| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Howard Hughes Medical Institute and Department of Neurobiology, Stanford University School of Medicine, Stanford, California 94305
Submitted 8 September 2003; accepted in final form 25 December 2003
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTSDISCUSSIONAPPENDIX: NORMALIZATION...ACKNOWLEDGMENTSREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONAPPENDIX: NORMALIZATION...ACKNOWLEDGMENTSREFERENCES |
|---|
; Gold and Shadlen 2001; Romo and Salinas 2001; Schall 2001). Neurophysiological investigations of decision mechanisms are typically conducted in behaving animals that are trained to perform sensory-discrimination tasks. Because the animal necessarily reports the outcome of the decision process with an operant movement of some sort, the neural signals underlying decision making can be closely related to, and may in some cases be identical with, "premotor" signals associated with preparation for a specific movement. In our own previous work, for example, rhesus monkeys were trained to discriminate between opposite directions of motion in a stochastic random-dot display and report the perceived direction with a saccadic eye movement to one of 2 targets aligned with the axis of motion (Horwitz and Newsome 1999, 2001a,b). Because the saccade targets were visible throughout the trial and bore a stereotyped geometric relationship to the axis of stimulus motion, the accumulation of sensory information favoring one or the other direction of motion could activate, nearly simultaneously, the premotor circuits that prepare one or the other eye movement (Gold and Shadlen 2000, 2003). Thus an abstract representation of the decision is not necessary to account for the behavior. To distinguish decision-related neural signals from purely motor signals in such paradigms, we and others have relied heavily on the intuition that decision-related activity should reflect to some extent the strength of the sensory evidence that led to the decision (i.e., the "certainty" of the decision), whereas strictly motor signals should be related more directly to the metrics of the operant movement irrespective of the evidence that led to the movement.
Using quantitative analytic techniques based on this intuition, we demonstrated the existence of decision-related signals in the superior colliculus (SC), a midbrain structure closely associated with the generation of eye and head movements. Using the same direction-discrimination task and similar analytic techniques, Shadlen and colleagues documented nearly identical decision-related signals in area LIP (lateral intraparietal) of the parietal lobe and area 46 of the frontal lobe, both of which have been implicated in oculomotor behavior (Kim and Shadlen 1998; Roitman and Shadlen 1998; Shadlen and Newsome 2001). A central question that has arisen from these studies is whether these neural structures contribute generally to the process of making perceptual decisions, or whether these structures are involved only because the logic of our "standard" direction-discrimination task enforces a stereotyped link between the direction of stimulus motion and the vector of the operant saccade.
To address this question, we trained 2 rhesus monkeys on a new version of our direction discrimination task—the "loose stimulus–response association task" (LSRA)—in which we broke the stereotyped link between motion direction and the vector of the operant saccade. The animal performed the same perceptual discrimination as in our previous studies and continued to report its perceptual decision with a saccadic eye movement. In the LSRA task, however, 
28 different operant saccades could be used to report a particular direction of motion, and the metrics of the required saccade were not specified until the end of the trial, well after the perceptual decision had been made. In this task we fully expected decision-related signals to be absent from the SC, probably being displaced to frontal lobe circuitry associated with "rule-based" performance (Miller et al. 2003). To our surprise, however, we found that a small population of neurons in the intermediate and deep layers of the SC exhibited modulations of activity that predicted the decision the monkey would make at the end of the trial. In each experiment this "predictive activity" was relegated to a specific region of the SC map of space—the spatial location toward which stimulus motion flowed. These observations, and the results of microstimulation experiments that we also report, suggest that the monkey remembers the direction of stimulus motion during each trial, at least in part, by referring the motion to the spatial framework of the oculomotor system.
A specific spatial framework that could be used for this purpose is an "object-centered" coordinate system in which a virtual object is considered to form a spatial link between the 2 saccade targets. The monkey would remember the direction of stimulus motion by referring it to the right or left side (for example) of the virtual object. Such object-centered representations have been demonstrated in the supplementary eye field (SEF) of the frontal cortex (Olson and Gettner 1999, 1995; Olson and Tremblay 2000; Tremblay et al. 2002), and our data are consistent with, but do not prove, the notion that object-centered coordinates are present in some SC neurons as well. To further investigate this possibility, we recorded from SEF neurons while one monkey performed the LSRA task. Confirming prior observations of Olsen and colleagues, SEF neurons exhibited predictive activity in a manner consistent with the existence of object-centered responses. Interestingly, predictive activity arose sooner in SEF, on average, than in the SC, suggesting that SEF may occupy an intermediate position in the processing chain that links sensory activity in occipital visual areas to the decision-related signals we have observed in the SC.
