Hastening orientation sensitivity

Abstract

Previous perceptual learning studies have shown that sensitivity to subtle orientation differences improves with practice at oblique axes but not with practice at cardinal axes. The cause of this anisotropy in angular resolution is uncertain, and it is not known whether the same anisotropy pertains to temporal resolution—the minimum stimulus duration needed to achieve a specified angular resolution. Here, we investigated the hypothesis that cardinal improvements were previously absent because long stimulus durations yielded maximal precision, even at the start of training. Accordingly, we exploited the relatively imprecise responses that occur naturally when masked stimuli are presented for extremely brief durations. After 110,000 trials were completed over seven daily sessions, temporal resolution improved by 51% at cardinal axes and by 86% at oblique axes. This hastening of the visual response was accompanied by significant improvements in angular resolution, which were specific to the trained axis. The data demonstrate plasticity in the response to cardinal orientations and indicate that sufficient initial levels of neural imprecision may be necessary for perceptual learning.

Introduction

A fundamental aspect of our visual experience is our sensitivity to variations in orientation. Orientation sensitivity is often measured with respect to angular resolution—the smallest angular difference that can be reliably discerned. However, another important aspect of orientation sensitivity is temporal resolution—the briefest stimulus duration needed to reliably see a given angular difference. This study was conducted to provide new information about the angular and temporal resolution of orientation sensitivity, using a perceptual learning procedure. Earlier perceptual learning studies on fine orientation sensitivity have tended to address angular resolution only. Those studies indicated that practice sharpened angular resolution at oblique (diagonal) axes but not at cardinal (horizontal or vertical) axes (Matthews & Welch, 1997; Vogels & Orban, 1985). Some insight about why improvements in orientation sensitivity may or may not be observed psychophysically can come from considering the underlying neural events, to which we now turn. 

In principle, practice-based improvements in orientation sensitivity could reflect any or a combination of several specific underlying neural changes. For example, at the population level, neuronal responses with varying degrees of relevance to the task could be selectively reweighted with practice (Dosher & Lu, 1998; Saarinen & Levi, 1995). This practice-based reweighting could occur in many brain regions, although human fMRI data suggest that a particularly important area for differences between oblique and cardinal orientation sensitivity is V1 (Furmanski & Engel, 2000). A study that directly recorded macaque V1 neurons indicated that practice-based improvements in oblique orientation sensitivity coincided with steepened orientation tuning curves for the individual neurons most relevant to the trained orientation (Schoups, Vogels, Qian, & Orban, 2001). Intriguingly, the number of neurons responding to the trained orientation did not increase, and there was no increase in the signal-to-noise ratio (firing rate mean/firing rate SD) of individual neurons. Instead, the data suggested that the behavioral improvements were at least partially due to steepened orientation tuning curves. 

Schoups et al.'s (2001) practiced-based steepening in orientation tuning curves occurred over several months. Other physiological data indicate that orientation tuning curves can sharpen over remarkably briefer time intervals (Ringach, Hawken, & Shapley, 1997). Specifically, Ringach et al. (1997) showed that after an oriented stimulus is presented, orientation tuning curves for macaque V1 output layer neurons are initially flat but become increasingly steep across the subsequent tens of milliseconds. This time-dependent sharpening at the neural level appears to be paralleled behaviorally, according to a recent human psychophysical study (Matthews, Rojewski, & Cox, 2005). In that study, the precision of angular judgments increased significantly with masked stimulus durations ranging from ∼8 to ∼142 ms. At the briefest stimulus duration (∼8 ms), sensitivity to a 4 deg angular difference was already reliably (albeit modestly) better than chance but comparable at cardinal and oblique axes. The oblique effect—the superiority of cardinal versus oblique orientation sensitivity—emerged gradually across the subsequent tens of milliseconds. 

The recently reported orientation dynamics at the psychophysical (Matthews et al., 2005) and physiological levels (Ringach et al., 1997; Schoups et al., 2001) inspired this study. Specifically, we exploited the fact that at brief masked stimulus durations, orientation sensitivity is comparable at cardinal and oblique axes and relatively imprecise (Matthews et al., 2005)—perhaps due to shallower tuning curves (Ringach et al., 1997; Schoups et al., 2001). We reasoned that significant practice effects could occur, even at cardinal orientations, if the neural response could be intercepted at a sufficiently early point, that is, when there remains considerable opportunity for increased precision. Indeed, given the recently reported orientation dynamics, we allowed for the possibility that the practice effects might be manifest as an improvement in temporal resolution—not just angular resolution—at brief stimulus durations. Accordingly, unlike the earlier perceptual learning studies, which used stimulus durations of 500 ms or more (Matthews & Welch, 1997; Vogels & Orban, 1985), the training regimen in the present procedure comprised stimulus durations between just ∼8 and ∼158 ms. 

