory processes supported by activity within MST.
While many studies have focused on the role of MST in visual motion perception, neurons in MST also respond to extraretinal signals, including eye position within the orbit (Squatrito and Maioli 1997), ocular following or pursuit eye movements (Kawano et al. 1994; Newsome et al. 1988), and vestibular input (Bremmer et al. 1999; Thier and Erickson 1992). Additionally, many neurons show selective responses to combinations of retinal motion image and extraretinal signals (Bradley et al. 1996; Duffy 1998; Upadhyay et al. 2000). Individual neurons in MST may be quite sensitive for one type of signal while being responsive but insensitive to another. This would result in a wide range of sensitivities when only a single stimulus type is explored and is consistent with our observations.
The diversity of sensory signals within MST need not imply that the representation of optic flow is "sparse," with only a small fraction of neurons actively contributing at any one time. Instead, large numbers of neurons are responsive to optic flow, with widely varying sensitivities relative to the monkey. Under these circumstances, two different pooling and decision rules are possible if the monkey is relying on the signals from MST for its perceptual judgment. First, the monkey could be using only a portion of the active neurons—those that have sensitivities close to his own. In both of our monkeys, these well-matched neurons are not the most sensitive ones, which exceed behavioral capabilities. Thus this rule would conform poorly to the "lower envelope principle" (Mountcastle et al. 1972). Alternately, the monkey could pool broadly over the population and use signal averaging to achieve higher sensitivity. The latter model is broadly consistent with our observations.
Missing choice probability
In several related experiments in the motion pathway, significant choice probabilities have been interpreted as a signature of neuronal involvement in the perceptual task. One study in particular is quite comparable to ours. Celebrini and Newsome (1994) measured both sensitivity and choice probability for MST neurons in the context of a linear direction-discrimination task. Their analysis of neuronal sensitivity revealed a pattern of results not very different from ours. Neurons were slightly more sensitive in that context, relative to the monkey, but again showed a range of sensitivities. However, choice probabilities were much larger in their study, suggesting an association between area MST and linear motion discriminations.
This difference in choice probability for two related tasks is puzzling. There are two main differences between the two tasks. The most conspicuous one lies in the stimuli: complex versus translational motion. The other difference is more subtle— our monkeys were given "matching" saccade targets at the end of the trial. These targets contained motion in the neuron's preferred and null direction. We do not believe that this latter task difference explains the lack of significant choice probabilities in our study. Significant choice probabilities have been seen during the sample period of a delayed match-to-sample task in area MT (T. Pasternak, personal communication), indicating that matching tasks do not eliminate the trial-by-trial correlation between neuronal activity and behavioral choice. Additionally, after initial inspection of the targets at the beginning of the block, the monkeys could use the position of the target for their response; target locations were constant throughout a block of trials. It seems probable that the monkeys switched to such a position strategy early in each block, since their response latencies were very short. Therefore it seems unlikely that the monkeys were performing a true matching task. Given this, we conclude that the difference in choice probabilities reflects a qualitatively different involvement in tasks requiring the analysis of complex motion.
The weak or nonexistent choice probability in our data also appears to contradict the results of microstimulation in MST during a heading discrimination task (Britten and Van Wezel 2002). Microstimulation influenced perceptual judgments of simulated heading, and those effects have been interpreted to support MST's involvement in complex motion perception. However, a number of significant effects in the microstimulation were backwards from the expected direction, suggesting that the relationship between MST activity and perception was not straightforward. Second, the horizontal headings used in those experiments contained substantial translational flow. As demonstrated by Celebrini and Newsome (1995), microstimulation in MST also affected linear discriminations. The micro-stimulation effects observed in the heading task may be partially due to the strong translational components in the stimuli. Taken together, these results suggest that the relationship between neuronal activity in MST and complex motion perception may be weaker than the relationship for linear motion analysis.
Interanimal differences
A curious feature of our data was that, while neuronal thresholds were on average similar for the two monkeys, behavioral thresholds were not. Monkey H had higher behavioral thresholds, resulting in lower threshold ratios. The differences between monkeys may be due to differing perceptual abilities or behavioral control issues. In three cases, monkey H had unusually high thresholds; these were sessions when the neuron's response properties dictated stimulus parameters that were less commonly encountered during training. For the remaining cases, monkey H's performance was consistent with thresholds obtained during the final stages of training, which were asymptotic against time. This indicates that the observed thresholds were an accurate reflection of perceptual ability. We also confirmed that monkey H was not strictly using local cues by testing with a foveally centered stimulus that should have encouraged such a strategy. Further support was provided by the slopes of the psychometric functions, which will be lower for monkeys inadequately motivated or performing poorly. Slopes were similar between the two monkeys and fell into an acceptable range. Therefore the observed differences between the two monkeys cannot be attributed to differing behavioral control, and presumably, they reflect differing perceptual abilities.
Although the perceptual thresholds differed between the two animals, the neuronal thresholds did not. This contrasts with previous observations for a linear direction discrimination task in area MT (Britten et al. 1992), where both neuronal and psychophysical thresholds varied between animals in a correlated manner—animals with more sensitive behavioral thresholds also had more sensitive neurons. The difference in perceptual thresholds for our task, unaccounted for by neuronal threshold differences, strongly suggests that other areas contribute to performance and might even limit it. The additional influences on performance may lie in downstream decision mechanisms or else in the contributions of other sensory representations of optic flow information. It could be argued that the decoupling between neuronal and perceptual thresholds we observe would be consistent with MST being completely uninvolved in this task. While this is possible, it would require that perception was not using a source of sufficient signals for performance, and we find this idea less than parsimonious.
