Spatial frequency tuning of perceptual learning and transfer in global motion

Perceptual learning is typically highly specific to the stimuli and task used during training. However, recently it has been shown that training on global motion can transfer to untrained tasks, reflecting the generalising properties of mechanisms at this level of processing. We investigated a) if feedback was required for learning when using an equivalent noise global motion coherence task, and b) the transfer across spatial frequency of training on a global motion coherence task, and the transfer of this training to a measure of contrast sensitivity. For our first experiment two groups, with and without feedback, trained for ten days on a broadband global motion coherence task. Results indicated that feedback was a requirement for learning. For the second experiment training consisted of five days of direction discrimination on one of three global motion tasks (broadband, low or high frequency random-dot Gabors), with trial-by-trial auditory feedback. A pre- and post-training assessment was also conducted, consisting of all three types of global motion stimuli (without feedback) and high and low spatial frequency contrast sensitivity. We predicted that if learning and transfer is cortically localised, then transfer would show specificity to the area processing the task (global motion). In this case, we would predict a broad transfer between spatial frequency conditions of global motion only. However, if transfer occurred as a result of backward generalisation, a more selective transfer would occur matching the low-pass broadband tuning of the area processing global motion. Our training paradigm was successful at eliciting improvement in the trained tasks over the five days. However, post-training transfer to trained or untrained tasks was only reported for the low spatial frequency trained group. This group exhibited increased sensitivity to low spatial frequency contrast, and an improvement for the broadband global motion condition. Our findings suggest that the feedback projections from global to local stages of processing play a role in transfer.

Perceptual learning has attracted much attention as a potential tool to aid 2 recovery of lost visual function for clinical populations [1]. The success of 3 perceptual training in amblyopia [2][3][4], presbyopia [4] and cortical damage [5][6][7] 4 has demonstrated sensory plasticity in adulthood. This evidence contradicts the 5 position that sensory development is restricted to a critical period early in 6 life [8,9] and that the visual system is hard-wired in mature systems [10]. While 7 it has repeatedly been established that training can improve perceptual 8 abilities [11], these benefits tend to be highly specific for both the perceptual 9 features of the stimuli [12][13][14] and the behavioural task used in training [15]. 10 This specificity severely limits the effectiveness of perceptual learning as a 11 general therapeutic tool. Resolving the conditions under which learning is tied 12 to the features and tasks used in training, and how much it can generalise to 13 new tasks and stimuli, is imperative for understanding the mechanisms of 14 perceptual learning [11,16]. Our study aims to identify the relative location of 15 one of the mechanisms involved when learning direction discrimination for a 16 global motion coherence task. We do this by evaluating the spatial frequency 17 tuning of improvements in performance for trained and untrained tasks. For the 18 purpose of this paper we dissociate between models and mechanisms of 19 perceptual learning using two general categories i) the How and ii) the What 20 and Where. Models of perceptual learning aim to resolve 'how' learning occurs. 21 Understanding which mechanisms underlie this learning address the 'what' and 22 'where' (and possibly 'why') questions. In this way we consider a mechanism a 23 feature or component of the learning model [17]. 24 Models of Perceptual Learning -How 25 The 'how' is often the primary focus for research and there are, broadly 26 speaking, two predominant models of perceptual learning. The first explains 27 perceptual learning in terms of a change in the neurons that code for that 28 feature [18,19]. The second position argues that learning is a result of the 29 change in weights of readout between the sensory representation and the 30 decision units [20][21][22][23][24]. The first position developed as a result of accumulating 31 evidence that improvements were highly specific for the features of the stimuli 32 used for training. Early studies identified that learning was specific for 33 orientation [13,25,26], spatial frequency [11,13,27], direction of motion [15], 34 retinal location [14,24,26,[28][29][30] of stimuli and the eye to which they are 35 presented [14,24,29]. On the basis of this feature-specificity it has been argued 36 that the underlying brain area responsible for the learning process must be 37 within the primary visual cortex [14] where the receptive fields of cells display a 38 high degree of specificity. The implication of this position is a potential degree 39 of plasticity in individual neurons within an area previously thought to be 40 incapable of structural change [14,19,31]. Although the failure of perceptual 41 learning to transfer across retinal location, orientation and other stimulus presented [61]. Zhang et al. (2010) [61] proposed a rule-based learning model to 107 account for the transfer, where higher level decision units learn and re-weight 108 the V1 inputs. They proposed that the absence of functional connections to the 109 untrained orientation or location prevents any potential reweighting. Double 110 training, through exposure, activates the functional connections at the new 111 locations or orientations, enabling transfer. This suggests that there may not be 112 a straightforward correspondence between stimulus features and the neural loci 113 involved in performing the task. Double-training has been a successful paradigm 114 in obtaining evidence of transfer [59,69,70], however task difficulty during 115 training, and observer confidence still play a vital role in the learning and 116 transfer process [34,71]. An early theoretical model of transfer is the Reverse 117 Hierarchy Theory [72]. This model proposes that transfer is a top-down process, 118 and the degree of transfer is dependent on the characteristics of the receptive 119 fields involved in performing the training task. Transfer occurs as a result of 120 modification of neurons found in the higher cortical levels, where receptive fields 121 generalise. In contrast, specificity, occurs at the lower cortical levels where the 122 receptive fields are localised [72]. As established, learning is highly specific to 123 the features of the stimulus and reflects the tuning properties of the relevant 124 receptive fields. However, what specificity is expected for tasks that are not 125 processed in early visual areas? For example, perceptual learning improves 126 detection and discrimination in tasks using global form [73] and global 127 motion [46,74,75]. These kinds of tasks are known to be processed at higher 128 levels of the visual cortex, for example areas V3 [76], V4 [73] and V5 [77,78]. 129 The Reverse Hierarchy Theory predicts that the feedback connections from 130 higher levels back to early visual processing areas may be involved in facilitating 131 learning that transfers, and that the key to understanding specificity and 132 November 28, 2018 4/51 transfer lies in the hierarchy of processing and the feedforward and feedback 133 connections between them [33,72]. Based on this, tasks that require higher level 134 integration or segregation (global processing) may therefore play an important 135 role in creating learning that generalises to other tasks and stimuli [33]. 136 Local vs Global Processing 137 The receptive fields of neurons in higher cortical areas integrate information to 138 represent global stimulus properties [19,[79][80][81][82]. Global processes are investigated 139 using stimuli or tasks that can only be resolved through integration and 140 segregation of coherent or conflicting information [81][82][83]. Cells higher in the 141 processing cortical hierarchy are involved in the perception of global aspects of 142 an image and generalise across individual features such as spatial frequency and 143 location. Based on the predictions of the Reverse Hierarchy Theory, modification 144 of these generalising receptive fields may produce perceptual learning that also 145 generalises over these stimulus parameters. Since perceptual learning occurs for 146 both global motion [46,75] and form tasks [83,84], this suggests that learning is 147 not restricted to the initial encoding of information in V1, and can occur at 148 higher levels of cortical processing. Learning that involves higher level global 149 aspects of perception is of particular interest since it has the potential to 150 produce more generalisable improvements [33]. Huxlin [36,44] suggest using perceptual learning experiments in which the sensory 160 representations required for two tasks are the same, but the decision stages 161 differ, or where both sensory representations and the decision stage are common 162 to both. Training using a global motion task has been found to transfer between 163 eyes [15,46]; increase sensitivity for detection and discrimination [5]; help 164 recover some of the blind field for cortically blind subjects [5]; and reduce 165 contrast thresholds for drifting stimuli [75]. Transfer of learning has also been 166 reported in a control population [75], where a post-training improvement was 167 found for contrast sensitivity. This is a particularly interesting result when 168 taking into account that a globally processed trained task improved contrast 169 detection, which is known to be processed in an early visual locations such as 170 4Cα, where cells are tuned for spatial frequency but not orientation [27]. The The perception of motion is hypothesised to occur as a two stage process [80,85]. 175 At the first stage, spatial frequency-and orientation-tuned mechanisms in V1 176 encode the motion signals that occur locally within the receptive fields of 177 individual neurons [54,81]. However, local and global stages of motion 178 processing differ in their spatial frequency tuning. In order to process more 179 complex motion, ambiguous or conflicting signals from the first stage need to be 180 integrated over a wider spatial area to provide a global representation of motion 181 and velocity [86,87].