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONAPPENDIX: NORMALIZATION...ACKNOWLEDGMENTSREFERENCES |
|---|
Two rhesus monkeys (Macaca mulatta) served as subjects in these experiments. Before data collection, each monkey underwent a pair of surgical procedures that were performed under aseptic conditions and general anesthesia. In an initial surgery, the animal was implanted with a head-restraint device and scleral search coil (Judge et al. 1980). After several months of behavioral training, a second surgery was performed to implant a stainless steel recording cylinder that permitted microelectrode access to the SC. One monkey was implanted with an additional recording cylinder over the SEFs. All experimental procedures conformed to the standards established by the National Institutes of Health and were approved in advance by the Institutional Animal Care and Use Committee of Stanford University.
Visual stimuli
Monkeys discriminated the direction of coherent motion in a stochastic random-dot stimulus that has been used extensively in this laboratory (Britten et al. 1992, 1996; Salzman et al. 1992). Stimulus movies lasted for 1 s and were presented on a CRT monitor with a 60-Hz refresh rate. Each frame of each stimulus movie consisted of a field of white dots presented within a 7° diameter aperture. Motion was created by replotting, after a delay of 50 ms, a randomly selected 51.2% of the dots (called "signal" dots) at a displacement of 0.15° in a single direction. Signal dots therefore moved coherently in a specified direction at a speed of 3°/s. The remaining dots (the "noise" dots) were replotted in random locations and thus served to mask the motion direction of the signal dots. Monkeys were trained to discriminate between 2 opposed directions of coherent motion and to report their judgments by making a saccade to one of 2 visual targets (see Behavioral paradigm 1: loose stimulus–response association task below). The density of dots in the visual stimuli was 15 dots/(deg2 · s-1), but the apparent density of dots in the stimulus was much higher because of persistence in the visual system.
Electrophysiology
Single-unit recordings were accomplished by means of tungsten microelectrodes (Microprobe or Frederick Haer) and standard electrophysiological amplifiers (Bak Electronics). Action potentials were detected using a time–amplitude window discriminator (Bak Electronics), time stamped (1-ms resolution), and stored on a magnetic disk for subsequent analysis.
Electrical microstimulation was accomplished using a pulse generator (Bak Electronics or Master 8 CP) in series with an optical stimulus isolation device (Frederick Haer). Stimulation pulses were biphasic, cathodal pulse leading. Each pulse was 0.2 ms in duration with a 0.1-ms pause between pulses. Stimulation currents ranged from 10 to 50 µA, frequencies ranged from 200 to 500 Hz, and train durations ranged from 70 to 400 ms. A saccade was classified as stimulation evoked if it was initiated within 50 ms of the onset of the pulse train.
Behavioral paradigm 1: loose stimulus–response association task
Figure 1 illustrates the LSRA task. This task is very similar to the direction-discrimination task used previously by Newsome and colleagues (Horwitz and Newsome 1999; Shadlen et al. 1996) with one critical difference: in the LSRA task the monkey does not know the metrics of, and therefore cannot plan, the operant saccade until late in the trial. Decision-related activity during the visual stimulus and delay periods is therefore unlikely to result from planning of a specific operant saccade.
|
The monkey received a liquid reward for initiating a saccade within 500 ms of fixation point offset that terminated in a 5 x 5° electronic window surrounding the correct target. Saccades that terminated in a 5 x 5° window surrounding the distractor target were counted as incorrect choices. Fixation breaks, failures to initiate the saccade within 500 ms of fixation point offset, or saccades to locations outside either target window were aborted immediately without reward.
Target locations were defined on a grid as shown in Fig. 1C. Spacing between target locations was fixed at 5° along the axis of motion and varied, between experimental sessions, from 4 to 8° in the perpendicular dimension. A target pair was defined as 2 target locations that were adjacent along the dimension parallel to the motion axis. Each target was thus a member of 2 target pairs, one in which it lay to one side of its mate and one in which it lay to the other side. In the context of the task this means that a saccade to a given target location could report motion in either direction, depending on where the other target appeared.
To prevent inaccurate saccades from being misclassified as incorrect choices we constrained target positions in 2 ways. First, we never used target locations >22° eccentric and rarely >15° eccentric. Second, we never used target pairs for which the 2 operant saccades were very similar in direction to avoid choice biases resulting from the path of the saccade to one target from passing directly over the other target. In practice, this exclusion meant that no target pairs were used whose midpoints were within 45° of the motion axis, accounting for the gap between the 2 clusters of targets in Fig. 1C. Saccades to targets satisfying both of these constraints were extremely accurate; only 5% of the saccades that fell within one of the 2 target windows landed closer to the midpoint of the target pair than either individual target. The number of allowable target locations varied from 12 to 40 across experimental sessions.
For each neuron, we adjusted the display geometry so that coherent motion on each trial flowed toward or away from the movement field (dashed curve). Thus in the example in Fig. 1, the monkey discriminated rightward from leftward motion (arrows) because the movement field lay directly to the right of the fixation point. As indicated above, each member of a target pair could represent motion flowing toward or away from the cell's movement field. We will refer to the target that corresponded to motion toward the movement field as "T1" and the other target as "T2" regardless of their positions on the computer monitor.