To briefly summarize the results, we found that orientation sensitivity hastened significantly with practice and that this improvement in temporal resolution occurred at cardinal and oblique axes alike. Additionally, the hastening was associated with sharpened angular resolution at each axis. This contrasts markedly with the absence of cardinal practice effects in previous perceptual learning studies that used longer stimulus durations. The difference implies that perceptual learning may require sufficient imprecision in the neural response before training begins. 

Methods

Apparatus, stimuli, and task

Some portions of the method used in this study were identical to those of Matthews et al. (2005). All portions of the method are described here for completeness. 

The experiment was conducted on a 21-in. (53.34-cm) ViewSonic P225 monitor that was controlled by a Macintosh G4 computer with a 733-MHz processor and software from the psychophysics toolbox (Brainard, 1997; Pelli, 1997). The vertical refresh rate of the monitor was 120 Hz. The spatial resolution of 1,024 × 768 pixels was such that a 1 × 1 cm square subtended 27 × 27 pixels. In a well-lit room, participants viewed the monitor through a circular tube that eliminated external cues to orientation, such as the monitor's borders, and had a 15-cm inner diameter. A chin rest helped to stabilize head position at 57 cm from the monitor. 

The discriminanda were Gabor patches, created by multiplying a sinusoidal luminance profile by a two-dimensional Gaussian envelope. Each Gabor patch was preceded and followed by a circular bull's-eye mask, that is, a radial square wave. All Gabor patches and masks were 8 deg in diameter. The spatial frequency of the Gabor patch was 1 cycle/deg, which was also the fundamental frequency of the bull's-eye mask. Both the Gabor patches and the masks had maximum (108.00 cd/m ²) and minimum (5.83 cd/m ²) luminances that rendered high contrast (Michelson contrast = 89.76%) within the apparently gray surround (56.91 cd/m ²). To eliminate positional cues that covary with changes in orientation, we randomized the phase of each Gabor patch. The phase of the bull's-eye masks alternated between light center and dark center, as pilot experiments indicated that orientation judgments were most disrupted when opposite-polarity masks preceded and followed each Gabor patch. Participants foveally viewed the stimuli, and a light, circular fixation dot approximately 11 arcmin in diameter (77.83 cd/m ²; 15.53% contrast with the surround) helped to stabilize eye position. 

The stimulus sequence is shown schematically in Figure 1. On every trial, two new Gabor patches were generated and presented successively at slightly different orientations. The specific orientation differences and the duration of the Gabor patches varied and will be detailed in the Procedure section. Regardless of those variations, however, each mask was 8.33 ms (one frame) in duration and the interstimulus interval was always 500 ms. On each trial, the participant's task was to report whether the second orientation was “clockwise” or “anticlockwise” to the first. 

Figure 1

 

Each trial consisted of two sequentially presented and foveally viewed gratings that were preceded and followed by opposite-polarity bull's-eye masks designed to limit oriented neural persistence from the gratings. The gratings differed slightly from each other in orientation, and participants judged the orientation of the second grating to be clockwise or anticlockwise to the first. The correct response is anticlockwise in the schematic above, where the magnitude of the angular difference and the spatial frequency have been selected for ease of viewing. The actual angular differences and spatial frequency are detailed in the Methods section.

Participants

Denison University's Human Subject Committee approved the study. Twenty naive adult participants from the Denison University community completed the study. Each participant had normal or corrected-to-normal vision. 

Procedure

Participants were instructed to make their clockwise or anticlockwise judgments as quickly as possible without sacrificing accuracy. To promote accuracy, participants proceeded at their own pace, initiating each trial with a button press when ready. Auditory feedback informed the participants whether their response was correct or incorrect immediately after each response to maintain motivation, and the computer announced the overall percentage of correct responses after each trial block. Although learning on lower level visual tasks can occur without feedback (Fahle, Edelman, & Poggio, 1995; McKee & Westheimer, 1978), practice effects can be larger when correct error feedback is provided (Herzog & Fahle, 1997). 