182
A number of areas within the visual cortical hierarchy play a functional role 183 in processing motion. Areas V2 and V4 have a role in processing moving 184 orientation signals [88,89]. V3A also plays a role in several aspects of motion 185 processing [62], with 76% of neurons being selective for orientation and 40% 186 showing strong direction selectivity. However, evidence from lesion 187 studies [90,91], extra-cellular recordings [77,92] and neuroimaging in 188 humans [93] support area V5 as a brain area that is heavily involved in 189 processing global motion [77,85,[90][91][92]94]. Most neurons in V5 are strongly 190 direction selective [93,95], and the evidence for the role it plays in spatially 191 integrating motion signals is well supported by non-human primate data, and 192 neuroimaging studies in humans [47,77,96,97]. Receptive fields in V5 can be up 193 to tenfold larger than those in V1 [65], with broad spatial frequency and 194 orientation tuning, allowing them to sum the responses of V1 neurons, across 195 space, orientation and spatial and temporal frequency [49]. Psychophysical studies have shown that global motion detectors have relatively 198 broad spatial frequency tuning [80]. This is consistent with single-cell recordings 199 from area V5 in marmoset monkeys that exhibit bandpass spatial and temporal 200 frequency tuning, with a preference for low spatial frequencies [96]. Amano et al. 201 (2009) [79] suggested that, within V5, there is a "motion-pooling mechanism" 202 with broadband, low-pass tuning. Pooling of visual sensory information has also 203 been found in studies investigating transfer in binocular stereopsis [98]. The 204 neural mechanisms of stereoscopic vision are also known to be processed on 205 multiple levels of the visual hierarchy [99], and the pooling of information across 206 spatial frequency mechanisms has been proposed as an important step in 207 estimating depth [100][101][102][103]. In the same way, pooling across low frequencies is 208 likely an important step for integration and segregation of coherent motion.
Feedback connections are argued to be fundamental to efficient cortical 216 organisation, and of all the feedback systems terminating in V1, the connections 217 from V5 to V1 have been found to cover the most territory [105]. Furthermore, 218 these feedback (or re-entrant) connections terminate in different combinations 219 within various layers throughout V1 [65,105] and are rapidly updated [106,107]. 220 Recordings from macaque monkeys have suggested that there is almost no delay 221 for information processed in V5 to be fed back to lower areas, and it has been 222 proposed that the feedback from V5 is present prior to the bottom up 223 information from the feedforward connections [106,107]. This has recently been 224 supported by evidence of an independent motion pathway in humans, finding a 225 direct link from lateral geniculate nucleus (LGN) to V5 [108]. Projections are 226 targeted retinotopically to V1 in locations lying within the V5 receptive 227 field [65]. Thus, as V5 responses are influenced by the bottom-up responses 228 from V1, the top-down responses from V5 shape the responses of the V1 cells 229 which provide the sensory input [65]. Sillito et al. (2006) [65] suggest that the 230 "iterative interaction" between the two stages could account for the selectivity at 231 both levels. Backward projections are a theoretically plausible route for 232 perceptual learning, and are consistent with the predictions from the Reverse 233 Hierarchy Theory. Overall this suggests a theoretically important role for 234 top-down information in obtaining transfer from global motion [105].