In the design described thus far, a small but systematic association remains between the direction of stimulus motion and the metrics of the operant saccade. Each target location in Fig. 1C, for example, is the rightmost member of one pair and the leftmost member of another pair—with the exception of targets at the right and left edges of the array. A saccade to one of these edge targets always indicates that motion flowed in the direction of the target. Across an entire experiment, therefore, the collection of saccades that the monkey makes to report rightward motion would have a small rightward component on average. Conversely, saccades that indicate leftward motion would have a net leftward component. Thus delay period activity in the SC might occur for a trivial reason: if the monkey "knows" that the required saccade is likely to have a component in the direction of motion, it may prepare to make a saccade in that direction on viewing the motion stimulus. To eliminate this cue to the metrics of the upcoming operant saccade, we presented "edge" targets only when they were distractors. Thus the distribution of correct saccade vectors was identical for both directions of motion, and the monkey had no basis for anticipating the metrics of the correct operant saccade.
It is important to note that on most trials neither saccade target fell within the neuron's movement field (e.g., only 2 of 20 target locations in Fig. 1C). This procedure appears odd at first glance because SC neurons are commonly thought to respond only for saccades to targets within the movement field. We chose this procedure so that the results of the experiments could be compared directly to those of our previous studies in which coherent stimulus motion always flowed toward or away from the movement field (Horwitz and Newsome 1999, 2001a,b). Thus the geometrical relationship of the stimulus aperture, coherent motion flow, and movement field location were identical in all studies.
Training history
In designing the LSRA task, we sought to create a behavioral context in which the monkeys made directional judgments without translating those judgments immediately into preparation for a particular saccade. Ideally, the delay period before onset of the saccade targets should provide a "window" for observing the representation of the decision in a relatively abstract form, divorced from a specific operant response. For this reason, naïve monkeys were selected for training on this task; neither animal had been employed previously in studies that used the standard version of our direction-discrimination task. In the first stage of LSRA training, however, both monkeys were briefly trained to perform the standard task along a single axis of motion (i.e., to make saccades in the direction of visual motion). As soon as each animal performed this discrimination reliably (
70% correct; 1 mo for monkey A, 8 sessions for monkey T), the 2 targets were displaced away from the axis of stimulus motion during an extended period of "generalization" training. For monkey A, training continued for a single new target pair until performance exceeded approximately 80% correct choices, whereupon a new target pair was selected. Monkey T was trained on sets of 4 to 6 target pairs in short blocks until performance was reliable for all positions. Only after reliable performance was attained at a variety of target positions were multiple target pairs interleaved within a single block of trials. The number of target pairs increased progressively and the intertarget distances decreased over the course of training for both monkeys. For monkey A, the axis of stimulus motion was rotated relatively early during generalization training, while new target pairs were being introduced. For monkey T, the axis of motion was changed only during the last month of training after the monkey had fully generalized to all target pairs for one axis of motion. In all, training took 14 mo for monkey A and 7 mo for monkey T.
Behavioral paradigm 2: memory-guided saccade task
We generally screened cells on the basis of their discharge during a memory-guided saccade task. In this task, a single eccentric saccade target was displayed for 200 ms while the monkey fixated a central point. After a random-length delay period of 1–1.5 s, the fixation point was extinguished and the monkey was required to initiate a saccade to the remembered location of the target within 500 ms of fixation point offset. The monkey received a liquid reward for saccades that terminated in an electronic window, usually 10 x 10°, surrounding the target location. If the monkey failed to initiate the saccade within the stipulated time interval, or if the saccade terminated outside of the window, the trial was aborted. At least 2 target locations, one inside and one outside of the movement field, were pseudorandomly interleaved from trial to trial.
Data analysis
During single-unit recording, monkey T responded correctly to 88% of 7,072 LSRA trials, and monkey A responded correctly to 88% of 6,427 LSRA trials. Performance was consistent across sessions for both animals (the SD of percent correct was 6% for monkey A and 7% for monkey T). Unless specified otherwise in the text, we analyzed data from correctly performed trials only.
Predictive activity
We classified individual neurons as "choice-predictive" if the firing rate during the stimulus presentation and initial delay period was significantly higher preceding reports of one direction of motion than the other (Mann–Whitney U tests: P 0.01). The evolution of predictive activity across a population of neurons was assessed using a technique based in signal detection theory (Green and Swets 1966). For each choice-predictive neuron, we calculated responses preceding T1 and T2 choices in 100-ms bins. We normalized these responses by dividing each by the maximal response across all time bins and then pooled the normalized responses across neurons. For each bin, we calculated a receiver operating characteristic (ROC) curve from the distributions of normalized responses. The area beneath this curve was defined as the predictive activity value for that time bin. Related methods have been used to measure predictive activity in previous studies (Kim and Shadlen 1998; Krauzlis and Dill 2002; Shadlen and Newsome 2001; Thompson et al. 1996).