Optic flow signals in other cortical areas
Although MST clearly contains suitable neuronal signals for optic flow processing, our results suggest that optic flow perception is likely to be based on activity in multiple cortical areas, including the ventral parietal area (VIP) and parietal area 7a. Both of these areas contain neurons with similar selectivities for spiral space motion as those in MST (Schaafsma and Duysens 1996; Siegel and Read 1997), and are also good candidates for the neuronal substrate of optic flow perception. Additionally, the anterior portion of the superior temporal polysensory area (STPa) contains neurons that respond selectively to optic flow components, although they appear to be more restricted to the cardinal axes of radial and rotary motion (Anderson and Siegel 1999). Perception of complex motions patterns may depend on activity in all four areas, either hierarchically or in parallel. MST projects to VIP, 7a, and STPa (Andersen et al. 1990; Baizer et al. 1991; Boussaoud et al. 1990), suggesting sequential processing. Intuitively, this would suggest that sensitivities in one of these other areas might be greater and that neurons in these areas might exhibit higher choice probabilities than we observed in MST. Alternatively, if signals from multiple cortical areas are pooled for the task, choice probabilities would be reduced due to lesser interneuronal correlation across the areas (Shadlen et al. 1996) Therefore the most sensible conclusion to draw from our results is that MST contributes to complex motion perception, but it is likely to do so in concert with other cortical areas.
|
ACKNOWLEDGMENTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
GRANTS
This work was supported by National Institutes of Health Grants EY-10562 and EY-12576 to K. H. Britten and MH-11700 to H. W. Heuer.
|
FOOTNOTES |
|---|
Address for reprint requests and other correspondence: K. H. Britten, Center for Neuroscience, Univ. of California, 1544 Newton Ct., Davis, CA 95616 (E-mail: khbritten{at}ucdavis.edu).
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONACKNOWLEDGMENTSREFERENCES |
|---|
Andersen R, Snowden R, Treue S, and Graziano M. Hierarchical processing of motion in the visual cortex of monkey. In: Cold Spring Harbor Symposia on Quantitative Biology, Vol. LV, The Brain. Plain View, NY: Cold Spring Harbor Laboratory Press, 1990, p. 741-748.
Baizer JS, Ungerleider LG, and Desimone R. Organization of visual inputs to inferior temporal and posterior parietal cortex in macaques. J Neurosci 11: 168-190, 1991.[Abstract]
Boussaoud D, Ungerleider LG, and Desimone R. Pathways for motion analysis: cortical connections of the medial superior temporal and fundus of the superior temporal visual areas in the macaque. J Comp Neurol 296: 462-495, 1990.[CrossRef][ISI][Medline]
Bradley A, Skottun BC, Ohzawa I, Sclar G, and Freeman RD. Visual orientation and spatial frequency discrimination: a comparison of single cells and behavior. J Neurophysiol 57: 755-772, 1987.
Bradley DC, Maxwell M, Andersen RA, Banks MS, and Shenoy KV. Mechanisms of heading perception in primate visual cortex. Science 273: 1544-1547, 1996.[Abstract]
Bremmer F, Kubischik M, Pekel M, Lappe M, and Hoffmann KP. Linear vestibular self-motion signals in monkey medial superior temporal area. Ann NY Acad Sci 871: 272-281, 1999.[CrossRef][ISI][Medline]
Britten KH, Newsome WT, Shadlen MN, Celebrini S, and Movshon JA. A relationship between behavioral choice and the visual responses of neurons in macaque MT. Vis Neurosci 13: 87-100, 1996.[ISI][Medline]
Britten KH, Shadlen MN, Newsome WT, and Movshon JA. The analysis of visual motion: a comparison of neuronal and psychophysical performance. J Neurosci 12: 4745-4765, 1992.[Abstract]
Britten KH and Van Wezel RJA. Electrical microstimulation of cortical area MST biases heading perception in monkeys. Nature Neurosci 1: 1-5, 1998.[CrossRef][ISI][Medline]
Britten KH and Van Wezel RJA. Area MST and heading perception in Macaque monkeys. Cereb Cortex 12: 692-701, 2002.
Celebrini S and Newsome WT. Neuronal and psychophysical sensitivity to motion signals in extrastriate area MST of the macaque monkey. J Neurosci 14: 4109-4124, 1994.[Abstract]
Celebrini S and Newsome WT. Microstimulation of extrastriate area MST influences performance on a direction discrimination task. J Neurophysiol 73: 437-448, 1995.
Chandler JP. STEPIT: University of Indiana Quantum Chemistry Program Exchange, 1965.
Crist CF, Yamasaki DSG, Komatsu H, and Wurtz RH. A grid system and a microsyringe for single cell recording. J Neurosci Methods 26: 117-122, 1988.[CrossRef][ISI][Medline]
DeAngelis GC and Uka T. Coding of horizontal disparity and velocity by MT neurons in the alert macaque. J Neurophysiol 89: 1094-1111, 2003.
Desimone R and Ungerleider LG. Multiple visual areas in the caudal superior temporal sulcus of the macaque. J Comp Neurol 248: 164-189, 1986.[CrossRef][ISI][Medline]
Dodd JV, Krug K, Cumming BG, and Parker AJ. Perceptually bistable 3-D figures evoke high choice probabilities in cortical area MT. J Neurosci 21: 4809-4821, 2001.
, population median (0.82). Solid bars, cells excluded from further analysis; median DI for included cells was 0.91. C: rose diagram of preferred directions for cells that passed inclusion criteria (n = 90). Each bin represents 15° in spiral space; size of wedge indicates number of cells whose preferred direction falls within that 15°. No correlation was found between DI and the preferred direction.
