Sequence learning in two-dimensional smooth pursuit eye movements in humans

Abstract

Sequence learning is common to all motor systems and is an essential aspect of human behavior necessary for the acquisition of motor skill. Many previous studies have demonstrated the ability to observe, store, and repeat sequences in a variety of modalities resulting in reduced reaction time. Recently, it has been found that subjects can make predictive smooth eye movements to a sequence of discrete horizontal target motions (C. J. Collins & Barnes, 2005). The present study extends that paradigm into two dimensions of motion in order to investigate qualitative and quantitative differences in sequences of vertical (V) and horizontal (H) eye movements. The subjects performed sequences of four discrete velocity ramps repeated either four or eight times in succession. Baseline measurements were obtained to discrete individual smooth pursuit velocity ramps to H and V predictable (PRD) and randomized (RND) targets. We found that subjects could rapidly learn and anticipate individual components of a four-ramp sequence in two dimensions. The results showed clear asymmetries in the eye movements made to horizontal and vertical targets. We found that the latencies to H targets were shorter than latencies to V targets in both the PRD and RND conditions. We also found higher initial eye velocity (50 ms after target onset) to H targets than vertical targets during the PRD condition. Because these differences in H and V eye movements are present in both RND and PRD trials, this suggests that the observed differences are not due to retention of information but are inherent asymmetries within the system.

Introduction

Learning and memory provide us with the fundamental ability to acquire new motor skills essential for development. Without this ability, all motor actions would employ reactive stereotypical behavior that inherently involves neural time delays. We would be unable to hit or catch a moving ball, and navigation in a world of moving objects could become very hazardous. The human motor system has devised ways of storing spatial and temporal information about a motor event, and subsequently utilizing this information to initiate an appropriate motor response prior to the stimulus presentation (Barnes & Asselman, 1991; Kao & Morrow, 1994; Kowler & Steinman, 1979). This early initiation avoids inherent neural delays within our motor processing system (Kowler & Steinman, 1979; Ohashi & Barnes, 1996). Previous studies have confirmed that during smooth pursuit eye movements, subjects can use timing cues to actively predict the onset of the next target when target motion is repeated (Barnes & Donelan, 1999). 

In addition to using cues for prediction, many motor skills also involve the concatenation of individual motor movements in a sequence, a process known as “sequence learning.” Sequence learning has previously been extensively studied in serial reaction time tasks, which clearly demonstrate a reduction in reaction time with increasing number of sequence presentations (Nissen & Bullemer, 1987). Other studies have shown that saccadic eye movements made to a familiar set of sequences also exhibit reduced onset latencies (Gaymard, Pierrot-Deseilligny, & Rivaud, 1990; Petit et al., 1996). This reduction in reaction time is thought to be brought about by subjects becoming explicitly aware of the temporal structure of a sequence (Nissen & Bullemer, 1987). Once this is achieved, subjects can actively predict future target motion based on prior stored information about the individual components of the sequence. This type of motor prediction is believed to involve a network of brain areas including the cerebellum, the basal ganglia, and frontal and temporal cortical brain areas (Penhune & Doyon, 2002). 

Previous studies in eye movements have found asymmetries between horizontal and vertical directions of motion, specifically in the gain of the responses for both saccades and smooth pursuit (Baloh, Yee, Honrubia, & Jacobson, 1988; Collewijn & Tamminga, 1984; Rottach et al., 1996), although no previous study has specifically analyzed vertical step-ramp stimuli or sequences of discrete anticipatory smooth pursuit eye movements to two-dimensional (2D) targets. It has consequently been suggested that these directions may act independently of each other, and thus may hold differing motor mechanisms (Churchland, Gardner, Chou, Priebe, & Lisberger, 2003). 

Recently, a visuomotor sequence learning task based on a series of smooth pursuit eye movements was developed for targets moving in the horizontal direction (Barnes & Schmid, 2002; Collins & Barnes, 2005). Collins and Barnes (2005) revealed that subjects could anticipate and scale appropriately each ramp component within a predictable sequence of horizontal target motion. Barnes and Schmid (2002) demonstrated that the concatenation and summation of the anticipatory responses to sequential target ramps allowed subjects to predictively track complex motion stimuli. The present study has used this paradigm and extended it to include (a) both horizontal and vertical sequence components, (b) randomized sequences, (c) increases in the numbers of repetitions, and (d) equivalent control measures of predictive and random smooth pursuit. This variety of tasks has allowed us to address commonalities and differences in these directions of motion and to investigate the differences between initiating pursuit within a sequence learning paradigm, compared with single repeated presentations. 