Cell screening
While advancing our electrode in search of a neuron, we typically had the monkey perform the memory-guided saccade task, or occasionally, the LSRA task. Each time we isolated a neuron from the intermediate or deep layers of the SC, we qualitatively assessed responses in both tasks. If the cell exhibited either maintained delay-period activity in the remembered saccade task or choice-predictive activity in the LSRA task, we recorded data for subsequent off-line analysis. We selected cells on the basis of the remembered saccade task for 2 reasons. First, at the outset of data collection, we did not know (or even expect) that we would find any choice-predictive neurons in the SC. It did not seem reasonable to select cells on the basis of a response we did not think existed. Second, we wanted to study roughly the same population of SC neurons that we had analyzed in our previous studies (Horwitz and Newsome 1999, 2001a,b). Our selection criterion in the previous studies correlated well with delay-period activity in the remembered saccade task. Saccade targets in the current study typically fell outside the SC movement field (Fig. 1C), so many cells that we encountered responded rarely or not at all during LSRA task performance. We quickly abandoned cells that failed to generate any responses during the LSRA task. Experiments were aborted if either electrical isolation or behavioral control was lost.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONAPPENDIX: NORMALIZATION...ACKNOWLEDGMENTSREFERENCES |
|---|
We recorded from 36 neurons from monkey A and 41 neurons from monkey T during performance of the LSRA task, selected according to the criteria described in METHODS (Cell screening). Firing rates were analyzed quantitatively off-line. Of the total sample, 17 neurons from monkey A and 18 neurons from monkey T were significantly more active preceding T1 choices (defined in METHODS) than T2 choices during the stimulus-presentation period and the initial delay period (Mann–Whitney U test: P 0.01). Recall that the saccade targets had not yet appeared during these epochs, and that most targets would ultimately fall outside the movement field. Thus the difference in firing rate preceding T1 and T2 saccades appears to reflect the monkey's perceptual decision per se, not preparation to report the decision with a particular saccade.
Two cells that were significantly more active preceding T2 choices than T1 choices were excluded from analysis. One cell whose response was clearly time-locked to small-amplitude saccades confined to the fixation window was likewise excluded.
The proportion of choice-predictive neurons in our data set (35/77) should not be taken as an estimate of the proportion of choice-predictive neurons in the SC as a whole. We deliberately restricted our recordings to a select population of SC neurons (see METHODS), and choice-predictive neurons are certainly less frequent in the SC as a whole than in our biased sample. Although we cannot provide a precise estimate, we believe that choice-predictive neurons in this task constitute roughly 10% of the neurons we isolated in the intermediate and deep layers of the SC.
Figure 2 illustrates the responses of a representative choice-predictive neuron on correctly answered trials. On trials in which the monkey chose T1, the firing rate rose steadily during the stimulus presentation period and the subsequent delay period. In contrast, the firing rate remained low throughout the trials on which the monkey chose T2. Thus the discharge of this neuron during the stimulus-presentation period and the delay period predicted the monkey's choice at the end of the trial despite complete uncertainty regarding the metrics of the correct operant saccade, which was randomized across 12 locations in both the contralateral and ipsilateral hemifields in this experiment (see following text).
|
Figure 3 shows the responses of the same neuron, segregated by perceptual choice and by target location, during 3 epochs of the trial (correct choices only). Left and right panels illustrate responses preceding T1 and T2 choices, respectively. The top panels depict the responses from 50 to 125 ms after the visual stimulus was extinguished. The firing rate of the cell was higher preceding T1 choices than T2 choices during this epoch (as noted in Fig. 2), and this relationship, of course, was independent of the (future) location of the targets.
|
The bottom panels illustrate the responses during a 75-ms epoch starting 400 ms before saccade initiation. Three aspects of the responses during this epoch are noteworthy. First, the response is considerably lower at this time than in the previous epoch, as indicated in Fig. 2. This is true preceding saccades to either T1 or T2 but is clearest preceding saccades to T1. Second, the response is only weakly related to the metrics of the impending saccade. Third, the firing rate is still significantly higher preceding T1 choices than T2 choices. Thus the signal carried by the cell at this relatively late point in the trial continues to reflect the perceptual decision made by the monkey and is inappropriate for specifying the metrics of the operant saccade.
Predictive activity: population analysis
To examine decision-related signals quantitatively, we normalized the responses for each choice-predictive neuron, pooled the data across the population, and performed the signal-detection analysis described in METHODS. We analyzed data from correct and error trials separately. Figure 4, A and B depicts the results of this analysis separately for the 2 monkeys. The development of predictive activity within the population of SC neurons followed closely the firing rate dynamics shown for the single example cell in Fig. 2. Predictive activity increased during the stimulus-presentation period and the delay period, peaked around the time of target presentation, and declined subsequently. The rightmost panels, aligned on the saccades, show that predictive activity was maintained through the time of saccade initiation. Comparison of A and B in Fig. 4 reveals that the time course of predictive activity was quite similar for the 2 monkeys, although the magnitude differed, reaching higher levels on average in monkey A.
|
Almost all (35 of 37) choice-predictive neurons preferred T1 choices to T2 choices. This demonstrates a close relationship between the choice preference of individual neurons and their movement field locations: choice-predictive SC neurons discharged more vigorously during trials in which the monkey reported visual motion flowing toward the movement field. We shall return to this point in the DISCUSSION.
