Eye movement training is most effective when it involves a task-relevant sensorimotor decision

Abstract

Eye and hand movements are closely linked when performing everyday actions. We conducted a perceptual-motor training study to investigate mutually beneficial effects of eye and hand movements, asking whether training in one modality benefits performance in the other. Observers had to predict the future trajectory of a briefly presented moving object, and intercept it at its assumed location as accurately as possible with their finger. Eye and hand movements were recorded simultaneously. Different training protocols either included eye movements or a combination of eye and hand movements with or without external performance feedback. Eye movement training did not transfer across modalities: Irrespective of feedback, finger interception accuracy and precision improved after training that involved the hand, but not after isolated eye movement training. Conversely, eye movements benefited from hand movement training or when external performance feedback was given, thus improving only when an active interceptive task component was involved. These findings indicate only limited transfer across modalities. However, they reveal the importance of creating a training task with an active sensorimotor decision to improve the accuracy and precision of eye and hand movements.

Introduction

Vision and eye movements guide most of our everyday actions. When we reach for an object, our gaze is shifted to the target long before the hand. This predictive gaze behavior serves to gather information about object identity and location, and marks future contact points for hand and fingers (Smeets, Hayhoe, & Ballard, 1996; Land, Mennie, & Rusted, 1999; Johansson, Westling, Bäckström, & Flanagan, 2001). When performing everyday actions, eye movements reflect the detailed evolution of the hand movement and the requirements of the task (Hayhoe & Ballard, 2005; Hayhoe, 2017). Gaze also leads the target when we interact with moving objects, such as when catching a ball, and thus reveals knowledge of the future target path (Diaz, Cooper, Rothkopf, & Hayhoe, 2013). Moreover, eye movements can be made predictively, for example during an occlusion interval, and scale with visual target properties such as acceleration (Bennett, Orban de Xivry, Barnes, & Lefèvre, 2007). In this context, it has been shown that more accurate tracking eye movements (smooth pursuit) result in better prediction of an object's trajectory (Bennett, Baures, Hecht, & Benguigui, 2010; Spering, Schütz, Braun & Gegenfurtner, 2011). Tracking a moving object with smooth pursuit is also associated with higher accuracy in hand movement planning (Leclercq, Blohm, & Lefèvre, 2013) and execution (Fooken, Yeo, Pai & Spering, 2016). It is important to note, however, that pursuit does not necessarily have to be aligned with the target until the moment of interception (Brenner & Smeets, 2011; López-Moliner & Brenner, 2016; de la Malla, Smeets, & Brenner, 2017). Rather, it seems that the initiation of pursuit and subsequent combination with saccades is sufficient to enable successful interception. 

Correspondingly, ball sport athletes commonly track the ball briefly using a combination of eye and head movements and then saccade to an anticipated bounce or contact location (Bahill & LaRitz, 1984; Diaz et al., 2013; Mann, Spratford, & Abernethy, 2013). Such a combination of tracking or gaze holding and prediction in eye movement behavior is considered a basic skill underlying superior athletic performance (Bahill & LaRitz, 1984; Land & McLeod, 2000; Mann et al., 2013; Uchida, Kudoh, Higuchi, Honda, & Kanosue, 2013; Vickers, 2016). 

Here we probe the relation between eye and hand movements in a rapid interception task using a perceptual training paradigm. First, we investigate whether eye-movement training enhances the ability to perform untrained goal-directed hand movements. Second, we assess whether eye movement training is sufficient to enhance hand movements, or whether it would be more effective if it was combined with hand movement training. 

The idea of transfer across modalities—from eye to hand—is based on the known tight link between eye and hand movements. On one hand, such transfer of training does not readily occur in perceptual or motor learning. For example, one of the hallmarks of perceptual learning is its specificity, i.e., lack of transfer to untrained tasks, visual locations, features, or across modalities (Polat & Sagi, 1994; Fahle, 2005). On the other hand, specificity depends on stimulus characteristics and task requirements. Transfer of perceptual learning from one location to another has been observed when the second location was previously sensitized via training an irrelevant task at that location (Xiao et al., 2008) or when the same stimuli were used (Porat & Zohary, 2016). Transfer across tasks, for example, from Vernier acuity and contrast detection to Snellen acuity, has been observed in patients with amblyopia (Levi & Li, 2009) and college baseball players (Deveau, Ozer, & Seitz, 2014), who even seem to improve on-field performance following such training (see also Faubert, 2013). The broadest transfer of perceptual learning beyond the trained task has been found after training with action video games (Green & Bavelier, 2012; Li, Chen, & Chen, 2016), resulting in improved selective attention (Green & Bavelier, 2003) and spatial resolution across the visual field (Green & Bavelier, 2007). Transfer of learning across modalities, from perception to eye movements, is possible if the task requires responses with similar underlying processing mechanisms. Szpiro, Spering, and Carrasco (2014) trained observers in a motion discrimination task during fixation and found generalization to untrained smooth pursuit eye movements. Transfer of learning has also been observed in other modalities, such as the somatosensory system: Tactile perceptual learning can transfer to untrained fingers (Dempsey-Jones et al., 2016). 