Based on previous findings from learning sequences of horizontal motion, we hypothesize that subjects should show an equivalent ability to anticipate and scale targets to a 2D sequence as for a single repeated velocity ramp. We also anticipate that increasing the number of presentations will result in an improved predictive response by displaying a modified latency, greater eye velocity, or both at target onset (Kowler & Steinman, 1981). 

Methods

Subjects

The authors received local ethical committee approval to conduct this research. All subjects acknowledged the declaration of Helsinki and gave informed consent to participate in the study. Subjects were healthy volunteers, with normal or corrected to normal vision, and had no known psychological or neurological disorders. Six subjects (mean age = 31 ± 7.2 years, four female) took part in the study and five subjects were naive to the experimental paradigm. Three subjects had variable levels of experience in performing eye movements under laboratory conditions, with three subjects having very little or no previous experience in eye movement experiments. 

Apparatus

Subjects sat, with their heads restrained, 1.5 m in front of a white screen in a dark room. Both the fixation cue and visual pursuit target consisted of a circle of red light emitting diodes (LEDs) projected onto a screen, subtending a diameter of 1.2° on the eye, and controlled by a mirror galvanometer. Eye movements were recorded using a video-based infrared pupil-tracking system (Chronos, Skalar Medical, BV). Data were sampled every 5 ms (200 Hz) and stored for later off line analysis. For the calibration, subjects tracked a target moving in a circle around the center of fixation at 0.4 Hz with a radius of 15°, prior to each stimulus block. 

Stimulus protocols

A variation of the remembered pursuit paradigm described previously (Barnes & Donelan, 1999) was used to elicit smooth eye movements in both the horizontal (H) and vertical (V) directions of motion. During each presentation, the fixation cue first appeared in the center of the screen for 80 ms before being extinguished. This fixation cue was followed by a temporal gap (200–600 ms) before the target began to move eccentrically from the center of the screen. The first 200 ms of motion was occluded; thus, when the target appeared it was displaced in the direction of target motion and then continued to move at a constant velocity for 650 ms. The introduction of a gap is known to reduce the latency of both predictable (Boman & Hotson, 1988) and randomized stimuli (Knox, 1996). The occlusion of the target in the initial 200 ms allows for a larger time window in which to observe anticipatory behavior (Collins & Barnes, 2006). Subjects performed the following experiments in four different sessions with each session lasting approximately 40 minutes. 

Control paradigm (CON)

Control trials consisted of blocks of eight presentations with the timing between successive presentations randomized between 1000 and 2000 ms. The target could move in one of four directions [left (l), right (r), up (u), or down (d)] and at two speeds (10°/s or 20°/s), resulting in eight conditions in total. The stimulus was presented in predictable (PRD) and randomized (RND) blocks. In the PRD block, the same stimulus was repeated eight times consecutively, with the timings of the visual cue, gap, and target duration all remaining constant. In the RND block, the duration and velocity of the stimulus, the duration of the fixation cue and the duration of the gap were all varied randomly between the eight presentations (see Figure 2a). Each subject performed eight PRD and eight RND blocks in total. This resulted in eight repetitions of each of the eight conditions (u10, u20, d10, d20, r10, r20, l10, and l20). 

Sequence learning four repeats (4×SEQ)

Each sequence comprised four individual presentations in which the target moved centrifugally in one of the eight conditions mentioned previously (u10, u20, d10, d20, r10, r20, l10, and l20). The fixation cue flashed twice in the center of the screen at the end of each sequence to signal that the sequence was over. This was necessary for subjects to identify the end of one sequence and the beginning of the next. The stimulus was presented in PRD and RND blocks, with each block comprising four sequence repetitions. In the PRD block, the same sequence was repeated four times consecutively, with the duration of the visual cue (80 ms), gap (400 ms), occlusion (200 ms), and ramp duration (650 ms) all remaining constant (see Figure 2b). In the RND block, the duration of the gap (200–600 ms), the direction, and the speed were all varied within and between each of the four sequence presentations (see Figure 2c). Each subject performed 16 different PRD sequences and 16 RND sequences with four ramp presentations within each, resulting in 128 presentations. This resulted in a total of four repeats to each of the eight velocity conditions. The sequences were pseudorandomly constructed ensuring all sequences comprised at least one horizontal and vertical component and that no sequence was comprised of all the same velocity targets. 