Collectively, the 20 participants completed 110,000 trials. Within a 2-week period, each participant completed 5,500 trails that were spread across seven 1-hr sessions. The seven sessions—each conducted on a different day—consisted of a pretraining session (Day 1; 1,000 trials), five training sessions (Days 2–6; 700 trials per day), and a posttraining session (Day 7; 1,000 trials). These will be described in turn. 

Pretraining session (Day 1)

The pretraining session began with a screening that was designed to establish that the task was understood, that is, could be performed at greater-than-chance levels. During the screening, cardinal and oblique trials were randomly interleaved. Additionally, the screening comprised stages of increasing difficulty. Specifically, although the angular difference during the screening was always 10 deg, the stimulus durations became progressively briefer. Initially, each participant was required to make five consecutive correct responses at each of the following three stimulus durations before proceeding: 500, 200, and 50 ms. In those initial-screening trials, there were no masks. Subsequently, the masks were added (see Figure 1), and each participant was required to make five consecutive correct responses at a stimulus duration of 500 ms. Lastly, when the stimuli were masked and reduced to just 200 ms, each participant was required to make 10 consecutive correct responses—which could occur by chance only once in every 1,024 tries. Each of the 20 participants successfully completed this screening, indicating that any performance limitations before the actual trials began were sensory rather than conceptual. 

After successfully completing the screening, each participant proceeded to the actual pretraining trials. The pretraining trials were designed to determine the participant's 84% angular-difference thresholds. Thresholds for each participant were measured separately at the horizontal axis (cardinal condition) and the axis 45 deg anticlockwise to horizontal (oblique condition). At each of those axes, thresholds were determined using the method of constant stimuli and an ensemble of 10 angular differences that formed the following geometric series: ±0.75, ±1.5, ±3, ±6, and ±12 deg. These 10 angular differences were randomly interleaved and presented 10 times per block, resulting in 100-trial blocks. Each participant completed 10 such 100-trial blocks, with 5 cardinal blocks being randomly interleaved with 5 oblique blocks. Across the trials within each block, the orientation of the first grating was randomly jittered across a ±5 deg range around either the horizontal axis (cardinal blocks) or the axis 45 deg anticlockwise to horizontal (oblique blocks). This jitter was used to minimize the participant's ability to base judgments on implicit estimates of the primary cardinal and oblique axes, as such estimates would likely be better for the cardinal axis. Instead, the participants were forced to compare the two explicitly presented orientations to each other. The ±5 deg range was also sufficiently small to prevent the cardinal and oblique categories from overlapping with each other, even when the maximum jitter summed with the largest angular differences. On all pretraining trials, a bull's-eye mask preceded and followed each grating (see Figure 1), and each grating was presented for 200 ms (24 frames). 

For each participant at each axis condition, the 10 angular differences were plotted on the abscissa of a psychometric function while the ordinate reflected the proportion of clockwise responses (500 trials per psychometric function, per participant). A least squares procedure was then used to fit the data with a sigmoid of the form  

$1/(1+\exp [-K(X-X_0)\left]\right),$

(1)

where K and X ₀ determine the slope and midpoint of the sigmoid, respectively. The correlation between the best fitting sigmoid and the data, as indexed by the Pearson correlation coefficient ( r), was statistically significant ( p < .001) for each participant at each axis. Because each fit was significant, it was possible to fairly interpolate from the sigmoid an 84% angular-difference threshold, which was defined as half the angular difference required to alter the response rate from 0.16 to 0.84. 

Thresholds so obtained were then used to divide the participants into two equally performing groups whose mean psychometric functions are shown in Figure 2. Each of the four psychometric functions in the figure is based on (10 participants per group  × 500 trials per participant at each axis) 5,000 trials. Visual inspection of the figure reveals that the proportion of clockwise responses (±1 SE) increased monotonically in the cardinal (blue squares, solid line) and oblique (red Xs, dotted line) conditions for participants who would subsequently train on cardinal (left panel) and oblique (right panel) axes. The monotonic trends indicate that, during the pretraining session, the participants' limitations were sensory, not conceptual. Additionally, an oblique effect is visually evidenced by the slope difference within each group. In fact, mean pretraining cardinal thresholds were significantly lower than mean pretraining oblique thresholds for participants assigned to the cardinal group, t(9) = 4.58, p = .001, and for participants assigned to the oblique group, t(9) = 5.14, p = .001. Most importantly, pretraining cardinal thresholds were virtually identical in the cardinal (2.66 deg, ±0.28 SE) and oblique (2.61 deg, ±0.50 SE) groups, and pretraining oblique thresholds were virtually identical in the cardinal (5.14 deg, ±0.50 SE) and oblique (5.13 deg, ±0.48 SE) groups. Given this initial equality between the groups, any subsequently observed group differences could be fairly attributed to differences arising from the training sessions. 