To address the question whether eye movement training is sufficient to enhance hand movement accuracy, we define and measure performance improvements in hand movement accuracy following isolated training (eye movements are trained, hand is not trained) or combined training (eye and hand movements are trained simultaneously). In an effort to address a bidirectional relation between eye and hand movements, we also investigate the effect of isolated or combined training on eye movement accuracy and precision. A comparison of effects of different training protocols will allow us to evaluate whether eye movement training alone is sufficient to improve the eye, or whether there are added benefits of involving the hand in training, indicating transfer. 

We assessed participants in a motion prediction task that required rapidly intercepting a moving object with the hand. In this task, participants viewed a target moving along a curved trajectory—akin to a lateral view of a batted baseball. Importantly, the target was shown only briefly, and participants had to extrapolate its motion trajectory before intercepting it in a dedicated area on the screen. Ultra-short target presentation durations make this task difficult (Fooken, Yeo et al., 2016) and therefore suitable for a training paradigm. Eye and hand movements were recorded simultaneously. Testing and training with different training protocols occurred over a period of five consecutive days; we also assessed the longer-term retention of learning in a follow-up session one week after training. Some of the training protocols involved an active movement towards the assumed target position and external performance feedback on whether the eye or hand successfully reached the target. Performance feedback is widely considered an important component in training across modalities (Swinnen, 1996; Herzog & Fahle, 1999; Gray, 2009; Wolpert, Diedrichsen, & Flanagan, 2011; Sigrist, Rauter, Riener, & Wolf, 2013). It might accelerate learning (Fahle & Edelman, 1993; Werner & Bock, 2007) and facilitate transfer (Salmoni, Schmidt, & Walter, 1984; Swinnen, Lee, Verschueren, Serrien, & Bogaerds, 1997; Deveau et al., 2014; Tanaka & Watanabe, 2017). We manipulated external performance feedback to investigate and compare feedback effects on eye and hand movement. 

Materials and methods

Participants

We recruited 50 right-handed undergraduate students (mean age = 24.2 years; SD = 3.5; 29 female, 21 male) with corrected-to-normal visual acuity and no history of disease interfering with normal eye movement function. The experimental protocol adhered to the Declaration of Helsinki and was approved by the UBC Behavioral Research Ethics Board. Participants gave written informed consent prior to participation and were unaware of the purpose of the experiment. Each participant was randomly assigned to one of five groups (n = 10 each): four training groups that were tested at least five times and one control group that was tested at least twice (see Table 1). Sample size per group is comparable to other studies investigating training effects on eye or finger movements (Dempsey-Jones et al., 2016; Porat & Zohary, 2016). Remuneration was $8 per hour for each session; a bonus of $20 was paid upon completion of all five sessions. All 50 observers completed the study, and 41 of them returned for the follow-up session one week later ($10 remuneration). 

Table 1

 

Training protocols for five groups. Notes: Training on days 2–4 could include eye movements only (eye), or combined eye and hand movements (hand); it either involved external performance feedback (FB) or not. Pre- and posttest were identical across groups. Number of observers during follow-up testing varied by group.

 

Visual stimuli and apparatus

Stimuli, apparatus, and task were based on methods developed in Fooken, Yeo et al. (2016), and are reproduced here for the reader's convenience. Observers had to track and intercept a briefly presented, small, black Gaussian dot with diameter 2° of visual angle, shown at a luminance of 5.4 cd/m². The stimulus moved across a uniform gray background equally divided into a lighter gray “tracking zone” on the left (35.9 cd/m²) and a darker gray “hit zone” on the right (31.5 cd/m²; Figure 1a). The physical trajectory of the target was simulated to be the natural flight of a batted baseball. Visual stimuli were back-projected using a PROPixx video projector (VPixx Technologies, Saint-Bruno, Canada) with a refresh rate of 60 Hz and a resolution of 1,280 (H) × 1,024 (V) pixels. Observers sat at a distance of 46 cm from the screen with their head supported by a chin and forehead rest. The screen was a 44.5 × 36 cm translucent display consisting of non-distorting projection screen material (Twin White Rosco screen; Rosco Laboratories, Markham, Canada) clamped between two glass panels and fixed in an aluminum frame. Stimulus display and data collection were controlled by a Windows PC with an NVIDIA GeForce GT 430 graphics card running MATLAB 7.1 (MathWorks, Natick, MA) and PsychToolbox 3.0.8 (Brainard, 1997; Pelli, 1997). 

Figure 1

 

Experimental procedure and design. Stimuli moved at one of three different speeds, resulting in three trajectory types (a), and were presented for either 100, 200, or 300 ms. An example trial sequence from the pretest track-intercept task is presented in (b). Each trial started with fixation in the “tracking zone,” followed by stimulus motion for 100–300 ms. Observers were instructed to track the target with their eyes even after it disappeared, and to intercept it anywhere within the “hit zone” with their index finger. Performance feedback was given after each trial. On training days (c), observers were instructed to either track the target with their eyes (groups 1 and 2), or to intercept with their finger (groups 3 and 4). Only groups 2 and 4 received performance feedback during training; group 5 was not trained and served as the control group.

Procedure, task and design

Testing sessions.