Sequence learning eight repeats (8×SEQ)

As above, each sequence comprised four individual presentations from the eight possible conditions. This paradigm used only PRD blocks in which the sequence was repeated eight times. Each subject performed four PRD blocks resulting in 16 repetitions of each velocity condition. No RND condition was required for these trials as the visual stimulus, and thus the response would be the same as for the RND 4×SEQ condition. 

Data analysis

All data were recorded at 200 Hz and analyzed off-line using custom-made Matlab software. Fast-phase intrusions were removed using threshold acceleration (750°/s ²) and velocity (40°/s) criteria, using a technique described previously (Bennett & Barnes, 2003). The gaps formed from saccadic removal were bridged using linear interpolation and were visually inspected for accurate time and position. The removal of blinks was not necessary as they did not interfere with the eye movement response and were isolated to the end of the presentations. Eye velocity was then low-pass filtered with a zero-phase digital filter at a cut-off frequency of 30 Hz. 

V0, V50, and V100

Smooth eye velocity at target illumination onset (V0), 50 ms after target onset (V50), and 100 ms after target onset (V100) were used as measures of anticipatory eye velocity. V0 and V50 represent the eye movement response prior to any significant visual feedback. An example of a response to an individual presentation is shown in Figure 1, for the purpose of demonstrating how the peak eye velocity and latency were acquired (see below). 

Figure 1

 

(Above) The diagram above shows an example of the method used for extracting the peak velocity and latency from each velocity profile. The abscissa (time = 0 ms) indicates target onset, the blue line shows eye displacement, the green line eye velocity, and the red dot marks peak velocity. The linear regression fitted to the eye velocity profile is shown in red and the point at which this crosses the y-axis is then subtracted from target onset giving the latency value in ms. This single example is taken from the PRD 4×SEQ block in Figure 2b.

Peak eye velocity (Pk)

Peak smooth eye velocity was calculated for the first peak of each response and then marked on the target and eye velocity profiles (see Figure 1). Visual inspection was again used to ensure a correct selection, and if the wrong peak was identified a new time window could be selected around the appropriate peak to correct this error. Note that the peak velocity was derived from a velocity signal that was effectively averaged by the filtering described above. 

Latency

Latency values were calculated by identifying 10% of peak velocity and performing a linear regression from this point back to the intercept y = 0 (see Figure 1). This regression was then superimposed on the actual eye velocity and position responses to ensure a good fit. If the line was not a good fit or the wrong starting point was chosen, the user could use 20% of peak velocity and regress back from this new point in order to avoid making an inappropriate assessment of initiation time (see Figure 1). Latency was found by subtracting time at eye movement onset from the time at which the target became visible. 

Gain

Gain of the eye movement to the target movement was calculated by dividing Pk by target velocity. 

Acceleration

Initial eye acceleration was found at time of target onset (V0 to V100) in PRD trials by calculating the change in eye velocity and dividing it by time (100 ms). 

Statistical analysis

All data were tested for normality and were found to be normally distributed. A repeated measures ANOVA was then performed on the data to identify significant differences between target velocity, direction, or condition. For CON paradigms, the last six responses (three to eight) to the stimulus were subsequently used for both PRD and RND trials to generate the mean and standard deviation, based on previous findings (Collins & Barnes, 2005). In the 4×SEQ trials, data for the third and fourth presentations of the sequence were averaged for each block in the PRD and RND trials in order to equate numbers in subsequent analysis. In the 8×SEQ trials, the last six presentations of the sequence (three to eight) were averaged for subsequent analysis. This resulted in a pool of 24 responses to each condition for the CON and 8×SEQ paradigms, and 32 responses for the 4×SEQ paradigm (of which 24 were randomly selected for subsequent statistical analysis). A repeated measures ANOVA was also performed to identify significant differences between paradigms (CON, 4×SEQ, and 8×SEQ). A priori contrasts were also carried out to identify individual differences between paradigms. 