235
Transcranial Magnetic Stimulation (TMS) has been used to stimulate the 236 re-entrant connections from V5 to V1, enhancing the perception of global 237 motion [109]. Using a novel paired cortico-cortical TMS protocol (ccPAS) to 238 induce Hebbian plasticity, observers' thresholds for motion detection were 239 reduced when the feedback connections from V5 to V1 were stimulated. 240 However, the improvement in perception was critically dependent on the timing 241 and direction of stimulation. There was no change when the feedforward 242 connections from V1 to V5 were stimulated [110]. This also suggests that these 243 re-entrant connections are malleable [110]. Using a similar method, a 244 direction-selective improvement was induced by pairing subthreshold stimulation 245 with the simultaneous presentation of direction-specific moving stimuli [111].

246
This provides additional support for the accumulating evidence that the 247 re-entrant connections from direction-tuned neurons in areas such as V5 play a 248 role in perceptual learning in global motion coherence tasks.

249
Our Study 250 Given the differences in spatial and temporal frequency tuning of processing 251 between the local and global levels [50,77,96], measuring the tuning of training 252 for these dimensions allows us to understand the role of each level in global 253 motion learning.

254
Prior to collecting the data for our main study, we also questioned whether 255 trial-by-trial feedback was a requirement for learning. Therefore, we first 256 investigated the necessity of feedback for perceptual learning to occur for our 257 specific stimuli. The specific nature of feedback and its role in perceptual learning is unclear, 260 and external feedback has been shown to improve learning and increase 261 efficiency [112,113]. However, some studies have found that learning occurs 262 without external feedback [11,15,30,36,78,114]. Recently, it has been found that 263 interleaving high accuracy (easy) trials and low accuracy (difficult) trials 264 resulted in perceptual learning without the need for feedback, even on difficult 265 trials [114]. Based on these results we predicted that we should find learning in 266 both conditions, as long as easy and difficult trials were interleaved. As detailed 267 in the following sections, our study found that learning did not occur in the 268 condition where no feedback was provided, even with easy trials. Learning only 269 occurred for the feedback condition. With this in mind our design for detectors in areas such as V5 and V3/V3A. There would thus be most transfer 288 when the training and test stimuli contain low spatial frequency components 289 (robust transfer is predicted from low frequency to low frequency conditions; 290 modest transfer from low to broad frequency stimuli, and from broad to low and 291 broad stimuli, and an unlikely but possible transfer from broad to high and high 292 to broad.). No transfer to contrast sensitivity would be predicted, since learning 293 would result in changes in weightings between direction-tuned global motion 294 mechanisms, and higher-level decision stages. If transfer occurs broadly across tasks and levels in a top down manner, as 298 predicted by the Reverse Hierarchy Theory [33], we would predict transfer to be 299 restricted by the frequency tuning of the global motion detectors, but to transfer 300 more broadly across tasks. We would expect most transfer when both training 301 and test stimuli contained low spatial frequency components (robust transfer 302 from low to low frequency stimuli; modest transfer from low to broad, and broad 303 to low and broad to broad, with an unlikely but possible transfer from high to 304 broad, and broad to high). We would also expect transfer to contrast sensitivity, 305 and for this to show the same low-pass spatial frequency tuning. This would 306 also be consistent with the transfer found by Levi et al. (2015) [75].

307
Finally, a mid frequency orientation discrimination condition was included as 308 a control task. Since none of the motion training stimuli contain any orientation 309 information, we predicted no transfer to occur from any trained condition or 310 either level of processing.   Global motion stimuli contained 100 Gaussian elements, each with a standard 333 deviation of 6.8 arc minutes (see Fig. 2a). Elements were presented within a Fig 1. (a) Expected improvement from training: We would expect robust improvement on frequency specific motion stimuli over the 5 days training, and for this to be reflected at the post-assessment as an improvement compared to the pre-assessment results. Predictions of transfer (b) Cortically localised global reweighting: In this case we may predict that transfer from frequency specific global motion would reflect the broad, approximately low-pass spatial frequency tuning of global motion detectors in areas such as V5 and V3/V3A. There would thus be most transfer when both training and test stimuli contained low spatial frequency components (c) Backward generalisation: Should transfer occur as a result of the backward projections from V5 to V1 (as predicted by the Reverse Hierarchy Theory [33]) we would therefore also expect robust transfer from low to low frequency stimuli; modest transfer from low to broad, and broad to low and broad to broad, with an unlikely but possible transfer from high to broad, and broad to high. Importantly, should there be transfer to contrast sensitivity we would expect this to reflect the low-pass spatial frequency tuning of global areas. mid-grey rectangle measuring 17.6 • x 17.6 • on a mid-grey background. Elements 335 moved 5 pixels per frame, and each element moved a fixed distance of 8.8 • . Dots 336 wrapped around the rectangle when approaching the edges. VIEWPixx/3D 23.6 inch monitor with a display resolution of 1920 x 1080 pixels, 357 with a 120 Hz refresh rate, using a Dell Precision T3610 PC running Windows 7. 358 One pixel subtended 1.6 arc minutes and stimuli were viewed from a distance of 359 570mm. Head position for testing was stabilised using a chin rest. Following 360 this, all observers undertook five consecutive days of global motion training in 361 one of three spatial frequency groups (broad, high or low). Training stimuli were 362 presented on a 19" monitor with a display resolution of 1980 x 1080 pixels and 363 60 Hz refresh rate, using a PC running Windows 7. One pixel subtended 1.7 arc 364 minutes. Stimuli were viewed from a distance of 500mm. All stimuli were 365 presented for 1 second.