Figure 2

 

Pretraining psychometric functions. Psychometric functions from the pretraining session are shown separately for the cardinal (left) and the oblique (right) training groups. Within each group, at the cardinal axis (blue squares, solid lines) and the oblique axis (red Xs, dotted lines), the proportion of clockwise responses increased monotonically. The monotonicity implies that during pretraining, the participants' limitations were sensory, not conceptual. Each group also produced significantly steeper slopes (i.e., lower angular thresholds) for cardinal stimuli than for oblique stimuli, indicating an oblique effect in each group before training began. At each axis, pretraining slopes were virtually identical in the two groups. Consequently, any between-group differences found in the posttraining session would be due to the axis-specific training regimen. Click here to view the corresponding auxiliary file for this figure.

The thresholds observed here were approximately twice those observed by Matthews and Welch (1997). This difference is most likely attributable to the fact that the stimuli in the previous study (Matthews & Welch, 1997) were not masked and were shown for a much longer duration—500 ms versus the present 200 ms. Despite the approximately twofold difference in thresholds, however, each study revealed a strong oblique effect in angular acuity before training began. Specifically, the ratio of oblique to cardinal thresholds was approximately 2.0 in this study and 2.3 in Matthews and Welch (1997). 

Training sessions (Days 2–6)

Within each of the five training sessions, each participant completed 700 training trials that were divided into ten 70-trial blocks. Each 70-trial block consisted of 10 trials—half clockwise and half anticlockwise, randomly interleaved—at each of seven stimulus durations. The seven stimulus durations were also randomly interleaved within each block and ranged between 8.33 ms (1 frame) and 158.33 ms (19 frames), inclusively, in 25-ms (three-frame) steps. 

Across all five training sessions, the angular difference to be judged was tailored to each participant's performance in the pretraining session. In a pilot study on the training durations (8.33–158.33 ms), we initially presented an angular difference equal to each participant's 84% threshold, as measured with the 200-ms stimulus durations from pretraining. However, pilot participants regularly failed to respond above chance levels when that angular difference was presented at the much briefer training durations. Pilot data also showed that doubling the 84% threshold that was obtained at 200 ms generated a reasonable performance range at the much briefer training durations. Consequently, we adopted this procedure for the actual experiment. Specifically, participants in the cardinal group always trained on an angular difference equal to twice their own 84% pretraining threshold at the cardinal axis (mean 84% masked cardinal threshold = 2.66 deg). Similarly, participants in the oblique group always trained on an angular difference equal to twice their own 84% pretraining threshold at the oblique axis (mean 84% masked oblique threshold = 5.13 deg). Therefore, unlike in the pretraining session, where the angular differences varied while the stimulus duration was constant, throughout each participant's training sessions, the angular difference was constant and the stimulus duration varied. 

As in the pretraining session, visual persistence was limited by presenting a bull's-eye mask before and after each grating. Also, the orientation of the first grating was randomly jittered across a ±5 deg range around either the horizontal axis (cardinal group) or the axis 45 deg anticlockwise to horizontal (oblique blocks), as in pretraining. Unlike in the pretraining session, however, in the training sessions, participants were exposed either to near-horizontal gratings exclusively (cardinal group) or to near-diagonal gratings exclusively (oblique group). Except for the difference in the training axis, all aspects of the training sessions were identical for the cardinal and oblique groups. 

Posttraining session (Day 7)

The posttraining session was identical to the pretraining session in all aspects except that the screening procedure was dropped from the posttraining session. The posttraining orientation sensitivity was important for addressing the issue of whether any learning that may have occurred in the training sessions was general or sensory. On the one hand, one would expect posttraining increases in orientation sensitivity to be comparable at the trained and untrained axes if the learning were general. On the other hand, sensory learning would be evident to the extent that posttraining increases in orientation sensitivity were specific to the trained axis. 

Data analysis

Hastening orientation sensitivity

The primary data analysis focused on the issue of whether fine orientation sensitivity could be hastened with practice. Such hastening would be evidenced by training-induced decreases in threshold duration, operationally defined as the stimulus duration required to achieve a given level of angular resolution. Here, we detail the steps used for estimating threshold durations. 