Regression analysis

The mean V50 value for each subject and for each condition (10°/s or 20°/s in four directions) and to each trial type (CON, 4×SEQ, and 8×SEQ) was plotted against the mean latency for each of the conditions mentioned above. We then carried out a regression analysis on these data to demonstrate the relationship between V50 and latency. 

Results

Single subject response

Figure 2 shows typical examples of an individual subject making smooth eye movements to the PRD and RND, CON, and 4×SEQ conditions. Note that the subject starts making anticipatory eye movements from the second or third presentation of the stimulus in the PRD CON ( Figure 2a, left side) and by the second presentation of the PRD 4×SEQ condition ( Figure 2b). Eye movements to the randomized targets start later and are not anticipatory (see Figure 2a, right side and Figure 2c). 

Figure 2

 

(Below) Examples of eye and target displacement data from a single subject is plotted against time. Panel a shows eye and target displacement for both a PRD (left) and RND (right) block for the CON condition. The vertical target and eye displacement are shown in black and red respectively and the horizontal target and eye displacement are shown in dark grey and blue respectively. An example of a PRD 4×SEQ block is shown in panel b, and the RND 4×SEQ in panel c. The dashed grey lines indicate the end of a sequence in the 4×SEQ paradigm and the end of a block in the CON condition.

Control trials (V50)

The control condition was developed in order to acquire baseline measures of smooth pursuit eye movements to H and V target motions using the paradigm mentioned above. Mean V50 values for each subject to the CON trials are presented in Figure 3 in response to PRD and RND targets moving in each of the four directions at either 10°/s or 20°/s. Each differently colored line represents the eye movement responses from a single subject (S1 to S6). For PRD trials, eye velocity increases as target velocity increases for all subjects ( p < .005): this trend is present in both the mean (grey squares) and individual responses (colored lines with symbols), demonstrating a robust finding both within and across subjects. 

Figure 3

 

The diagram above shows mean eye velocity 50 ms after target illumination (V50) for each of the six subjects (S1–S6) to the control (CON) experiment. The experienced subjects are shown as triangles and the nonexperienced subjects as circles. V50 to the vertical (V) target motion is shown on the left-hand side of the graph and to the horizontal target motion (H) on the right-hand side, for each target velocity. The mean of all the subjects to PRD targets is shown as a grey square and to RND targets as black squares.

There is, however, a difference in the magnitude of the response for H (right) and V (left) targets. V50 is higher for H than for V target directions, an effect that was found to be highly significant ( p < .005). This asymmetry appears more pronounced in the downward and leftward directions (negative values) of the target motion. By contrast, the RND trials (black squares) show negligible V50 values in all directions, with no significant differences between the directions of motion or between subjects. We also measured eye velocity (V50) orthogonal to the actual target motion and found no significant differences between the PRD V50 and RND V50 in all subjects ( p < .001). This shows that there is negligible carry-over of prior vertical responses onto horizontal responses and vice versa. 

4×SEQ trials (v50)

In the 4×SEQ RND condition, subjects responded after visual feedback of the target motion had been acquired, whereas for the PRD paradigm subjects were able to make anticipatory eye movements to each ramp component by the second presentation of the four-ramp sequence. Figure 4 shows V50 values for individual subjects (colored lines) and mean V50 values (grey squares) in response to each target velocity for a sequence of four ramps repeated four times in a row. A pattern similar to the results in CON trial above was found, but with some key differences. ANOVAs revealed higher V50 values to the higher target speed ( p < .000); however, the ANOVA indicated that V50 values were significantly reduced compared with the CON condition ( p < .05). We also found a significant difference between horizontal and vertical eye velocity ( p < .001) for the 4×SEQ condition, as found for the CON paradigm. Again, the RND condition shows negligible values of V50 in the 4×SEQ condition, suggesting that subjects are not anticipating the target in the RND condition. We also examined V50 in the direction orthogonal to target motion and compared PRD and RND responses. We found no significant differences for any condition for any of the subjects. Orthogonal V50 was negligible in all cases. 

Figure 4

 

Mean eye velocity 50 ms after target illumination (V50) for each of the six subjects to the sequence learning paradigm repeated four times (4×SEQ). V50 values are means of repetitions three and four. As in Figure 3, the experienced subjects are shown as triangles and naive subjects as circles. V50 to the vertical (V) target motion is shown on the left-hand side of the graph and to the horizontal target motion (H) on the right-hand side, for each target velocity. The PRD mean of all the subjects is shown as a grey square and RND mean as black squares.