366
Training Stimuli 367 Global Motion: Broadband stimuli were the same as previously described. For 368 low-frequency stimuli, the elements were circularly symmetric Gabor patches. the spatial frequency of luminance modulation, f , was 1 cycle/degree. For each 371 element, the luminance profile was defined as a function of horizontal and 372 vertical position (x, y) as: where x 0 and y 0 is the central position of the element, and A determines the 374 contrast. Elements for the high frequency stimuli were defined in the same way, 375 but had a standard deviation of 7.48 arc minutes and a spatial frequency, f , of 4 376 cycles/degree. For all stimuli, the spatial frequency and the speed of motion  Contrast) were presented. 396 3. Orientation Discrimination: Gabor patches were presented for 100ms on a 397 mid-grey background, measuring 7.9 • x 7.9 • , and presented 6.6 • either 398 above or below the central fixation point. Orientation was centred at zero, 399 at which point lines were vertical. There was a total range of 1.8 • 400 difference in tilt, between ± 0.9 • , with 7 linearly spaced points within the 401 range. The spatial frequency of the Gabor patch was 1.85 cycles/ • and the 402 standard deviation of the Gaussian envelope was 0.2 • (see Fig. 4c).

403
Contrast was fixed at 50% Michelson Contrast.  Measures were taken for global motion (high, broad and low spatial frequency), 411 contrast sensitivity (high and low spatial frequency) and orientation 412 discrimination (angle of tilt from the vertical) of an oriented Gabor patch.

413
Responses were captured on the DataPixx response box for contrast sensitivity, 414 and left and right arrows on the keyboard for global motion and orientation 415 discrimination. The presentation order of trials was randomised for direction 416 and coherence (global motion), orientation and contrast (contrast sensitivity) or 417 orientation (orientation discrimination). There were 20 repetitions for each of 418 the seven levels, for each condition. Testing was performed in a darkened room, 419 before and after training. those from different observers) that allow inference to a larger population [120]. 430 The linear component of the model is given by: where α and β are the intercept and the slope of the fixed effects parameters 432 respectively. The link function converts the expected outcome variable (the 433 proportion of correct responses, p), to the linear predictor [121]. Here, a logit 434 function was used, adapted to take account of chance performance, such that: Selecting the model: All response data were analysed using a GLMM likely to be true values, thus providing a fit closest to reality [120]. The best 445 regression model will provide the lowest value of AIC.
R is the observer response (proportion correct) T is time, the day or session 447 of testing, L is the stimulus level (degree of coherence, level of contrast, or 448 degree of orientation) and O is the observer. To verify the spread of easy and difficult trials across the stimulus levels 459 suggested by Liu et al. [114], accuracy for the first day was evaluated across all 460 observers for each level of the stimulus and is reported in (Table 1). The 461 accuracy across levels confirms that there is an even distribution of easy trials 462 (85% and above) and difficult trials (65% and below) with the balance around 463 75%. and an interaction between these two predictors (see Table 2). For the Feedback 472 Group, there was a significant negative effect of day however there was a 473 significant positive day × coherence interaction indicating there was a change in 474 slope across the 10 days. The same analysis was undertaken for the No 475 Feedback group, and found a main effect of day with a significant negative 476 interaction. The no feedback group performed worse after 10 days training (see 477 Fig. 5(b)). Performance over the 10 days is illustrated in Fig. 5 The number of correct responses from the daily training was calculated for each 486 observer, for each day and level of coherence. Training data were analysed 487 independently for the three groups trained on different frequencies (Broad, Low, 488 High) and were modelled as a fixed effect of day and coherence and an 489 interaction between these two predictors.

490
Results were analysed using the method previously described (see Table 3). 491 Performance improved for the broad and low trained groups and there was a 492 significant positive coherence-by-day interaction found for both conditions. For 493 the group trained with high spatial frequency stimuli there was a significant 494 main effect of day. This shows that there was a significant increase in the total 495 number of correct responses over the five days (see Fig. 3). These results thus 496 show an improvement in performance during the training phase for all three 497 spatial frequency conditions.   Learning is often measured through monitoring performance at a particular 500 threshold, which is expected to shift the psychometric function leftwards if 501 performance is improved, (see Fig. 7(a)). The psychometric function describes 502 performance in terms of accuracy as a function of the strength of the 503 stimulus [122], which are usually positively correlated, and it is expected to 504 reach asymptotic performance at the highest stimulus intensity [123]. Inspection 505 of the pre-and post-assessment data revealed that in some conditions 506 performance did not reach perfect accuracy, asymptoting at a proportion of 507 correct responses that was less than 1; this resulted in a poor psychometric fit of 508 the observer response data using the GLMM. To accommodate this, a nonlinear 509 generalised mixed effects model (NLME) was used to include an additional 510 parameter in order to model variability in the asymptotic performance at high 511 signal levels, as has been applied in other perceptual learning studies [124][125][126]. 512 Nonlinear regression analysis 513 The nonlinear regression provides three measures to assess a change in 514 performance over time. Firstly, like the GLMM a leftward shift in the curve 515 indicates an improvement in threshold (7(a)). An increase in slope indicates an 516 increase in the rate at which performance increases with increase signal level 517 (7(b)). Finally a change in the asymptote indicates a significant change to the 518 performance at the highest level of stimulus intensity (7(c)). These changes are 519 independent aspects of the psychometric function fit, and may not necessarily be 520 congruent. For example, it is possible to obtain an increase in one measure and 521 a decrease (or no change) in another.