Saccade-related discharge
The single-neuron data in Figs. 2 and 3, and the population data in Fig. 4, suggest that choice-predictive neurons in the SC carry signals related to the monkey's perceptual decision up to and including the time of the saccade, irrespective of target location. The question naturally arises whether these neurons ever exhibit the spatially localized movements fields that are typical of SC neurons.
Figure 5 illustrates the responses of the example cell of Fig. 2 during a perisaccadic epoch, defined as 50 ms preceding saccade initiation until 25 ms afterward. Responses are segregated by perceptual decision and target location as in Fig. 3. To determine whether these 2 factors significantly influenced perisaccadic firing rate, we performed a randomization test analogous to an ANOVA (Edgington 1995). Unlike a conventional ANOVA, this randomization test does not assume normally distributed firing rates. The data were permuted 10,000 times for each analysis.
|
Of the 17 choice-predictive neurons in monkey A, 14 exhibited significant spatial tuning at the time of the saccade and 13 were modulated by the perceptual decision during the same interval (randomization ANOVA, P < 0.01). Of the 18 choice predictive neurons in monkey T, 15 were spatially tuned and 3 were modulated by the decision. Thus at the time of the saccade the large majority of choice-predictive neurons indeed exhibited conventional spatially tuned movement fields, and roughly half of these neurons carried signals related to the perceptual decision as well. Neurons without significant spatial tuning may have had small movement fields that did not overlap with any saccade target in the array.
Saccade metrics
The effect of the perceptual decision on perisaccadic activity might be artifactual if the saccades to specific target locations varied systematically with the direction of stimulus motion that instructed the saccade. We therefore asked whether saccade endpoints varied with motion direction. Only experiments in which 
3 saccades for each direction of motion were made to 3 targets were considered in this analysis. For each target, we calculated the signed distance between the mean endpoints of T1 and T2 saccades, projected onto the axis of motion (positive values indicate displacements in the direction of motion, and negative values indicate displacements in the opposite direction). Under the null hypothesis, the signed distance should be distributed symmetrically around zero. One-sample t-tests (P < 0.05) showed that this was not the case for 16/35 experiments from monkey A and 13/33 experiments from monkey T. The mean endpoint displacement across targets and experiments was 0.15° for monkey A and 0.11° for monkey T in the direction of motion. Saccade endpoints were thus weakly biased in the direction of motion. We obtained qualitatively similar results irrespective of whether we subtracted initial eye position.
This small bias in saccade metrics clearly cannot account for the effect of choice on perisaccadic activity in Fig. 5 (20 spikes/s for several target locations), but it could in principle account for some of the small effects we observed. The randomization ANOVA described in the preceding section was blind to this dependency because firing rates were regressed on target location, not on precise saccade metrics. We therefore reanalyzed the effect of choice on perisaccadic activity using a second randomization test, which is similar to the ANOVA, with the important distinction that it takes precise saccade metrics into account.
For each cell, we fit 2 separate regressions, one for T1 choices and one for T2 choices, modeling perisaccadic firing rate as a third-order polynomial of saccade endpoint coordinates. The fitted equation was
 |
Finally, we considered the possibility that small saccades made within the fixation window during the stimulus presentation and delay periods were biased by the motion direction. Such a bias could produce artifactual predictive activity. A circular analog of the 2-sample unpaired t-test, the Mardia–Watson–Wheeler test (Batschelet 1981), revealed significant differences in fixational saccades (P < 0.05) in only 2/36 experiments in monkey A and 3/41 experiments in monkey T. The directions of fixational saccades thus do not appear to be strongly affected by the motion direction.
Electrical stimulation
We considered the possibility that predictive activity in the SC may be related to saccade planning even though it cannot, by virtue of the task design, be related to the planning of the operant saccade. For example, the monkeys may have planned saccades during the delay period that were congruent with the direction of motion during the stimulus period. Such saccade plans might even have served as a mnemonic device for remembering the direction of stimulus motion. This explanation accounts for the general pattern of selectivity we observed (T1 choice preference) because T1 choices were consistently associated with motion toward the movement field.