Threshold durations were estimated separately for each of the five training sessions and each training group. For each threshold duration, the estimation process first required plotting the group's mean orientation sensitivity at each of the seven stimulus durations. At each stimulus duration, signal detection procedures (Green & Swets, 1966) were used to compute orientation sensitivity (d′) as follows: Hits and false alarms were operationally defined as clockwise responses made when the second Gabor patch was, respectively, clockwise or anticlockwise to the first. The standard deviation of the function that converted the proportion of hits and false alarms to z scores was set to 0.5. Consequently, d′ = 1.0 corresponded to 84% correct without response bias. A least squares procedure identified the best fitting power function that related orientation sensitivity (d′) to stimulus duration. In all cases, the power function provided a statistically significant fit (p < .001) to the data. Because each fit was significant, we could fairly interpolate the stimulus duration (X-intercept) corresponding to a given level of orientation sensitivity (d′), which we call the criterion sensitivity (d′). The criterion sensitivity (d′) for each group was set at a level that would minimize “floor” and “ceiling” effects. Specifically, for each group, the criterion sensitivity (d′) was operationally defined as the average of the d′ values empirically observed at the extremes of our training conditions. These extremes were the first day of training at the briefest stimulus duration—where d′ would likely be lowest—and the final day of training at the longest stimulus duration—where d′ would likely be greatest. The resultant criterion sensitivity (d′) values were 1.26 for the cardinal group and 1.41 for the oblique group. With those values in hand, we could track daily changes in threshold duration. Significant downward trends would signify hastening in orientation sensitivity (d′). 

Evaluating angular resolution

Angular resolution was evaluated in two ways. First, within-subject ANOVAs and r ² values were used to compare orientation sensitivity ( d′) in the first training session with that in the final training session. Such comparisons assessed the extent to which the training regimen sharpened angular resolution. Second, we evaluated the extent to which the training regimen may have sharpened angular resolution in an axis-specific manner. To accomplish this, we conducted within-subject ANOVAs and r ² comparisons on pre- versus posttraining orientation sensitivity ( d′) for each group at each axis. Comparable r ² values at the trained and untrained axes would suggest that the training produced a general form of learning. By contrast, axis-specific improvements—an indicator of sensory learning—would be evident to the extent that the r ² values at the trained axis exceeded those at the untrained axis. That is, large differences between the r ² values at the trained and untrained axes would implicate modifications limited to the neural events most responsible for orientation judgments at the trained axis. Finally, because the statistical comparisons of orientation sensitivity ( d′) were planned (a priori) and limited to the overall effect of training (averaging across the separate angular differences), the alpha level was not adjusted for multiple comparisons (Keppel, Saufley, & Tokunaga, 1992). 

Auxiliary data files

The raw data from individual participants can be found in the auxiliary files. Each auxiliary file bears the name of the corresponding figure or movie. Click on the corresponding images to view the auxiliary files. 

Results

Our main finding is the hastening of orientation sensitivity, which is readily seen in Figure 3. Visual inspection reveals that the threshold duration decreased monotonically with training in the cardinal group (left) and the oblique group (right). Indeed, for the cardinal group, the threshold duration decreased by 51.1% between the first training day (56.5 ms) and the last training day (27.6 ms), and the declining linear trend across all five training days was significant, r(3) = .978, p = .004. For the oblique group, the threshold duration decreased by 86.7% between the first training day (58.7 ms) and the last training day (7.8 ms), and the declining linear trend across all five training days was again significant, r(3) = .984, p = .002. In short, the data in Figure 3 indicate that fine orientation sensitivity at cardinal and oblique axes can be hastened with practice. 

Figure 3

 

Hastening orientation sensitivity. The hastening of orientation sensitivity is evident for the cardinal training group (left) and the oblique training group (right), as practice across the five daily training sessions generated significant decreases in threshold duration. Threshold duration was operationally defined as the stimulus duration associated with a specified level of orientation sensitivity ( d′) that was held constant across training sessions (see Methods section). The conventions here are the same as those in Figure 2. Click here to view the corresponding auxiliary file for this figure.