522
Analysis of the pre-and post-assessment data was undertaken using a where p is the proportion of correct responses, A determines the asymptotic 526 level of performance, K defines the slope and C 0 defines the threshold. d is the 527 post-training change in performance, where A d , K d and C 0d determine the 528 change in asymptote, slope and threshold respectively. C is the coherence level, 529 and S is a dummy variable, taking on values of 0 (pre) or 1 (post) training We predicted that should transfer occur cortically, this would be located at the 537 global motion processing level, and transfer across frequencies would most likely 538 occur for stimuli with frequency properties that reflect the broadband low-pass 539 frequency tuning of global motion detectors (see Fig 11(a)). Firstly, we Bottom: Change statistics for the asymptote, threshold and slope, respectively. Plots show the median performance and the 95% confidence intervals for the change in performance between pre-and post-assessments. The red horizontal line at zero represents no change, confidence intervals crossing the zero line reflect no significant improvement. Points above the reference line show an improvement in performance and those below reflect a decrease in performance. Bottom: Change statistics for the asymptote, threshold and slope respectively. Plots show the median performance and the 95% confidence intervals for the change in performance between pre-and post-assessments. The red horizontal line at zero represents no change, confidence intervals crossing the zero line reflect no significant improvement. Points above the reference line show an improvement in performance and those below reflect a decrease in performance. Bottom: Change statistics for the asymptote, threshold and slope respectively. Plots show the median performance and the 95% confidence intervals for the change in performance between pre-and post-assessments. The red horizontal line at zero represents no change, confidence intervals crossing the zero line reflect no significant improvement. Points above the reference line show an improvement in performance and those below reflect a decrease in performance.
trained frequency and (ii) robust transfer to broad spatial frequency global 548 motion.

549
• For the high frequency trained condition we predicted; (i) moderate 550 learning on their own trained frequency and (ii) an unlikely but possible 551 transfer to the broad frequency condition.

552
Conversely, should transfer occur as a result of backwards generalisation 553 using the re-entrant connections from global motion processing areas to V1 we 554 would predict that transfer would reflect the pooling of spatial frequencies and 555 likely attenuation of high spatial frequencies. Transfer to global motion: broad-frequency trained group 570 We had predicted robust to moderate improvement for the broad trained group, 571 however this group showed no consistent transfer to trained or untrained 572 conditions (see Fig. 8). The post-assessment on the trained task found no 573 significant change in slope (proportion correct as a function of frequency) or 574 threshold (coherence required to obtain 75% correct), and a significantly lower 575 asymptote (proportion of correct responses at the highest coherence). Assessing 576 transfer to untrained tasks, there was no significant change for any measure 577 (intercept, slope or asymptote) for the low frequency test. Finally for the 578 high frequency test, there was no change in slope, a significant reduction in 579 asymptote (indicating worse performance at higher coherence), and a small but 580 significant reduction in threshold. The increase in the proportion of correct 581 responses at the lower coherence levels was the only significant improvement for 582 this group, however neither of the other two measures were consistent with an 583 improvement.

584
Transfer to global motion: low-frequency trained group:

585
The low-frequency trained group was the only condition to exhibit transfer to 586 trained and untrained motion conditions (see Fig. 9). In the post-assessment on 587 November 28, 2018  Fig. 1. We predicted a) that if transfer was local to the processing mechanism transfer would occur for global motion conditions with shared spatial frequency properties. (b) Performance improved for all groups during the five days training, with robust improvement for the low and broad trained groups, and moderate improvement for the high frequency group. However, at the post-assessment stage, when tested without the presence of trial by trial feedback, neither the broad nor high groups reflected a change from the pre-assessment measures. Improvement was restricted to the low trained group. (c) When assessing transfer to untrained frequencies, only the low spatial frequency group showed evidence of transfer, and performance improved for broad frequency motion. In addition, the low spatial frequency trained group was the only group to show an improvement on contrast sensitivity and this was exclusive to the low frequency contrast condition.
the trained task, while there was no significant change to the slope (proportion 588 correct as a function of frequency) there was significant reduction in threshold 589 (coherence required to obtain 75% correct), and a significantly higher asymptote 590 (proportion of correct responses at the highest coherence). Assessing transfer to 591 untrained tasks, there was a modest improvement for the broad frequency 592 test reflected by the was a significant increase in asymptotic performance.

593
However, there was an increase in threshold and no change to the slope. This 594 shows reduced sensitivity at low signal levels, but an increase is performance at 595 higher levels. Finally, the high frequency test stimuli show a small reduction 596 in threshold but a significantly shallower slope. Suggestive of increased 597 performance at lower coherence but decreased performance at higher stimulus 598 coherence. There was no change in asymptotic performance, and overall no 599 evidence of transfer. The spread of scores for the broad/broad and broad/high condition show that 611 one observer performed particularly poorly at higher coherence levels, but not 612 persistently so. Therefore, the lack of overall improvement for the broad trained 613 group is unlikely to be as a result of an outlier. However, the significantly worse 614 performance in asymptotic performance may, to some degree, be as a result of 615 this one individual.

616
Transfer to global motion: Summary

617
In summary, only the low frequency trained group provided reliable evidence for 618 transfer to trained and untrained motion conditions, with a moderate transfer to 619 the broadband frequency condition.