We used an electrical-stimulation technique to determine whether the monkeys formed saccade plans before target presentation. To a first approximation, saccades elicited by SC stimulation depend only on the position of the stimulating electrode within the collicular map (Robinson 1972). Several groups have shown, however, that the exact metrics of an electrically elicited saccade can be modified when the monkey simultaneously prepares a saccade to a different region of space (Gold and Shadlen 2001; Kustov and Robinson 1996; Sparks and Mays 1983). For example, electrical stimulation of an SC site that normally generates leftward saccades will produce a saccade with an upward component (in addition to the standard leftward component) if the monkey is preparing an upward saccade at the time of electrical stimulation, and a downward component if the monkey is preparing a downward saccade. If our monkeys consistently plan to make saccades in the direction of stimulus motion in the LSRA task, we thus expect systematic deviations in stimulation-evoked saccade vectors to reflect this fact. On the other hand, if the signals we have reported are not related to saccade planning, we expect stimulation-evoked saccades to be minimally affected by the direction of stimulus motion.
Figure 6 illustrates the sequence of events during these experiments. To ensure that the monkey would continue to perform the perceptual discrimination task, only one-ninth of the trials were interrupted with electrical stimulation. On these trials, the fixation point disappeared at the moment that the targets would normally have appeared, and 80 ms later a train of electrical stimulating pulses into the SC evoked a saccade. In half of the stimulation trials visual motion flowed in one direction; in the other half motion flowed in the opposite direction. The axis of motion was selected to be roughly orthogonal to the direction of the stimulation-evoked saccades. Rewards were given 200 ms after the onset of electrical stimulation on all stimulation trials and for correct decisions, as usual, on nonstimulated trials.
|
|
We performed this analysis for 14 stimulation experiments from monkey A and 23 from monkey T in which we obtained at least 10 stimulation-evoked saccades for each direction of visual motion. For each experiment, we conducted a randomization test to determine whether the separation between saccade endpoint centroids was statistically significant. Only 4 experiments yielded significant effects (P < 0.05). This number of significant effects is, itself, not significantly greater than the false-positive rate expected by chance. Thus any effect of predictive activity on saccade-generating circuitry in the SC was too weak to be measured within individual experiments.
We then performed a population analysis to gain additional statistical power. In this analysis we asked whether the direction of the stimulation effect agreed with the hypothesized effect of predictive activity on saccade preparation (as in Fig. 7), irrespective of whether the effect was significant in individual experiments. For each experiment, we arbitrarily defined one of the 2 motion directions as the "reference direction," the other being the "nonreference direction." We then calculated a vector from the centroid of saccade endpoints obtained from nonreference trials to that obtained from the reference trials. We will refer to the direction of this vector as the "direction of the stimulation effect." In the example of Fig. 7, upward was the reference direction, and the direction of the stimulation effect (the vector from the green centroid to the red centroid in Fig. 7B) was very close to the reference direction. Figure 8 plots the angular difference between the reference direction and the direction of the stimulation effect for each experiment for monkey A (8A) and for monkey T (8B). Under the null hypothesis that no relationship exists between the 2 quantities, these differences should be uniformly distributed about the circle. Visual inspection shows, however, that they tend to cluster near 0°, as we would expect if saccades were deviated in the direction of visual motion. Treating each data point as a vector of unit length, we calculated an average vector for each monkey. The length of each average vector exceeded the upper 95% confidence bound (dashed circle) computed from a Rayleigh test (Batschelet 1981), verifying that the data points are not distributed uniformly. The population effect illustrated in Fig. 8 is consistent with the saccade-planning hypothesis, but the individual effects were very small and rarely significant, raising doubt about their role in a hypothesized saccade-planning process. We therefore compared these data to similar microstimulation effects obtained when the monkey performed a delayed-saccade task—a task that explicitly encourages saccade planning. In the delayed-saccade task, a single saccade target appeared 400 ms after the monkey achieved fixation and remained visible throughout a 2- to 3-s delay period. The saccade target appeared at one of 2 locations arranged symmetrically about the fixation point. On 8 out of every 9 trials, the monkey obtained a reward by making a saccade to the target within 500 ms of fixation point offset. On the remaining one-ninth of the trials, the targets were extinguished along with the fixation point, and after an 80-ms pause, electrical stimulation was applied to elicit a saccadic eye movement.
|
To compare stimulation effect magnitudes, we sorted saccade endpoints by stimulus condition (motion direction in the LSRA task and target position in the delayed-saccade task), and calculated the centroid of each group of endpoints. Figure 9A is a scatterplot of intercentroid distances (in normalized units; see APPENDIX). Every point lies above the identity line, indicating that stimulation effects were consistently larger in the delayed-saccade task than in the LSRA task. We thus conclude that choice-predictive activity in the SC during the LSRA task does not constitute a saccade "plan," at least not in the standard sense that applies in the delayed-saccade task. We will return to this issue in the DISCUSSION.
|
Since the stimulation effects during LSRA task performance do not appear to reflect saccade plans, we must ask whether the effects are related to task performance at all. It is possible, for example, that the effects result solely from the presence of the visual stimulus at the point of fixation, perhaps related to very slow pursuit or optokinetic nystagmus responses elicited by the foveal motion stimulus. Figure 9B shows that the stimulation effects are in fact related to task performance. Here we plot the monkey's performance in each experiment (in percentage of correct choices) as a function of the size of the microstimulation effect. Across the database of microstimulation experiments, the monkey's performance varied from 71 to 100% correct, with an average of 89% correct. (These percentages indicate performance on nonstimulated trials only.) The size of the stimulation effect was positively correlated with task performance. The correlation is statistically significant when the data are pooled across the 2 monkeys (r = 0.49, P < 0.005), suggesting that the observed deviations in the metrics of stimulation-evoked saccades, although very small, are linked to LSRA task performance. When considered separately, the correlation was considerably more robust in monkey A (r = 0.68, P < 0.005) than in monkey T (r = 0.42, P < 0.05), consistent with the fact that choice-predictive activity, although present in both animals, was stronger in monkey A than in monkey T.