Observers were tested before training (pretest) and after training (posttest and follow-up test) using the following procedure and task (Figure 1b): Each trial began with fixation on a stationary ball presented 14° to the left from screen center; fixation duration was randomized (500–700 ms). The ball then moved to the right at one of three speeds (25°/s, 30°/s, or 35°/s) and disappeared after a short presentation duration (100, 200, or 300 ms) before reaching the hit zone (see Figure 1a). Speed and presentation duration were randomized within each block of trials. Observers were instructed to track the ball smoothly with their eyes, to continue to track it after it disappeared, and to intercept it with their index finger at an assumed position anywhere within the hit zone. Interceptions were always made in the ipsilateral (right) half of the screen and observers returned their hand to a table-fixed resting pad after each trial. Feedback on manual interception performance was provided at the time of interception; a red dot indicated interception location and a black dot indicated true target position (Figure 1b). Performance feedback was visible for 500 ms; the next trial started in immediate succession. Each pre- or posttesting session (162 trials) took approximately 20 min. 

Training sessions.

Training protocols (groups 1 through 4; days 2–4) differed with regard to response modality (eye alone or combined eye-hand) and visual feedback (Table 1; Figure 1c). Groups 1 and 2 did not move their hand during training. Group 1 was instructed to track the target with their eyes into the hit zone; the target never reappeared, and observers received no performance feedback. Group 2 was asked to track the target and to actively intercept it with their eyes. To be recognized as final eye position (“hit” or “intercept”), observers had to hold their gaze within a 1.4° radius of the assumed target position in the hit zone for 200 ms. They received feedback about where the target was relative to their eye: A red dot indicated their final eye position, and a black dot showed true target position. Observers in groups 3 and 4 were asked to track the target with their eyes and to intercept it with their finger, just as they did during testing. Group 3 received no feedback (only interception position, not true target position, was shown) and group 4 received the same full performance feedback as during testing. Training sessions included three blocks of 162 trials each (486 trials total) and took approximately 60 min. to complete. Group 5 served as a control and received no training; these participants were only tested twice and did not come into the lab on training days. 

Eye and hand movement recordings and analysis

We recorded right eye position with a video-based eye tracker (tower-mounted Eyelink 1000; SR Research Ltd, Ottawa, Canada) at a sampling rate of 1,000 Hz. At the start of each block of trials, the measurements were calibrated by asking observers to fixate on a small visual target appearing successively at nine different locations on the screen; this procedure was repeated for validation. All data were analyzed off-line using custom-made routines in MATLAB for the detection of saccades and pursuit onset (Fooken, Yeo et al., 2016). Saccades were excluded from pursuit analysis and analysed separately. We computed relative eye velocity (calculated as gain: frame-by-frame eye velocity divided by target velocity in the interval 140 ms after pursuit onset to interception) as the parameter defining pursuit quality. We also computed the cumulative amplitude of catch-up saccades, i.e., the total distance covered by saccades. To investigate how closely the eye landed relative to the final position of the target at the end of the trial, we analyzed the interception error of the eye. For this measure, we defined the target position, p_tar, and the final eye position p_eye at the time the trial finished. The trial finished either when the target was intercepted manually (pre- and posttest; groups 3 and 4), the target was intercepted with the eyes (group 2), or the target reached the end of its trajectory (group 1). Eye interception error (or accuracy) was then calculated as the Euclidean distance d_eye = ‖p_tar − p_eye‖. 

During all testing sessions, and during training sessions that involved the hand (groups 3 and 4), movements of observers' right index finger were tracked with a magnetic tracker (3D Guidance trakSTAR, Ascension Technology Corp, Shelburne, VT) at a sampling rate of 240 Hz. A lightweight sensor was attached to the observer's fingertip with a small Velcro strap. Calibration of measurements obtained with the trakSTAR relied on a 5-point procedure in which observers were asked to point to the location of a visual target appearing successively at five locations with their index finger and keep their finger at this location until the next target appeared. This procedure immediately followed eye tracker calibration at the start of each trial. We recorded the 2D finger interception position in x- and y-screen-centered coordinates for each trial. To analyse manual interception accuracy, we first defined the finger hit position, p_finger, as the 2D position of the finger when it first makes contact with the screen. The target position at that time is denoted as p_tar. Interception accuracy is the interception error, calculated as the Euclidean distance d_finger = ‖p_tar − p_finger‖. We also analysed interception precision, defined as the statistical variance, indicating spatial variability of x- and y-interception positions. All trials were manually inspected, and those with eye blinks (4.3% of all trials), hand movement onset prior to target onset (1%), or undetected finger end position (2%) were excluded from further analysis. 

Hypotheses and statistical analyses

Hypotheses are graphically presented in Figure 2. First, we expected that movement accuracy would improve in all training groups as compared to the control group. This improvement might be limited to the trained modality; i.e., hand movement accuracy will increase when training involves moving the hand, and eye movement accuracy will increase after training that involves the eye (Hypothesis 1a; Figure 2a). Alternatively, the improvement might transfer across modalities; i.e., hand movement accuracy will increase after eye movement training, and eye movement accuracy will increase after hand movement training (Hypothesis 1b; Figure 2a). Second, we expected that external performance feedback based on an active interceptive movement might boost movement accuracy for each modality (Hypothesis 2; Figure 2b). Alternatively, if transfer was found (Hypothesis 1b), feedback might enhance performance across modalities. 

Figure 2

 

Schematic predictions of training effects. The graphs in (a) illustrate expected improvements after training as compared to the control group within each modality (Hypothesis 1a, H₀ not shown) and potential boost across each modality (Hypothesis 1b, transparent bars). Graphs in (b) illustrate the expected effect of feedback on performance (Hypothesis 2, transparent bars). If feedback had no effect on eye and hand movement, accuracy results would be equivalent to panel (a).