Fig 12.
Change scores (the difference between the number correct on the pre-and post-assessments) are plotted for each individual. Each observer is illustrated by a colour grouped by training and test condition (broad, low and high spatial frequency motion). Of particular interest is the broad trained group, which had a decline in performance post training. One observer performed worse at higher coherence levels after training, which explains the decline in asymptotic performance for the broad trained group. However, the overall spread of data suggest that on average post-assessment results did not differ significantly.
Transfer to contrast sensitivity 621 We predicted that, should contrast sensitivity improve as a result of training on 622 global motion it could only occur as a result of backward generalisation, as the 623 two tasks are processed in cortically separate locations. We predicted transfer 624 would be to limited conditions containing low frequencies, namely the broad and 625 low trained groups.

626
Data were analysed for all but the same 2 observers who did not complete 627 the final session using the generalised linear effects model previously outlined. 628 The measures of analysis are the intercept and slope, and an interaction between 629 the two predictors. For low spatial frequency contrast there was a a significant decrease in intercept 632 (fewer correct responses at the lowest coherence), but significant increase in slope 633 (an increase in the proportion of correct responses as a function of frequency). 634 Because of this ambiguity we plotted 75% thresholds for pre and post 635 performance (see Fig. 13 top left). Post-assessment thresholds shifted rightwards 636 suggesting worse performance overall.

638
The group trained with low frequency global motion showed a significant 639 decrease in intercept, and a significant increase in slope. Again, we plotted 75% 640 thresholds for pre and post performance (Fig. 13 left middle). In this instance 641 the steeper slope, and leftward shift of the threshold, suggests improvement High-frequency trained group:

644
The high frequency trained group showed no significant change for any condition 645 (see Fig. 13 bottom left).

647
This task was added as an untrained control, for which we predicted no 648 improvement, as no orientation information was present in any of the training 649 stimuli. Data were analysed for all but 3 observers who did not complete the 650 final session, using the generalised linear effects model previously outlined. The 651 measures of analysis are the intercept and slope , and an interaction between the 652 two variables. For analysis, the orientation levels were converted to absolute 653 values. Results are plotted in Fig. 14, and the fixed effects are listed in tab 5. 654 Neither the broad-or low-frequency trained group showed any significant change 655 on this task. However, the high spatial frequency trained group showed a 656 significant decrease in intercept over session, and a significant increase in slope 657 following training. However, the 75% threshold at post-assessment shift 658 rightwards, suggesting worse performance overall.

660
Damage to the visual cortex, as a result of stroke or other brain injury can result 661 in dramatic changes to connectivity between areas. It has been proposed that 662 and may even create novel connections [1][2][3][4][5][6][7]. While specificity is a limitation to 664 the effectiveness of perceptual learning as a tool for therapy, there is evidence 665 that this specificity is reduced for visually impaired populations [127][128][129][130][131][132]. 666 Huxlin et al. (2009) [5] proposed that using global motion may tap into "islands 667 of activity" within V1 through the feedback connections from V5 to V1.

668
However, it is still unclear how much the brain is able to compensate for 669 damage, and whether recovery involves building new connections or changes in 670 the functional connectivity within the existing pathways [133]. Understanding 671 how the cortical hierarchy is organised in terms of the nature of the feedforward 672 and re-entrant pathways is central to developing theories of perception [134]. The purpose of the first experiment was to establish the necessity of 675 performance feedback when undertaking a period of training on a task to 676 discriminate the direction of global motion. Fahle and Edelman (1993) [28] 677 predicted that internal reinforcement could act as the teaching signal when 678 performance feedback was absent and when the confidence is high [34,135]. 679 Therefore, perceptual learning should occur when training procedures include a 680 mixture of easy and difficult trials. This was the case for the study conducted 681 by Liu et al. (2012) [114] where learning occurred for easy and difficult trials 682 without feedback. For this specific task, consistent with Seitz et al. (2006) [136], 683 our results found that feedback was a requirement for learning. After ten days 684 of training the group who received feedback improved significantly, while the 685 performance for the no feedback group deteriorated.  Re-weighting Model (AHRM) [21]. When external feedback is provided, the 691 post-synaptic activation is shifted further in the correct direction, enforcing 692 appropriate weight changes in the decision unit. However, when external 693 feedback is absent the model uses the observer's internal response. In this 694 situation learning is dependent on the level of difficulty of the task, and uses the 695 observer's internal confidence to update the weights [114]. Where a task is easy, 696 the weights still move, on average, in the correct direction [137]. For a difficult 697 task, the neural signal is weak and a clear indication of the appropriate changes 698 that are required is absent. As a result the process of updating the decision 699 weights is ineffective, and learning does not occur.

700
The results of our study showed that learning did not occur without feedback, 701 even when easy trials were presented. While our task spanned the 65-85% and at no other levels, throughout the experiment. Ours, on the other hand, did 706 not include an adaptive staircase. Stimuli were randomly presented using 707 MOCS. This method involves random stimulus selection from a predefined range 708 of stimulus magnitudes [138]. In our experiment observers were randomly 709 presented 1 of 7 threshold levels. Interleaving these levels may influence how 710 decision weights are updated and reduce the observer's confidence in their 711 judgements, for example by disrupting observers' meta-cognitive judgements of 712 perceptual confidence [139] and thus their ability to selectively weight 713 high-confidence trials in perceptual learning [34].

714
Our findings bear some similarity to those found by Seitz et al. (2006) [136], 715 who also used MOCS. They found that learning did not occur without feedback, 716 even when easy trials were presented. Observers trained on one of two tasks, 717 either to discriminate the direction of low luminance motion stimuli, or to 718 discriminate the orientation of a bar that had been masked in spatial noise.

719
After training, both groups who had received feedback showed an improvement, 720 while the groups without feedback did not improve. Seitz et al. (2006) [136] 721 note that many of the experiments investigating perceptual learning use 722 adaptive staircase procedures, and even though easy trials were present, this 723 difference may contribute to the difference in findings. Seitz et al. (2006) [136] 724 propose that interleaving easy and difficult trials within the staircase may allow 725 for "better bootstrapping" from easy to hard, compared to trials that are 726 randomly presented.