Stimulus-response compatibility affects saccade latency
The microstimulation experiments described above showed the monkeys' perceptual decisions influenced the precise metrics of saccades evoked with electrical stimulation. We now show that decisions were likewise manifest in the latencies of the operant saccades made to the targets on interleaved (nonstimulated) trials. Figure 10A illustrates the basic finding presented in the preceding sections: stimulus motion in a particular direction excited a population of SC neurons whose movement fields (dashed circles) lay in the direction of stimulus motion (arrows). For a given direction of motion, however, the correct operant saccade may be directionally compatible or incompatible depending on the location of the pair of saccade targets on any given trial. In Fig. 10A, for example, motion down and to the right, which elicited predictive activity in a population of neurons with movement fields near the dashed black circle, could be followed by an operant saccade in the same general direction (target locations C and D) or in the opposite direction (target locations A and B).
|
To assess this effect across a population of behavioral experiments, we performed a multiple regression analysis of each data set, modeling the latency of correct saccades as a weighted sum of target position, the direction of stimulus motion, and the interaction between these 2 variables. The interaction term is of particular interest because it captures differences in the slope of the relationship between target position and saccade latency that depend on the motion direction. The data in Fig. 10B clearly illustrate such a slope difference, and the interaction term for these data was indeed significant (P < 0.05), indicating that the slope difference is not attributable to random variability.
Figure 11 shows a scatterplot of the 2 regression slopes (e.g., the black and gray lines in Fig. 10B) obtained in each of the 86 experiments we performed. Data from monkey A appear in Fig.11A and data from monkey T appear in Fig. 11B. Experiments with significant interactions, which constitute a majority, are indicated with filled symbols. The relationship between slopes was generally consistent across sessions: motion flowing in the reference direction yielded negative slopes, whereas motion in the opposite direction yielded positive slopes. In other words, for either direction of motion, saccade latencies decreased as the target was presented with increasing eccentricity in the direction of motion. On average, a displacement of the target by 1° in the direction of motion decreased saccade latency by an average of 1.2 ms for monkey A and by 1.0 ms for monkey T. Thus on correct trials saccades in the direction of motion have shorter latencies than saccades in the opposite direction.
|
10 errors were made in each direction. This analysis revealed that saccade latencies were related to the direction judgments, not to the direction of motion in the stimulus. Thus for error trials saccade latency increased with displacement of the target in the direction of motion (1.1 ms/°). This effect was significant for monkey T (Wilcoxon test on regression slopes: P < 0.01) and nearly so for monkey A (P < 0.1). This effect is a further indication that the monkey's perceptual judgment influenced the state of the oculomotor system.
Predictive signals in SEF
Where do the predictive signals that we observed in the SC originate? The supplementary eye field (SEF) is a good candidate for providing such signals, given that Olson and colleagues have documented predictive activity in the SEF in a task that is similar to our LSRA task (Olson and Gettner 1999, 1995; Olson and Tremblay 2000; Tremblay et al. 2002). To explore this idea, we recorded from SEF neurons in monkey T while he performed the LSRA task.
The SEF was identified using conventional criteria: the medial surface of the frontal lobe from which saccadic eye movements could be elicited with electrical stimulation currents 60 µA (Mitz and Godschalk 1989). Neurons were isolated for study while the monkey performed either the LSRA task or the memory-guided saccade task. For each neuron isolated, the preferred saccade direction was estimated from response during a block of delayed-saccade trials or from the vector of saccades elicited with microstimulation. Each neuron was then tested for responsiveness in the LSRA task, first using an axis of dot motion aligned with the cell's estimated preferred direction, and second, with dot motion along an orthogonal axis.
Figure 12 illustrates predictive activity in a population of 25 choice-predictive SEF neurons (thick curve). We included only correct trials in this analysis because the animal generally performed the task very well, generating too few errors for analysis. For comparison, Fig. 12 also shows predictive activity from the population of 18 SC neurons from the same animal (thin curve). Early in the trial, SEF neurons were more predictive than SC neurons. The difference in predictive activity was statistically significant from 300 ms after stimulus onset until 300 ms into the delay period (randomization tests: P < 0.01; shaded area, Fig. 12). The population of SC neurons was never significantly more predictive than the population of SEF neurons. Nine of 23 SEF neurons tested also exhibited statistically significant delay activity in the memory-guided saccade task (Mann–Whitney U test, P < 0.05). In every case, there was reasonable agreement between preferred directions in the delayed-saccade task and in the LSRA task.