Differences between groups, and the hypothesis that each training group improved within the given response modality more than the control group, were examined using one-way ANOVA with factor group. Intergroup differences were examined using a posthoc Dunnett's test to account for potential dependencies in multiple comparisons between each training group and the control group (Dunnett, 1955). With n – k degrees of freedom (sample size minus number of groups), α = 0.05, five groups, and a sample size of n = 10 per group, the critical value above which a Dunnett's test would be significant is t = 2.89. We also compared training groups with regard to response modality and whether or not feedback was given using repeated-measures ANOVA with factors modality and feedback. Changes across training days were investigated with repeated-measures ANOVA with factor day. In all groups and sessions, speed and presentation duration were varied to increase stimulus uncertainty. Based on previous studies using the same paradigm, we expected that these stimulus conditions would systematically affect eye and hand movements (Fooken, Yeo et al., 2016; Kreyenmeier, Fooken, & Spering, 2017), and we replicate those results here. Effects of speed and presentation duration were analysed using repeated-measures ANOVA with factors speed and duration; this analysis was conducted on data averaged across training groups for the pretest. However, we had no specific hypotheses regarding interactions between training groups and stimulus conditions, and thus pooled across speeds and presentation durations when testing our main study hypotheses. Statistical analyses were conducted in IBM SPSS Statistics Version 24 (SPSS, Inc, Armonk, NY). 

Results

Our results focus on the effects of training on eye and hand movement accuracy (Hypothesis 1, Figure 2a) and describe how performance changed from pretest to posttest. We also consider the role of external feedback in task versions requiring active interception (Hypothesis 2, Figure 2b), and day-to-day improvements across training days. 

Effects of target properties on eye and hand movements

We varied speed and presentation duration—two variables that can be expected to strongly influence eye and hand movements. Replicating previous results (Fooken, Yeo et al., 2016), pursuit was more accurate when the target was presented for a longer duration and when it moved at a slower speed: Relative eye velocity increased with increasing presentation duration, F(2, 98) = 115.4, p < 0.001, η² = 0.70, and with decreasing speed, F(2,98) = 247.9, p < 0.001, η² = 0.84. Manual interceptions were most accurate (smallest interception error) when the target was presented for a longer duration, F(2, 98) = 58.3, p < 0.001, η² = 0.54, and when it moved at a medium speed, F(2, 98) = 38.53, p < 0.001, η² = 0.44, indicating a speed-range effect. 

In general, observers tended to intercept early in the hit zone and close to the medium-speed target trajectory, in line with a speed-range effect. These patterns were observed across all groups, despite high individual variability in interception strategy (e.g., early vs. late in the hit zone). Figure 3 shows 2D interception positions during pre- and posttest for four representative observers: two observers from the control group, for whom we would expect the smallest performance increase, and two from group 4 (eye-hand training with feedback), for whom we would expect the largest performance increase. 

Figure 3

 

2D interception positions of two representative observers in group 5 (no training; top row) and two observers in group 4 (eye-hand training with feedback; bottom row); each data point denotes interception position in one trial in the pretest (closed symbols) or posttest (open symbols). Target speeds are denoted by line color and presentation durations by symbol type.

Whereas absolute interception position and timing (later interceptions were made further into the hit zone) differed between individuals, they did not differ significantly between groups. This observation was confirmed using a multivariate ANOVA with group as fixed factor, yielding no significant group effects on differences between x-interception position, F(4, 45) = 1.71, p = 0.16, η² = 0.13, or y-interception position (F < 1) in pretest versus posttest. Our main analyses therefore focus on interception accuracy and precision, rather than on absolute position. 

Training the hand enhanced hand movement accuracy and precision irrespective of external feedback

We calculated finger interception error, i.e. the 2D distance between finger and target end position at time of interception, to evaluate hand movement accuracy. Interception error was overall smaller after training as compared to before training in all groups (Figure 4a). The largest percentage improvements can be seen in groups that trained while using their hand (Figure 4b and c). These observations were confirmed by significant comparisons (Dunnett's t) between hand-training groups with the control group (group 3: p = 0.05, Cohen's d = 1.2; group 4: p = 0.02, d = 1.41), implying that training the hand in a manual interception task improved interception accuracy. Performance in the eye-training groups also improved as compared to the control group, but these changes were nonsignificant (group 1: p = 0.53, d = 0.27; group 2: p = 0.27, d = 0.65), reflected in an overall nonsignificant effect of the factor group, F(4, 45) = 2.2, p = 0.08, η_p² = 0.16. These results indicate that hand movement accuracy only improved when the hand was engaged during training, whereas isolated eye movement training did not benefit the hand. 

Figure 4

 

Training effects on manual interception accuracy. (a) Interception error (in degrees) during pretest versus posttest. Each data point is the mean 2D interception error for one observer across all trials; larger data points with 2D error bars are group means. Data points falling above the unity line indicate higher error during the posttest; data points below the unity line denote higher error during the pretest. (b) Interception accuracy increase (degree) in comparison to the pretest, with error during pretest set to zero, for all testing and training days. (c) Percentage accuracy increase (error decrease) in posttest relative to pretest. Asterisks denote significant results of Dunnett's t test, *p < 0.05. (d) Absolute interception error averaged across observers that came for the week follow up for post- (darker) compared to week-test (lighter). Asterisks denote significant results of pairwise t test, *p < 0.05, ***p < 0.001. Error bars in all panels denote standard errors of the mean.