727
A second difference between the study conducted by Liu et al. (2012) and 728 ours, was the type of task used. We used a global motion coherence task and 729 not an orientation discrimination task [114]. These tasks are known to be 730 processed at different levels of the visual hierarchy. Learning without feedback 731 has been obtained for local motion tasks. Ball and Sekuler (1982) [15] found 732 that observers did not need feedback to improve on their direction 733 discrimination task. In their study, observers were required to make a 734 same/different judgement for two rapidly presented trials. In the "same" trials, 735 November 28, 2018 31/51 motion took the same direction, and in the "different" trials the direction of 736 motion varied by 3 • . However, while the no feedback group did not receive 737 trial-by-trial response feedback, they were rewarded with two cents for a correct 738 response and had one cent deducted for each incorrect response, which may be 739 construed as end of block feedback [136], which has been found to be as effective 740 as trial-by-trial feedback [113].

741
Learning without feedback has also been found for global motion coherence 742 tasks [78,140]. However, none of these examples used the equivalent noise 743 coherence task, but rather used the ratio of signal-to-noise coherence task. Levi 744 et al. [75], using an equivalent noise method of global motion, also found 745 perceptual learning, however they used an adaptive staircase for training paired 746 with trial-by-trial feedback. When investigating task-irrelevant learning, 747 Watanabe et al. (2002) [141] found that task-irrelevant local motion improved 748 passively, but did not find the same for task-irrelevant global motion. They Having established the necessity of feedback for learning, the objective of the 762 main study was to investigate the specificity of spatial frequency tuning in 763 perceptual learning for global motion. We trained three groups of observers on a 764 global motion task with stimuli tuned to three different spatial frequency ranges 765 (Broad, Low and High) and performed pre-and post-training assessments for all 766 frequencies and high and low spatial frequency contrast detection. Based on the evidence for the frequency tuning of V5 [79,80,96,104], we 773 predicted that we might have obtained greater improvements (and transfer) 774 between broad and low spatial frequencies, but limited or no transfer for the 775 high spatial frequency trained group. Furthermore, we predicted that should 776 transfer occur as a result of the reweighting of the cortically localised global 777 motion mechanisms, then any transfer would reflect the broad, approximately 778 low-pass spatial frequency tuning of global motion detectors. There would thus 779 be most transfer when both training and test stimuli contained low spatial 780 frequency components, and a moderate improvement for those trained or tested 781 with broad frequency components. Since contrast discrimination is not 782 processed in the same cortical location as global motion we predicted no transfer 783 to contrast sensitivity would occur from any trained condition. Theory [33,72], we predicted transfer to show specificity to the frequency tuning 787 of the global motion processing areas. Given the broadband low-pass frequency 788 tuning of the motion processing areas, such as V5 and V3A, we predicted robust 789 to moderate improvement where stimuli contain low spatial frequency 790 components (learning to transfer from low to low; modest transfer of learning 791 from low to broad, and broad to low and broad). In addition, we predicted  Notably, the broad trained group performed worse than they did at 816 pre-assessment stage, and there was no change for the high trained group.

817
Furthermore the only evidence of transfer to untrained conditions, was from the 818 low frequency trained group on the broad test condition. Pre-and post-training 819 assessment for contrast sensitivity found that there was a significant 820 improvement exclusively for the low trained group on the low spatial frequency 821 contrast condition. No further improvement was found for any other trained or 822 tested frequency. Finally, the orientation discrimination control task showed no 823 improvement from any trained frequency.

824
Evaluating the transfer from global motion 825 This study explored the effects of training on global motion and its transfer to 826 other spatial frequencies and tasks. When assessing the post-training transfer to 827 trained and untrained global motion frequencies, the low spatial frequency 828 trained group was the only group to 'transfer to trained task' and transfer to 829 another condition. Although the 75% threshold was worse for the broad test 830 after training, the asymptotic performance was significantly better. Interestingly, 831 for the high spatial frequency test, there was a significant improvement to the 832 threshold although the slope was significantly shallower. This suggests that the 833 sensitivity for lower coherence levels increased. Performance at the highest levels 834 reached an asymptotic performance close to 1 at pre-assessment and remained 835 unchanged at the post-assessment stage. Finally the shallower slope suggests a 836 reduction in correct responses as a function of stimulus intensity.

837
The high spatial frequency trained group showed no improvement and no 838 transfer to any other spatial frequency.

839
The most surprising results were those obtained from the broad frequency 840 trained group. Asymptotic performance was significantly worse at the post test 841 stage for their own trained frequency and the high frequency test, with no 842 significant change in the low frequency test. There was no improvement in the 843 slopes for any condition, and a small but significant improvement was found in 844 75% threshold for the high frequency test. Viewing individual performance 845 revealed that the reduction in asymptotic performance may be accounted for by 846 an outlier, however the outlier was unlikely to account for the overall absence of 847 improvement in performance across all levels and measures, as there was no 848 clear improvement evident for the other observers.