The hastening that is readily seen in Figure 3 coincided with practice-based improvements in angular resolution at the trained stimulus durations. This relationship is made explicit in Movie 1, where successive training days are shown on successive frames. Within each frame, the group's daily threshold duration is centered above the two graphs. The left graph cumulatively displays the threshold durations across the training days, as in Figure 3. The right graph displays the daily levels of orientation sensitivity ( d′ ± 1 SE) across the seven training stimulus durations, as well as the associated best fitting power functions. The power functions indicate that orientation sensitivity ( d′) increased significantly with stimulus duration on each day, r(5) ≥ .962, p < .001. Across days, the best fitting power functions slide upward along the ordinate, indicating that practice enhanced the participants' angular resolution. Correspondingly, on successive training days, there is leftward movement in the X-intercept, that is, the threshold duration (solid black “drop lines”). This dynamic relationship between the hastening (leftward-shifting X-intercepts) and the sharpening in angular resolution (upward-shifting power functions) occurred in the cardinal and oblique groups alike. The group similarities occurred although the criterion sensitivity in the cardinal group ( d′ = 1.26) differed from that in the oblique group ( d′ = 1.41), as noted in the Methods section. An additional analysis revealed that the practice-induced declines in threshold duration remained significant even after swapping the criterion sensitivity ( d′) values between the two groups. Overall, the results indicate that the practice-based hastening in temporal resolution and sharpening in angular resolution occurred whether the training axis was cardinal or oblique. 

 

Movie 1

 

Relation between hastening and angular resolution. For successive training days, the threshold durations from Figure 3 are dynamically replotted (left) and juxtaposed with the data set from which each daily threshold duration was interpolated (right). Threshold durations in the left panel correspond to the X-intercepts in the right panel. The panel on the right also shows that the relationship between orientation sensitivity ( d′) and stimulus duration is positive and well described ( p < .001) by a power function on each training day. Across training days, the best fitting power functions slide upward, whereas the drop lines that mark the threshold duration slide leftward. These practice-based improvements in angular resolution and threshold duration are evident for the cardinal group (blue squares, solid lines) and the oblique group (red Xs, dotted lines). Click here to view the corresponding raw data for this movie.

The sharpening in angular resolution, which occurred during the training sessions, is seen most directly in Figure 4, which contains data from the right panel of Movie 1, replotted in static view. Visual inspection immediately reveals that, for both the cardinal (left) and the oblique (right) groups, orientation sensitivity ( d′) is greater on Training Day 5 (closed green circles) than on Training Day 1 (open orange circles). ANOVAs confirmed that this training effect was significant in the cardinal group, F(1,9) = 5.728, p = .04, and in the oblique group, F(1,9) = 17.791, p = .002. Consistent with visual inspection, a comparison of r ² values indicated a larger training effect for the oblique group ( r ² = .664) than for the cardinal group ( r ² = .389) despite virtually identical performance levels in the two groups before training (see Figure 2). Still, the finding that cardinal orientation sensitivity ( d′) improved significantly (albeit more modestly) with practice is surprising given the fact that previous studies revealed no such improvements at cardinal axes after extensive training (Matthews & Welch, 1997; Vogels & Orban, 1985). 

Figure 4

 

Practice-based improvements in angular resolution. Orientation sensitivity ( d′ ± 1 SE) and the best fitting power functions for the cardinal training group (left) and oblique training group (right) are plotted for the seven stimulus durations that were presented during training. The training generated a significant sharpening in angular resolution, as orientation sensitivity ( d′) on Training Day 5 (solid green circles) significantly exceeded that on Training Day 1 (open orange circles) for each group. Click here to view the corresponding auxiliary file for this figure.

The sharpening in angular resolution, which occurred during the training sessions ( Figure 4), was specific to the trained axis. The specificity is visually evident in Figure 5, where the difference between the pretraining (open orange squares) and posttraining (closed green squares) data is readily apparent at the trained axes (top panels) but not at the untrained axes (bottom panels). Indeed, at the cardinal axis (left panels), orientation sensitivity ( d′) increased significantly with training in the cardinal group (top left panel), F(1,9) = 5.12, p = .05, r ² = .363, but not in the oblique group (bottom left panel), F(1,9) = 0.284, p = .607 ( ns), r ² = .031. Conversely, at the oblique axis (right panels), orientation sensitivity ( d′) increased significantly with training in the oblique group (top right panel), F(1,9) = 12.074, p = .007, r ² = .573, but not in the cardinal group (bottom right panel), F(1,9) = 0.321, p = .585 ( ns), r ² = .034. Stated another way, at each trained axis, the practice effects were significant and accounted for more than one third of the variance; by contrast, at each untrained axis, the practice effects were nonsignificant and accounted for less than 4% of the variance. Overall, the axis-specific improvements seen in Figure 5 argue against a general learning account that predicts practice effects that are not restricted to the trained axis. Moreover, the data are inconsistent with a simple pretest–posttest account of the improvements because the pre- and posttraining procedures were identical for conditions that generated improvement (top panels) and those that did not (bottom panels). 