Notably, the interception performance level achieved after training remained stable or continued to increase one week later (see Figure 4d). A repeated-measures ANOVA with within-subjects factor time (post vs. week test) and between-subjects factor group revealed a significant main effect of time, F(1, 35) = 5.93, p = 0.02, η² = 0.15, and a significant group × time interaction, F(4, 35) = 2.97, p = 0.03, η² = 0.25. Note that these statistics are based on observers who participated both in the posttest and in the week test (n = 41). Hence, they do not exactly match posttest means shown in Figure 4b for n = 50. 

Whereas feedback did not modulate hand movement accuracy after training—no significant main effect of feedback on posttest performance (see Figure 4c), and no feedback × modality interaction (all F < 1)—it did affect the time course of learning. Figure 4b shows improvements in interception accuracy relative to pretest for each day. When feedback was given (group 4), interception accuracy increased across training days, saturated on day 3, and remained stable during posttest. When no feedback was given and observers had no knowledge of the target end position (group 3), interception accuracy decreased across training days, indicating that observers hit further away from the target in the absence of feedback. Performance in this group recovered during posttest when observers received feedback, reaching the same performance level as the group that had received feedback throughout training. These observations were confirmed by a repeated-measures ANOVA with factors feedback and day, yielding a significant main effect of day, F(5, 65) = 7.3, p < 0.001, η_p² = 0.36, and a feedback × day interaction, F(5, 65) = 4.13, p = 0.003, η_p² = 0.24. In sum, hand movement accuracy increased significantly following training that involved the hand. These improvements occurred irrespective of feedback, even though feedback modulated performance during training days. 

It is noteworthy that performance increased even in the absence of any training (see control group data in orange in Figure 4), posing the question whether the pretest alone might have been sufficient to provide full and extensive training of the task. We investigated this possibility by analysing average interception error during the pre- and posttest in bins of nine trials. Figure 5 shows average binned interception errors for two groups—group 5 (no training) as the group that improved the least (Figure 5a), and group 4 (eye-hand training with feedback) that improved the most (Figure 5b). For both groups, interception error decreased during the first half of the pretest, and then remained stable at a level of around 3° on average across all speeds. Even though the control group's performance improved at the start of the posttest in the absence of training, overall interception error was higher (M = 2.4° ± 0.13°) than for the eye-hand training group (M = 2.1° ± 0.16°), indicating that additional training was useful. This group continued to improve on training days, and reached an overall higher level of accuracy. Most importantly, neither group achieved ceiling performance at any time. 

Figure 5

 

Mean interception error in response to three different speed conditions across time; (a) Group 5 (no training), and (b) group 4 (eye-hand training with feedback). Values for pre- and posttest are averaged within bins of nine trials. Values for training days are averaged across session (n = 486 trials). Filled symbols denote group mean; open symbols are means per subject.

Finally, we investigated effects of training on interception precision, the spatial variability of x- and y-interception positions. Figure 6 shows mean variance for each group before and after training and reveals large improvements in precision for the four training groups, especially along the horizontal axis (timing; i.e., how far into the hit zone observers intercepted; see Figure 3). The control group also improved, but to a much smaller extent (Figure 6e), maintaining relatively large spreads in horizontal and vertical directions (see also Figure 3a). 

Figure 6

 

2D interception variance (precision) in pretest (outer ellipse) as compared to posttest (inner ellipse) for all groups averaged across conditions and observers.

Training the eye differentially enhances eye movement performance

Training effects on eye movements were quantified by calculating two main performance measures: interception error of the eye (i.e., eye position relative to final target position), and relative eye velocity (i.e., eye velocity relative to target velocity). 

Similar to the pattern observed for manual interception error, the eye interception error decreased with training: main effect of group on accuracy improvement in posttest relative to pretest, F(4, 49) = 2.96, p = 0.03, η_p² = 0.21 (Figure 7a). When no feedback was given and observers had no knowledge of the target end position (group 1), eye interception accuracy decreased across training days (Figure 7b). Performance in this group recovered during posttest when observers received feedback, reaching the same performance level as the group that had received feedback throughout training. However, we observed no significant overall intergroup differences due to large improvements in the control group (Figure 7c). Across all groups, observers' final eye and hand positions were closely linked, indicating that observers pointed where they looked, within 1.9° ± 0.32° in the pretest, and 1.5° ± 0.25° in the posttest. 

Figure 7

 

Training effects on eye interception accuracy. (a) Interception error (in degrees) during pretest versus posttest. Each data point is the mean 2D interception error for one observer across all trials; larger data points with 2D error bars are group means. Data points falling above the unity line indicate higher error during the posttest; data points below the unity line denote higher error during the pretest. (b) Interception accuracy increase (degree) relative to the pretest, with error during pretest set to zero, for all testing and training days. (c) Percentage accuracy increase (error decrease) in posttest relative to pretest.