8×SEQ trials (V50)

The V50 results for the 4×SEQ trial were not as high as expected. To assess whether this was due to insufficient repetitions, we conducted further trials in which the sequence was repeated eight times. Subjects again began making predictive eye movements to each target in a sequence of four ramps by the second presentation of the sequence. Figure 5 displays the results for each subject to each of the target velocities (eight conditions). We revealed an increase of eye velocity (V50) with target velocity as shown previously in the CON and 4×SEQ trials. We found subjects show significantly lower V50 values for the 8×SEQ than the CON condition ( p < .05). However, the 8×SEQ condition does not significantly differ from the 4×SEQ condition. Significant asymmetries for the H and V directions of motion were found for all subjects ( p < .005) with V V50 values having lower magnitudes than H V50. V50 in the orthogonal direction revealed no significant difference between conditions in any subject. Finally, orthogonal V50 was compared to the 4×SEQ RND V50 and no significant difference was observed in any subjects. 

Figure 5

 

Mean eye velocity 50 ms after target illumination (V50) for each of the six subjects to the sequence learning paradigm repeated eight times (8×SEQ). V50 values are means of repetitions 3 to 8. As before, the experienced subjects are shown as triangles and other subjects as circles. V50 to the vertical (V) target motion is shown on the left-hand side of the graph and to the horizontal target motion (H) on the right-hand side, for each target velocity. The mean is shown as a grey square.

V0, V50, and V100

We found an incremental increase in eye velocity toward actual target velocity between target onset and 100 ms later as revealed by values of V0, V50 and V100 (see Figure 6). For the predictable trials, subjects were achieving ∼50% of target velocity by 100 ms after target onset (V100). This finding is supported by the latency results reported later and implies continuing acceleration in the first 100 ms of PRD responses as found previously (Barnes & Donelan, 1999; Kao & Morrow, 1994). In support of this, we also found significant differences (p < .001) between the velocities taken at 50 ms time intervals (V0, V50, and V100). During the RND condition, we found negligible values of V0 and V50. However, in the RND CON condition, we found small eye movements toward target velocity by V100 that were not observed at V100 during the sequence learning (4×SEQ) trials. A regression analysis at the time of target onset was also performed on the initial eye acceleration (from V0 to V100) in PRD trials. We found no correlation between initial eye acceleration and latency. 

Figure 6

 

Mean eye velocity for all subjects at target illumination (V0), 50 ms after target illumination (V50) and 100 ms after target illumination (V100) for both predictable (red) and randomized conditions (blue) to each target velocity.

Latency from target onset

Latency was measured from time of target illumination to time of eye movement onset (see Figure 7). The mean (± SD) latency for Subjects 1 to 6 to the RND condition in the CON trials was 85 (±23) ms, 79 (±28) ms, 92 (±28) ms, 90 (±32) ms, 101 (±28) ms, and 104 (±19) ms, respectively. By contrast, RND latency to the 4×SEQ trial was significantly longer for all subjects at 108 (±24) ms, 101 (±31) ms, 121 (±46) ms, 115 (±51) ms, 110 (±37) ms, and 115 (±34) ms, respectively. In both the RND CON and RND 4×SEQ trials, the latency to V targets was significantly longer ( p < .005) than to H targets (see Figure 7). In contrast to the RND trials, the PRD trials reveal a negative latency; that is, they all start prior to target illumination onset. The PRD responses also reveal a significant difference in latency between the V target directions and the H target directions in both the sequence learning and control trials ( p < .005), the H responses starting earlier. This asymmetry between V and H latency was more pronounced in the PRD conditions than in the RND trials. We found a significant difference between the sequence trials (4×SEQ and 8×SEQ) and the CON condition ( p < .02). No significant difference was found between the 4×SEQ and 8×SEQ latencies. 

Figure 7

 

Latency of eye movement onset to both predictable (PRD) and randomized (RND) trials for the vertical (V) and horizontal (H) target motions for each of the test conditions (CON, 4×SEQ and 8×SEQ). Subjects are labeled S1–S6 with the three experienced subjects shown as triangles and the naive subjects as circles. The overall mean from all subjects is shown as a grey square.