Figure 5

 

Lack of generalization to untrained axes. Orientation sensitivity ( d′ ± 1 SE) from the pretraining (open orange squares) and posttraining (closed green squares) phases is plotted for the cardinal (left) and oblique (right) axes. From pre- to posttraining, orientation sensitivity ( d′) improved significantly at the trained axes (top panels) but not at the untrained axes (bottom panels). Additionally, the practice effect accounted for more than one third of the variance at each trained axis (top panels), yet less than 4% of the variance at each untrained axis (bottom panels). Given the axis-specific practice effects, it is unlikely that a general form of learning can account for the sharpening in angular resolution seen in Figure 4. Click here to view the corresponding auxiliary file for this figure.

Discussion

This study was conducted to provide new information about the angular and temporal resolution of our orientation sensitivity, using a perceptual learning procedure. We were motivated by several observations from the literature on perceptual learning and fine orientation sensitivity. First, the previous studies had addressed practice-based improvements in angular resolution but not temporal resolution (Matthews & Welch, 1997; Vogels & Orban, 1985). Those studies used relatively long stimulus durations (i.e., 500 ms or more) and indicated that angular resolution improved with practice but only at oblique axes. The oblique practice-based improvements may have been, in part, due to a steepening in the underlying orientation tuning curves, according to single-unit recordings from macaque V1 (Schoups et al., 2001). Single-unit recordings also indicate that orientation tuning curves are dynamic, gradually sharpening across the tens of milliseconds after stimulus presentation (Ringach et al., 1997). Those physiological dynamics were paralleled psychophysically in a recent study that indicated that orientation sensitivity improves across the tens of milliseconds after masked stimulus presentation (Matthews et al., 2005). Moreover, that study also demonstrated that orientation sensitivity is comparable (if modest) at cardinal and oblique axes at very brief masked stimulus durations. Synthesizing those observations, we attempted here to generate practice-based improvements, at oblique and cardinal orientations, by training at extremely brief stimulus durations. We reasoned that significant improvement could occur, even at cardinal orientations, if the neural response could be intercepted at a sufficiently early point, that is, when there remains considerable opportunity for increased precision. 

Our results on temporal resolution indicated that, at each axis, practice significantly hastened the visual system's response to subtle orientation differences. Specifically, the minimum stimulus duration needed to achieve criterion performance decreased by 51.1% at the cardinal axis and by 86.7% at the oblique axis. Angular resolution ( d′) also improved significantly at both the cardinal and oblique axes. Although significant practice effects at each axis would be consistent with a general learning explanation (e.g., improvements in attention), there are two reasons why such an explanation fails to account for the present practice effects. First, the practice effects did not generalize to the untrained axis. Second, the practice effects tended to be smaller at the cardinal axis than at the oblique axis. These observations are contrary to a general learning account that predicts no dependence on axis. By contrast, a sensory learning account predicts axis-specific learning, as well as greater learning at the oblique axis. This latter point follows from previous indications that there is greater internal noise—and, hence, greater opportunity for improvement—at oblique axes than at cardinal axes. That anisotropy has been inferred from external noise studies, to which we now turn. 

In a previous psychophysical study that used external orientation noise (Heeley, Buchanan-Smith, Cromwell, & Wright, 1997), participants judged subtle orientation differences while the orientation bandwidth of the stimuli was varied. The data were analyzed separately at cardinal and oblique axes with the following variance summation model: 

${\sigma }_{\mathrm{o}}=\sqrt{({\sigma }_{\mathrm{i}}^2+\frac{{\sigma }_{\mathrm{e}}^2}{N})},$