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONAPPENDIX: NORMALIZATION...ACKNOWLEDGMENTSREFERENCES |
|---|
, 2001a,b). In that paradigm the monkey could prepare the operant saccade nearly simultaneously with the perceptual accumulation of motion signals originating in the visual cortex. Thus in the standard task it is reasonable to expect decision-related signals to appear in oculomotor structures, and this is indeed the case (Gold and Shadlen 2000; Horwitz and Newsome 2001b; Kim and Shadlen 1998; Roitman and Shadlen 2002; Shadlen and Newsome 1996). In contrast to the standard task, the LSRA task used in the current study did not permit the monkey to know the metrics of the operant saccade during formation of the perceptual decision. We therefore expected that decision-related activity would be eliminated from the SC (and probably from other oculomotor structures as well) during performance of this task. To our surprise, however, a population of intermediate and deep layer SC neurons, as well as a population of SEF neurons, discharged in a choice-specific manner during the LSRA task. This decision-related discharge, or "predictive activity," began up to several seconds before the monkey could plan the operant saccade, allowing an experimenter to predict the monkey's perceptual report irrespective of the direction and amplitude of the operant saccade. Thus the activity of a subpopulation of SC and SEF neurons during this task appears to be related to the perceptual judgment per se, and not simply to the operant saccade.
Predictive activity rose steadily after the motion stimulus was extinguished, peaking around the time that the targets were illuminated. This observation is inconsistent with the idea that the SC buildup activity reflects the integration of motion information (Gold and Shadlen 2001), but rather points to more complex dynamics (e.g., Wang 2001). Remarkably, predictive activity persisted after the presentation of the 2 saccade targets and even extended through the time of the saccade itself for some neurons (Fig. 5). Thus even after the monkey was able to plan and execute the correct saccade, the discharge of some SC neurons was not strictly related to saccade metrics, but rather, was strikingly modulated by behavioral context (i.e., the perceptual decision).
A possible (uninteresting) explanation for our data is that choice-predictive activity reflects covert saccade planning. In this scenario, the monkey covertly prepares a saccade in the direction of stimulus motion, accounting for our observation of choice-predictive activity in an oculomotor structure. The saccade plan would then be changed to the correct operant saccade on appearance of the pair of targets. Three lines of evidence militate against this interpretation. First, the LSRA task was specifically designed to discourage the monkey from planning saccades in advance of the target presentation; at the time in the trial that predictive activity evolved, the monkey could not predict the metrics of the required saccade because the distribution of correct target locations was identical for both directions of stimulus motion (see METHODS). Second, the proportion of neurons that predicted choices in the LSRA task (about 10%) was considerably smaller than the proportion of neurons exhibiting predictive activity in our standard direction-discrimination task or exhibiting delay-period activity in a memory-guided saccade task (both about 30%). If a saccade plan can be thought of as a well-defined, unitary entity, we would expect similar numbers of neurons to contribute to a plan in all 3 tasks. Finally, stimulation-evoked saccades were influenced only weakly by the direction of stimulus motion in the LSRA task (Figs. 7, 8, 9A). In contrast, saccade plans formed during a delayed-saccade task exerted striking effects on the direction of stimulation-evoked saccades. If the monkey actively planned saccades in the LSRA task, we would expect stimulation-evoked saccade vectors to be affected similarly. Together, these observations build a strong case that decision-related activity in the LSRA task does not result from covert saccade planning.
Spatial representation of the perceptual decision
A priori, perceptual decisions made during the LSRA task might be represented in a completely "abstract" fashion, that is, unrelated to a particular motor system or to particular regions of visual space. For example, the monkey's decisions might be represented in categorical "rightward" or "leftward" neurons in a nontopographic region of the frontal cortex. That decision-related signals are present in a spatially organized structure like the SC raises the alternative possibility that decisions in the LSRA task are in fact coded with respect to space. This notion is supported by 3 different observations presented in this paper: electrophysiological, microstimulation, and behavioral. First, decision-related signals recorded electrophysiologically were systematically related to the spatial map in the SC: most choice-predictive SC neurons discharged most vigorously when the monkey reported motion flowing toward their movement fields. Thus stimulus motion and the accompanying decision selectively excited a topographic region of the SC lying in the direction of motion (i.e., stimulus motion "pointed" toward the neuron's spatial movement field). Second, small deviations in electrically evoked saccades were systematically related to the direction of stimulus motion. On trials in which rightward or leftward motion was presented, for example, the endpoints of evoked saccades were displaced slightly to the right and left, respectively, showing that decision-related signals influenced the oculomotor system in a spatially specific manner. This relationship was particularly strong during sessions of good performance (Fig. 9B). Finally, the behavioral effects of stimulus-response compatibility (Figs.