Smooth eye velocity gain

No consistent differences were found in the gain of peak eye velocity for either RND or PRD trials in CON, 4×SEQ, and 8×SEQ conditions with V or H target motion. However, significant differences were found between the two target velocities (10°/s and 20°/s) for all subjects and for all conditions ( p < .001 in all cases) (see Figure 8). The gain of the peak eye velocity was significantly lower to the 20°/s target than the 10°/s target and subjects tended to overshoot the 10°/s target (gain > 1) and undershoot the 20°/s target (gain < 1). 

Figure 8

 

Peak eye velocity gain to both predictable (PRD) and randomized (RND) trials for the 10°/s and 20°/s velocities for each of the conditions (CON, 4×SEQ and 8×SEQ). As before subjects are labeled S1–S6 with the three experienced subjects shown as triangles and the naive subjects as circles. The overall mean from all subjects is shown as a grey square.

Relationship of latency and eye velocity

The observed differences between V50 for the H and V eye movements could be associated with differences in the latency in these directions. 

If anticipatory movements are initiated earlier to horizontal target motion, it might be expected that V50 values would be greater if eye acceleration remains the same. However, we failed to find a correlation between mean latency and V50 when all conditions were grouped together. 

Segregating the resultant data into 10°/s and 20°/s target velocities and combining V and H directions (see Figure 9), we found that for all trials (CON, 4×SEQ, and 8×SEQ) there was a consistently stronger correlation between eye velocity and latency for the 10°/s target than for the 20°/s target motion. It was also clear that the 8×SEQ and 4×SEQ conditions revealed a greater correlation ( p < .0001) between eye velocity and latency than the CON condition ( p < .005). The 8×SEQ condition also revealed the greatest significant difference between the 10°/s and 20°/s target motions. 

Figure 9

 

Mean eye velocity (V50) for each subject for both PRD and RND trials for the 10°/s (squares) and 20°/s (triangles) velocities for each of the conditions: CON (green), 4×SEQ (blue), and 8×SEQ (red) is plotted against latency. The R ² values for each condition (colors as before) are also shown for the 10°/s target on the left and the 20°/s target on the right. The solid lines indicate the regression lines for the 20°/s targets and the dashed lines for the 10°/s targets.

Discussion

General remarks

We have successfully isolated differences in eye movement responses to stimuli presented randomly or as part of a predictable sequence. As hypothesized, we found that all subjects showed anticipation for upcoming target motion in all directions for all conditions (CON, 4×SEQ, and 8×SEQ) if the sequence of target motion was predictable. No anticipatory movements were present in random conditions. Subjects were therefore able to use prior information about target motion in 2D and then release this information in the appropriate order during sequences of four discrete velocity ramps. However, the magnitude of V50 values was lower in the 4×SEQ than the CON trials. Contrary to our hypothesis, we found that increasing the number of repetitions of a four-ramp sequence from four to eight gave no consistent improvement in V50 or latency. We also found asymmetries between eye movements made to V and H target motion with H motion eliciting shorter latencies and higher eye velocity (V50) in all PRD test conditions. 

Effects of sequence learning on eye movements

The observed differences between the responses to the control and the sequence learning paradigms are of considerable interest. In a previous study, in which only horizontal target motion was presented, no significant difference was found in V50 between the control and the sequence learning task using four and six ramp sequences (Collins & Barnes, 2005). Here, however, we found that subjects show a significantly lower V50 in the 4×SEQ and 8×SEQ trials when compared to the CON trial. The total number of test conditions was comparable to the easier of the tasks (4× ramp components) in the previous experiment, with the same requirement to reproduce four choices from a total of eight. The eight conditions were slightly different (two speeds and four directions versus four speeds and two directions), but there is no obvious reason why this alone should give an impaired performance in the current experiment. We suspect that the poorer performance in 2D may be associated with the relative independence of the horizontal and vertical eye movement control pathways in the brain. When all motion was in the horizontal or vertical direction (as in the CON condition), a single control system was utilized, whereas with mixed horizontal and vertical eye movements two systems must be controlled, which may lead to increased delays in initiating anticipatory activity and consequent reduction in V50 values. 