(2)

where σ_o is the behaviorally observed threshold, σ_e² is the variance of the external noise, N is the number of samples presumed to be taken by the visual system, and σ_i² is the variance of the internal noise. The model's optimum values indicated that internal noise was greater at oblique (σ_i² = 3.26) than at cardinal (σ_i² = 2.08) orientations. Comparatively high internal noise along oblique axes was also observed in a recent study on the angular resolution of motion sensitivity (Dakin, Mareschal, & Bex, 2005). In that study, participants judged subtle angular differences between two directions of motion, while the external directional noise was varied. A variance summation model similar to Equation 2 revealed that the internal noise for oblique directions was nearly twice that for cardinal directions. Taken together, these external noise studies indicate that the neural response along oblique axes is comparatively imprecise. Intriguingly, it may be that sufficient initial levels of neural imprecision are necessary for perceptual learning. This would explain why perceptual learning has been greater for oblique than cardinal orientations (Matthews & Welch, 1997; Vogels & Orban, 1985) and why perceptual learning across a variety of visual tasks tends to be greater when external noise is added (Fine & Jacobs, 2002). Here, we attempted to enhance perceptual learning, not by explicitly adding external noise but by exploiting the imprecise neural response associated with the brevity of our stimuli. 

Although the brevity of our stimuli—and the correspondingly imprecise neural response—may explain why cardinal practice effects occurred in this study but not in the previous studies (Matthews & Welch, 1997; Vogels & Orban, 1985), other explanations are also possible. For example, because masks were used in this study but not in the previous studies (Matthews & Welch, 1997; Vogels & Orban, 1985), the present practice effects may reflect an improved ability to disregard the masks. Two aspects of such practice-based mask inhibition are noteworthy. First, because the present practice effects did not transfer to the untrained axis, any practice-based mask inhibition in this study would, in itself, be a form of orientation-specific perceptual learning. That is, within the circular bull's-eye masks, only the orientation band most relevant to the trained orientation would have been disregarded with practice. Second, practice-based mask inhibition would effectively lengthen the stimulus duration—thereby permitting a more precise response (Matthews et al., 2005). Thus, the practice effects shown here could reflect a hastening of the direct response to the Gabor patches, or the equivalent of an extended stimulus duration arising from practiced-based inhibition of the masks, or both. 

Mask inhibition has also been associated with visual hastening on other tasks that have interesting similarities to the present fine orientation task. The most closely related studies are those that investigated perceptual learning on texture discrimination tasks (Ahissar & Hochstein, 1993; Karni & Sagi, 1991; Schwartz, Maquet, & Frith, 2002). The texture discrimination tasks, like the present fine orientation task, required an orientation-related response and contained brief masked stimuli. In one of the texture discrimination tasks (Karni & Sagi, 1991), practice significantly hastened the visual system's response to texture orientation, and this hastening was contingent on the orientation of the line elements within the surrounding mask. 

Despite the aforementioned similarities between this study and the previous texture discrimination studies (Ahissar & Hochstein, 1993; Karni & Sagi, 1991; Schwartz et al., 2002), there are two important stimulus differences. First, unlike the foveally presented stimuli here, targets in the texture discrimination tasks were presented in the periphery. Second, unlike the subtle angular differences (0.75–12 deg) tested here, the orientation differences to be detected or identified in the texture discrimination tasks ranged between 30 and 90 deg. It is unlikely that such large angular differences would generate practice-based hastening if foveally viewed. This is because, under foveal viewing conditions, even untrained naive participants can identify 12 deg orientation differences with 75% accuracy at masked stimulus durations as brief as 8.33 ms (Matthews et al., 2005). It therefore seems likely that the previous texture discrimination tasks, while important in their own right, tapped into neural events different from those mediating the present fine orientation task. Given these differences in the underlying neural events, it is intriguing (and perhaps coincidental) that the previous texture discrimination tasks and the present task each generated orientation-specific practice effects and a hastened sensory response. 

Supplementary Materials

Fig_2_Raw_Data - Supplementary Data 

Fig_3_Raw_Data - Supplementary Data 

Fig_4_Raw_Data - Supplementary Data 

Fig_5_Raw_Data - Supplementary Data 

Movie_1_Raw_Data - Supplementary Data 

Acknowledgments

This research was supported by Anderson Scholarship awards to Kristen Strong and Kei Kurosawa and by a Denison University Research Foundation award to Nestor Matthews. 

We thank an anonymous reviewer for the observation that practice-based inhibition to the mask would effectively lengthen stimulus duration. 

Kristen Strong and Kei Kurosawa contributed equally to this work. 

Commercial relationships: none. 

Corresponding author: Nestor Matthews. 

Email: [e-mail]. 

Address: Denison University, Department of Psychology, Granville, OH 43023, USA.