Figure 8 shows mean eye velocity profiles for all test and training days for each group. Observers commonly initiated smooth pursuit in anticipation of the predictable motion direction; the frequency of occurrence of these anticipatory movements increased with exposure to the task (compare pre- and posttest velocity values at time 0, when the target started moving). At around 100–150 ms after target onset eye velocity increased rapidly, but never reached target speed, decreasing again after reaching a peak at around 300 ms. Pursuit's transient nature in this task is due to limited target presentation duration (100–300 ms), and the peak at 300 ms indicates that observers attempted to maintain pursuit up to the maximum visible target duration. Accordingly, relative eye velocity (gain) values were on average 0.25 ± 0.11. The eye velocity profiles reveal an increase in eye velocity from pretest to posttest in training groups 2 through 4 (compare colored dotted and dashed lines in panels b through d). In these groups, performance also improved across training days from day 1 to day 3 (compare light and dark gray lines). By contrast, group 1 showed only a slight increase, and the control group showed no increase in eye velocity. 

Figure 8

 

Mean eye velocity traces as a function of time relative to target motion onset across all observers (n = 10 per group; one panel per group). Saccades were replaced by linear interpolation. Line style denotes testing or training day.

We quantified the observed effects of training by calculating relative eye velocity, i.e. the mean of the ratio of eye versus target velocity at each time frame between 140 ms after onset to time of interception, as a measure of eye movement accuracy. Training effects emerged early during the pursuit response and built up over time, reaching a peak at around 250 ms after stimulus onset, just before the eye started to decelerate. Correspondingly, relative eye velocity increased up to 20% in the training groups, whereas the control group's relative velocity slightly decreased from pre- to posttest (Figure 9a through c). These findings were confirmed by a significant main effect of group on relative velocity improvement in posttest in comparison to pretest, F(4, 49) = 3.65, p = 0.01, η_p² = 0.25. Similar to what we observed for hand movement accuracy, eye movement performance differences were significant in both hand-training groups regardless of feedback (group 3: p = 0.01, d = 1.32; group 4: p = 0.006, d = 1.16). Accordingly, we found no main effect of feedback on posttest relative velocity in training groups, F(1, 36) = 1.88, p = 0.18, η_p² = 0.05, and no feedback × modality interaction, F(1, 36) = 1.82, p = 0.19, η_p² =0.05. 

Figure 9

 

Training effects on eye movement accuracy. (a) Relative eye velocity during pretest versus posttest. Each data point is the mean relative velocity for one observer across all trials; larger data points are group means. Data points falling above the unity line indicate higher relative velocity during the posttest; data points below the diagonal denote higher relative velocity during the pretest. (b) Relative velocity change in comparison to the pretest, with relative velocity during pretest set to zero, for all testing and training days. (c) Percentage velocity increase in posttest relative to pretest. Asterisks denote significant results of Dunnett's t test, *p < 0.05, **p < 0.01. (d) Relative velocity averaged across observers that came for the week follow up for post (darker) compared to week-test (lighter). Error bars in all panels denote standard errors.

Eye movement improvements were also significant in the eye training group that received feedback (p = 0.005, Cohen's d = 1.63), but not in the eye training group without feedback (p = 0.13, d = 0.81). Again, feedback modulated the time course of learning (compare green lines in Figure 9b). This observation was confirmed by a main effect of feedback on relative eye velocity for the two eye-movement training groups, F(1, 18) = 17.21, p = 0.001, η_p² = 0.49, but not for the hand-training groups (F < 1). This differential modulatory effect of feedback on eye and hand movement training was reflected in a significant feedback × modality interaction, F(2, 36) = 25.07, p < 0.001, η_p² = 0.59, during training days, and across all testing and training days, F(1, 10) = 6.01, p = 0.03, η_p² = 0.38. 

Similar to what we observed for hand movement performance, eye accuracy remained unchanged compared to posttest when tested one week later (see Figure 9d). A repeated-measures ANOVA with within-subjects factor time (post vs. week test) and between-subjects factor group revealed no effect of time, F(1, 35) = 1.58, p = 0.22, η² = 0.04, and no group × time interaction (F < 1). Again, these statistics are based on subjects who participated in the week test (n = 41), resulting in the difference to posttest data shown in Figure 9b for n = 50. In sum, eye movement accuracy increased significantly either when training involved the hand, or when external performance feedback was given, as was the case in task versions that involved a sensorimotor decision. 

It is important to note that performance differences between the two isolated eye training groups (Figure 9b and c) could result either from the fact that only one of these groups received external performance feedback, or from differences in eye movement behaviour. Whereas group 1 merely tracked the target, group 2 was asked to intercept with their eyes, triggering a goal-directed movement at the assumed target position. As a result of this instruction, observers in groups 1 and 2 produced eye movement patterns that differed both qualitatively and quantitatively, resulting in overall higher eye interception accuracy in group 2 (see also Figure 7b). Observers in group 1 tracked the target for longer periods of time, using a combination of smooth pursuit and saccadic eye movements (see example in Figure 10a and c), whereas observers in group 2 made an early saccadic eye movement towards the target, thus terminating the trial (Figure 10b and d). Congruent with these observations, observers in group 1 initiated their final saccade on average 500 ms later than group 2 (latency group 1: M = 1169 ± 39 ms; group 2: M = 503 ± 12 ms). Moreover, observers in group 1 made on average twice as many saccades as compared to group 2 (group 1: M = 4.3 ± 0.24; group 2: M = 2.4 ± 0.14), resulting in a larger cumulative saccade amplitude (group 1: M = 31° ± 1.9°; group 2: M = 15° ± 0.5°). These results indicate that performance differences between both groups are likely modulated by differences in task, rather than the fact that external performance feedback was given. 