A reduction was also observed in all subjects in the RND and PRD latency between the CON and the sequence learning trials. Latency to the CON trial was consistently shorter for all subjects (see Figure 7), leading to the suggestion that for all sequence learning trials the additional cognitive demand of learning the sequence (regardless of whether it was predictable or not) increased the latency to all the targets within that sequence. Previous experiments in this lab have shown that limitations do exist on short-term memory storage when the number of sequence components in a stimulus comprising individual horizontal smooth velocity ramps, presented in a continuous sequence, exceeds six ramp components (Barnes & Collins, 2002). 

Further, we found that performance did not improve systematically with increasing number of presentations of the sequence. As observed previously (Collins & Barnes, 2005) optimal performance is reached within the first two presentations and is then maintained but not greatly improved. This rapid learning is advantageous for prediction and anticipation but is clearly inadequate for longer term motor skill acquisition. These results suggest that the short-term memory systems used to drive the individual component of this sequence are akin to a buffer memory system. Storage in the buffer is limited and unless this action becomes consolidated in a longer term store this information will be lost. 

It would be of interest to show in the future whether information related to the sequence order is stored independently from information related to the initiation of the individual anticipatory responses. We found the 4×SEQ RND condition has a significantly longer latency (∼20 ms) than the CON RND condition. The stimulus presentation for both RND blocks shows no differences apart from the cognitive awareness that one is presented in a sequence of four ramps repeated four times and the other eight individual randomized ramps. The added sequence component may be the cause of this increase in time required to process the information. 

Asymmetries in eye movement responses

Velocity

A general observation is that subjects are able to anticipate H target motion better than V target motion. This difference was found to be significant and is in agreement with Soechting, Mrotek, and Flanders (2005), who found that subjects tracking step-ramps with abrupt changes in directions did not grade the y-component (V) velocity but did grade the x-component (H). Previous studies suggest that VOR (vestibular–ocular reflex) suppression, which has also been closely associated with pursuit, is greater in the downward direction (Lott & Post, 1993). Our results show a significant difference between upward and downward motion during the predictable trials in all conditions (CON, 4×SEQ, and 8×SEQ), with the downward motion showing lower velocities at V50. The reason for this is not clearly understood but could be related to environmental factors as downward pursuit is probably less frequently used. 

A significantly higher velocity was found for the 20°/s targets than the 10°/s. This shows that in the CON condition, subjects are actively predicting not only the direction and onset of the target but also the velocity; that is, they are scaling their eye velocity to upcoming target velocity. The scaling is less pronounced for the sequence learning paradigms. The added cognition and memory required for the 2D sequence learning paradigm is probably detrimental to the ability to scale eye velocity with target velocity. This is probably due to limitations of the short-term memory store discussed earlier (Barnes & Collins, 2002). 

In addition, we also found a significant difference between the upward and downward V50 values for PRD trials. Subjects performed better to upward targets than to downward targets. This finding is supported by Baloh et al. (1988) and Jagla, Zikmund, Mashonkina, and Yakimoff (1992) who found downward saccades to be more inaccurate. 

Latency

The RND conditions revealed shorter latencies than observed in most pursuit experiments, although Krauzlis and Miles (1993) have also recorded very short latency pursuit. To the authors' knowledge, no previous study has examined responses to the particular stimulus used here (i.e., with a gap and an initial ramp occlusion). Part of the difference could be due to the gap, which is known to reduce the latency of smooth pursuit and saccadic eye movements (Knox, 1996; Krauzlis & Miles, 1993; Saslow, 1967), and also to the occluded step toward the target direction. We found that all subjects revealed directional asymmetries in the latency to RND target locations, with vertical pursuit taking longer to initiate than horizontal pursuit. This asymmetry in V and H eye movement initiation during discrete smooth velocity ramps has not been reported previously, but again this particular stimulus has not been used before. These results lead to the assumption that our subjects showed an inherent bias toward the initiation of responses to horizontal target motion. Interestingly, this bias does not appear to be generated specifically in association with a motor plan in the short-term memory system but is clearly present in the initiation of all smooth eye movements made to discrete horizontal ramp stimuli. It must be acknowledged that this bias may be the result of training effects. However, this is highly unlikely as all six subjects exhibited this effect and three subjects were naive to oculomotor experiments. 