849
Comparisons to Levi et al.

850
Since Levi et al. (2015) reported improved contrast detection for low frequency 851 drifting targets, we had questioned if the improvement was specific to the 852 temporal and spatial features of the training stimuli. Half the cells in layer 4Cα 853 of V1, where improvement in contrast has been argued to take effect [27], are 854 tuned for direction of motion [142]. It might be expected therefore that 855 improvement in contrast sensitivity would be limited to moving stimuli.   [75] reported the biggest improvement for low spatial 899 frequency stimuli exclusively in moving targets, this may suggest improvement 900 was specific to the temporal and spatial features of the training stimuli. In 901 contrast, our results suggest that when trained on low frequency global motion, 902 robust learning and transfer occurred for stimuli that contained low frequency 903 elements, including to a static contrast sensitivity task. Both of these findings 904 are consistent with the low-pass frequency tuning of V5. We predicted that should global motion training improve contrast detection, 909 given the frequency tuning of the early visual areas, we would expect specificity 910 to the spatial frequency of training. Improvement was restricted to the low 911 frequency contrast condition for the low trained group. The specificity of 912 transfer to low frequencies is consistent with the low-pass frequency tuning of 913 global motion areas such as V5 and V3/V3A influencing processing at V1 914 through feedback loops [109]. These results suggest however that there is some 915 specificity for low frequency elements. This remains compatible with the view 916 that global motion detectors pool via low frequency information. This may 917 suggest that the re-entrant connections to V1, after training on global motion, 918 only update low frequency channels.

919
Transfer 920 We predicted that if transfer occurred as a result of backward generalisation, 921 then improvement would reflect the frequency tuning of the global motion 922 detectors. The group trained on high spatial frequencies showed no improvement 923 and no transfer. Since the representation of high spatial frequency content is 924 attenuated in V5 [104], this may explain the lack of improvement in any 925 post-assessment from the high or broad frequency trained groups. This may 926 suggest any stimuli containing a high frequency element are attenuated for the 927 purpose of the backward projections to V1 from global motion processing areas. 928 The broad trained group performed worse overall for almost all 929 post-assessments. For the low spatial frequency contrast detection, there was 930 evidence of a significantly steeper slope, however (a change in the rate of 931 increase in proportion of correct responses with contrast), however this did not 932 lead to a robust improvement, and the average detection threshold was worse. 933 Transfer was only found for the low frequency trained group. This group 934 showed improvement in low spatial frequency contrast detection, a significant 935 increase in asymptotic performance for the broad frequency global motion test, 936 and also improved on their own task. This may suggest that after training on 937 global motion only the low frequency channels are updated. The low frequency 938 trained group will have experienced a high level of correlation in the activity in 939 frequency channels between the feedforward and feedback connections. This 940 joint activity may be important for perceptual learning, and is consistent with 941 the proposal that there is an iterative interaction between the global motion 942 processing areas and V1, since feedback from higher to lower areas engage the 943 connections that mirror the features of the stimulus [65].
What our findings suggest for the models of perceptual learning 945 The two main theories of visual perceptual learning suggest that learning either 946 occurs through physiological changes to sensory neurons [18,19,31], or as a 947 reweighting [22] of the decision units. Our findings are not able to distinguish 948 between learning as a result of a physiological change or a higher-level 949 computational reweighting. However neither model is able to account for the 950 breadth of learning (and non-learning) and transfer within our results including 951 the general absence of improvement and the occasional decline in performance at 952 the post-assessment stage. The only difference between these components of the 953 experiment was that trial-by-trial feedback was present for the training phase 954 and absent for the testing phase. Our pilot study found that, for our methods, 955 learning occurs with feedback, but not without. Additionally, Herzog and Fahle 956 (1997) [113] suggested that once feedback is removed, performance plateaus 957 around the level last obtained with feedback, so our expectation was that 958 observers would maintain the improvement achieved for their trained frequency. 959 Neither model alone accounts for the ambiguous performance at the 960 post-assessment stage or the decline in performance by the broad trained group. 961 On the other hand, the specificity to low spatial frequencies is consistent with 962 both models, given the frequency tuning of the global motion detectors.

963
While the two positions are pitched as competing theories, they may not be 964 mutually exclusive. In an attempt to combine the two models, Solgi and Weng 965 (2013) [143] model learning as a two-way process (descending and ascending).

966
The model attributes training effects to the reweighting of connections between 967 early and higher sensory areas, but crucially also as a result of increased 968 neuronal activity representing the trained feature [143]. For transfer to occur, 969 Solgi and Weng (2013) [143] argue there is an important role for the re-entrant 970 connections from higher levels. Similarly, Bejjanki et al. (2011) [144] propose a 971 model that is computationally similar to the reweighting models, however it 972 illustrates how changing the population codes in the early sensory areas can 973 create similar changes in response to those made by high-level re-weighting 974 models. Bejjanki et al. (2011) [144] argue that their model captures the 975 characteristics for perceptual learning for both behavioural and physiological 976 changes. Sensory inputs improve the decision weights in the feed-forward 977 connections, and improved probabilistic inference in early visual areas as result 978 of the increased neural activity in the feedback network. This is compatible with 979 our findings that suggest there the high level of consistency experienced for the 980 group trained and tested on the low frequency conditions. Ultimately, a model 981 of perceptual learning needs to account for multiple cognitive influences. 982 Maniglia and Seitz (2018) [145] argue that perceptual learning is not a detached 983 or isolated process, instead multiple mechanisms react and interact to produce 984 learning.

985
Afterthoughts, Speculation and Unresolved Points 986 We offer some speculative ideas that may help account for the general absence of 987 improvement at the post-assessment stage. There is emerging evidence that 988 interleaving random stimuli may disrupt perceptual learning [146]. Yu et al.

989
(2004) [56] found that when interleaving random contrasted Gabor stimuli, such 990 as when using MOCS, learning did not occur, but did when only one contrast 991 was presented at a time. Yu et al. (2004) [56] termed the interleaving of contrast 992 stimuli as "contrast roving", and suggest that the effect of non-learning as a 993 result of contrast roving implies that a temporally organised pattern of stimulus 994 presentation may be required for perceptual learning to occur. Additionally, 995