Figure 10

 

(a) Horizontal and vertical eye position for a representative observer in group 1 (eye training no feedback) for two single trials during a training day. Black line denotes average target path shown until time of interception, the vertical gray line denotes point of target disappearance at 200 ms. (b) Eye position of a representative observer in group 2 (eye training with feedback) in two trials in which the target disappeared after 200 ms. In both panels the target entered the hit zone at 1,000 ms after target onset. (c) Horizontal and vertical eye position as a function of time for one representative trial from the same observer as in (a). (d) Horizontal and vertical eye position for the same observer as in (b).

Discussion

This study investigates under which circumstances perceptual-motor learning transfers across modalities, and the mechanisms underlying performance improvements in eye and hand movements. Using a motion prediction task and comparing five different types of training, we report three key findings. First, eye movement training does not transfer to hand movements, despite known close links between both modalities in tasks that involve goal-directed hand movements. Second, eye movements improve most when training involves an interceptive movement (either eye or hand). Third, external performance feedback has relatively little influence on training outcome. 

Eye movement training does not transfer to hand movements

Eye movements are usually made spontaneously when observers engage in visually-guided manual tasks such as reaching, grasping, pointing, or hitting. Eye and hand movements are spatially and temporally coordinated: gaze leads the hand by up to 1 s (Ballard, Hayhoe, Li, & Whitehead, 1992; Smeets et al., 1996; Sailer, Flanagan, & Johansson, 2005; Land, 2006), and gaze locations are anchored to future contact points on the target, indicating strong spatial coupling (van Donkelaar, Lee, & Gellman, 1994; Neggers & Bekkering, 2000; Gribble, Everling, Ford, & Mattar, 2002; Brenner & Smeets, 2011; Cesqui, Mezzetti, Lacquaniti, & d'Avella, 2015; Vazquez, Federici, & Pesaran, 2017). Many of these studies have focused on the saccade-to-reach relationship. Using the same motion prediction task as in the current study, we recently extended these findings to smooth pursuit, revealing a close relationship between the accuracy of pursuit and the accuracy of manual interceptions (Fooken, Yeo et al., 2016). This link was closest at the time of interception, indicating a common spatiotemporal framework for the control of smooth pursuit and interceptive hand movements. One potential consequence of such common mechanisms would be that improvements in one domain—the eye—should transfer to the other—the hand. Yet, the current study showed that training eye movements alone was not sufficient to improve hand movements, revealing no transfer from eye to hand (Figure 3). This result was obtained regardless of the type of eye movement training employed (i.e., with or without feedback). Transfer of learning across modalities might only be possible if task requirements are strongly aligned and rely on the same processing mechanisms. Szpiro et al. (2014) observed transfer from perception to pursuit in a motion discrimination task that required perceptual estimation of the target's motion direction. There is considerable overlap in the neural mechanisms underlying motion perception and smooth pursuit (Lisberger, 2010; Osborne, Lisberger, & Bialek, 2005; Spering & Montagnini, 2011), facilitating transfer from motion perception to motion tracking. Even though there is evidence for interdependency between the neural control of eye and hand movements, particularly within posterior parietal cortex (Snyder, Batista, & Andersen, 1997; Buneo & Andersen, 2006; Cui & Andersen, 2007; Battaglia-Mayer, Ferrari-Toniolo, & Visco-Comandini, 2015), both types of movement are ultimately controlled by effector-specific networks. Moreover, there is little research on the neural mechanisms underlying pursuit-hand coordination, and the extent of overlap between the cortical architecture underlying each type of movement is unclear. Finally, our current task was more complex and required not only processing of sensory motion information, but also trajectory prediction, based on past experience. Lack of transfer could indicate that the process of integrating sensory with experience-based information might differ for pursuit and hand movements. 

In sum, our finding of lack of transfer from eye to hand is congruent with much of the perceptual learning literature indicating specificity of learning (Fahle, 2005). It emphasizes the importance of designing training tasks whose requirements mimic real-world requirements across diverse areas of application (e.g., a particular type of move in sports, or clinical rehabilitation). 

Eye movements improve most when training involves a sensorimotor decision

We found that training smooth pursuit eye movements alone was also not sufficient to improve eye movement accuracy. Our training group 1 merely viewed the target briefly and tracked it with their eyes; these observers never saw the target end position in the hit zone. As a result, neither eye nor hand movements improved as compared to the control group. This finding is consistent with literature on eye movement training, showing only marginal improvements in pursuit velocity gain after training (Guo & Raymond, 2010; Szpiro et al., 2014). Yet, some studies have reported beneficial effects of smooth pursuit training. Experimentally naïve monkeys showed higher pursuit velocity and fewer catch-up saccades after many rigorous training sessions (Bourrelly, Quinet, Cavanagh, & Goffart, 2016). Moreover, engaging in contralesional pursuit improved functional recovery in human patients with visual neglect (Kerkhoff et al., 2013; Kerkhoff et al., 2014). These studies differ from ours in many aspects—study subjects, type of stimulus and task, and probably most importantly, duration and intensity of training. It is possible that prolonged pursuit training over many weeks could have increased pursuit performance in our study. Congruent with this assumption, group 1 was the only group displaying a small trend towards further improvement when tested again one week after training had been completed (Figures 3d and 5d). 