The bias toward the horizontal direction of motion in all subjects could be acquired via environmental experience. It seems plausible that the world consists of more horizontally moving targets than vertical. It could therefore also be conceived that this has led to better tuning of horizontal motion than vertical motion. Contrary to this hypothesis, a recent study performed in 5-, 7- and 9-month-old children found clear vertical and horizontal asymmetries in smooth pursuit which persisted in all age groups (Grönqvist, Gredebäck, & Hofsten, 2006). This indicates an inherent neuronal difference from birth that could arise from asymmetries in early visual areas (V1 and V2) (Berardi & Fiorentini, 1991). Another plausible explanation for the difference between subjects could lie in attentional and motivational differences that have previously been shown to affect this type of eye movement (Wyatt & Pola, 1987). 

Gain

We found no significant difference in peak eye velocity gain for the V and H target motions. However, peak eye velocity gain was significantly higher in 10°/s targets than in 20°/s targets in all conditions (see Figure 8). This finding is in agreement with other studies (Collewijn & Tamminga, 1984; Kao & Morrow, 1994; Rottach et al., 1996; Wyatt & Pola, 1987), which also found higher peak velocity gain to the lower target velocities during reactive conditions. Like Rottach et al. (1996), we found no consistent asymmetry in the H or V peak velocity gain during RND trials. In agreement with previous research, we also found that PRD eye movements show the same pattern of peak eye velocity gain, suggesting that both PRD and RND smooth pursuit modulate gain in the same way. This is because at the time of peak velocity the system is receiving visual feedback of target position. The reasoning for lower gain to a higher 20°/s target velocity may be due to a velocity saturation mechanism operating at higher eye velocities (Barnes, 1993). Others, however, suggest this gain difference may be due to the usual inclusion of a head movement for tracking a target of this amplitude. Therefore, the motor plan may use gaze–velocity coordinates and hence incorporate a head movement for the higher target velocity, resulting in a lower eye velocity gain (Carey & Lisberger, 2004). Studies using continuous sinusoidal wave stimuli have shown gain to be significantly higher for H than for V eye movements (Baloh et al., 1988; Rottach et al., 1996), and also that eye movements to this type of stimulus typically undershoot target velocity. Future research needs to address differences in the mechanism of responses between continuous sinusoidal wave stimuli and discrete smooth target presentations. 

Effects of latency on initial eye velocity

We plotted eye velocity against latency and isolated the results according to target velocity (see Figure 9). These results showed a good correlation between eye velocity and latency in both the 10°/s and 20°/s target motions in all conditions (CON, 4×SEQ, and 8×SEQ) but better correlation for the 10°/s than the 20°/s target velocity. This is due to the inevitably higher variability in acceleration to a 20°/s target than to a 10°/s target. As no correlation was observed between acceleration and latency, we suggest that the higher V50 is the result of an earlier eye movement onset (i.e., shorter latency) and not due to acceleration differences. 

Conclusions

To summarize, subjects show an ability to anticipate a four-ramp sequence in 2D, but latency is longer and V50 lower than eye movements made to a single repeated ramp stimulus. Increasing the number of repetitions of the sequence did not improve performance. We also found that the visuomotor system in our subjects appears more biased toward horizontal smooth eye movements during both RND and PRD target motion. We suspect this bias may be the result of an inherent brain mechanism, the result of environmental biases, or both. It was also found that the bias was present even when the information was not stored in memory, thus indicating that it is not specifically related to memory. We agree with Collewijn and Tamminga (1984) that this bias could be due to environmental factors and based on experience as subjects seem to show varying levels of this bias. We cannot, however, exclude the possibility that there is an inherent bias that is possibly amplified by environmental factors. 

We also found subjects quickly acquired a plateau of optimal performance regardless of the complexity of the task (by the third presentation as in 1D discrete movement) (Barnes & Schmid, 2002). 

There was a clear correlation between V50 and latency during all conditions, suggesting that earlier onset of anticipatory pursuit results in a higher V50. 

Acknowledgments

We would like to thank the MRC for funding this research, and Dr. Sue Collins for her support and advice throughout the project, as well as the participants. 

Commercial relationships: none. 

Corresponding author: Melanie Rose Burke. 

Email: [e-mail]. 

Address:Faculty of Life Sciences, The University of Manchester, Manchester M60 1QD, United Kingdon.