Interestingly, when pursuit training was paired with the instruction to make an eye movement toward the assumed target position in the hit zone (group 2), eye movements improved considerably. This improvement could be due to a combination of factors, including the preparation and execution of a goal-directed saccadic eye movement (Figure 6), the predictive sensorimotor decision underlying this saccade, or motivational aspects related to external performance feedback. Feedback did not differentially affect eye movement performance in groups where training involved the hand (Figure 5c). It is therefore unlikely that the difference between the two eye-training groups was entirely driven by feedback. Instead, differences in eye movement behavior, most notably an early goal-directed saccade and overall shorter and more accurate tracking in group 2, could underlie the finding that training in group 2 was more effective. Making an interceptive saccade comprises a different behavioral goal than just tracking a moving target without any task related to its trajectory; it requires prediction based on experience from previous trials (location of feedback) and integration with current sensory information (initial launch trajectory). Eye movements generally reflect the requirements of visual-motor tasks such as reaching, grasping, or walking (Hayhoe & Ballard, 2005; Hayhoe, 2017). Along the same lines, the behavioral goal of intercepting the target with the eyes might determine the need for accurate trajectory prediction, which in turn can only be achieved with accurate pursuit (Spering et al., 2011). 

Finally, we acknowledge that we attributed beneficial effects of training that engaged the hand to the fact that the hand performed a goal-directed movement, involving a sensorimotor decision. Alternatively, improvements might be possible just based on engaging the hand in any kind of movement. This possibility could be tested by including a control group that engages in an independent pointing or hand movement task, not training the observer in the primary task. 

External performance feedback did not modulate training

Providing external performance feedback generally boosts effects of training in visual (Deveau et al., 2014) and motor tasks (Swinnen, 1996; Sigrist et al., 2013). Gray (2009) compared swing accuracy in a baseball batting simulator when giving visual, auditory, or tactile feedback, respectively, to a no-feedback condition. Swings were more accurate when feedback was given, and visual feedback produced the greatest accuracy. However, feedback might not be critical for learning to occur: Beneficial effects of training have been repeatedly reported in the absence of feedback (Herzog & Fahle, 1997; Liu, Lu, & Dosher, 2010; Szpiro et al., 2014). In our study, feedback significantly altered performance on training days (Figures 3b and 5b) but did not significantly modulate training outcome. 

It is possible that effects of external performance feedback were attenuated by the existence of both visual as well as internally generated feedback. In versions of the task that involved hand movements, observers were always able to see their own hand and thus received a strong visual feedback signal. Moreover, with every movement we make, the brain sends a copy of the movement command—an efference copy or corollary discharge—back to sensory brain areas, which then integrate these feedback signals with sensory input (Bridgeman, 1995; Crapse & Sommer, 2008; Sommer & Wurtz, 2008). Another source of internal feedback information comes from proprioceptive information signaling the position of the eye in the orbit, or the arm in space (Vercher, Gauthier, Cole, & Blouin, 1997; Ren et al., 2006). Visual and internally generated feedback information might have boosted training across all training groups in our study, rendering external performance feedback less important than in studies involving perceptual judgments, where no visual or internally-generated feedback is automatically available. 

Seeing the hand while it moves, especially close to the time of interception, provides important information needed to correct movements online (de la Malla, López-Moliner, & Brenner, 2012) and boosts spatial accuracy (de la Malla & López-Moliner, 2012). Further evidence for beneficial effects of visual and internally generated feedback on motor task performance comes from the neglect literature. Patients with hemispatial neglect (e.g., following stroke) show severe biases in processing visual information presented in their contralesional hemifield. Yet, they can be surprisingly accurate when performing motor tasks, such as goal-directed hand movements, in their blind hemifield (Harvey et al., 2001; Himmelbach & Karnath, 2003; Harvey & Rossit, 2012). This finding has been attributed in part to the availability of visual information (seeing the arm move in space); it could also be due to internally generated feedback (efference copy and proprioceptive feedback). 

Conclusion

Our results highlight the importance of a naturalistic task design to successfully train observers' sensorimotor performance. Learning is optimal when the training task involves a sensorimotor decision, here an active interception of the target either by eye or hand movement. We found no direct transfer of training from eye to hand, indicating that cross-modality transfer likely requires coactivation of the neural networks underlying trained effectors. Our results also revealed only little influence of external performance feedback on training outcome, indicating that internally generated feedback during learning may be sufficient to boost eye and hand movement accuracy and precision. 

Acknowledgments

This work was supported by an NSERC Graduate Scholarship to KML, and an NSERC Discovery Grant (RGPIN 418493) and a Canada Foundation for Innovation John R. Evans Leaders Fund equipment grant to MS. Data were presented in preliminary form at the 2016 Vision Sciences Society meeting in St Pete's Beach, FL (Fooken, Lalonde, & Spering, 2016). 

Commercial relationships: none. 

Corresponding author: Jolande Fooken. 

Email: . 

Address: Visual Performance and Oculomotor Mobility Lab, Blusson Spinal Cord Centre, Vancouver, Canada