The goal of the present study was to investigate whether it is possible to reduce the effects of visual crowding through training and how such a reduction is mirrored at the level of neural responses, particularly in the BOLD signal in early visual cortex. Determining whether or not there is a substantial reduction in crowding after training has important implications for clinical practice, as well as for our understanding of peripheral vision and object recognition. Here we argue that training can modify the anisotropic feature of crowding manifested not only in behavioral responses but also at the level of neural signal changes.
Performance on different tasks is known to improve via practice (i.e., perceptual learning). This improvement usually occurs fairly quickly and enhances performance on that task even for day or several weeks or months after cessation of training.
43,44 It is suggested that the improvement in visual perceptual functions is achieved by promoting neural plasticity as early as in the striate
45–48 and/or in extrastriate
49 visual cortex.
Training-induced reductions in the strength and extent of crowding have been demonstrated previously in the periphery of normal-sighted subjects,
27,29,50,51 as well as in amblyopic individuals.
28 Chung
24 found a reduction in the radial–tangential anisotropy in the zone of the preferred retinal locus (PRL) in patients with age-related macular degeneration. The change in the shape of the crowding zone at the PRL was associated with more intense usage of the PRL. Similarly, after training, we observed shrinking of the crowding zone along the radial axis in our behavioral experiments, as the shape of the crowding zone became less elliptical (
Fig. 3B). This indicates that the radial–tangential anisotropy of the crowding zone can indeed be altered by training.
Several studies have shown that the anisotropic feature of visual crowding is replicated in BOLD responses.
25,31 Likewise, in our pre-training study, we observed anisotropy in BOLD responses, as adding a target to equally spaced radial flankers induced a larger reduction in the BOLD signal compared to tangential flankers. After behavioral training, the BOLD crowding index (C
BOLD) was reduced significantly for the radial flanker condition but did not change for the tangential flanker configuration. The results, presented in
Figure 5C, demonstrate that training led to a radial-specific reduction of the crowding effect (positive values of C
BOLD indicate a weaker crowding effect, whereas negative values indicate a stronger crowding effect). The results of ANOVA analysis for C
BOLD based on the fMRI BOLD response point to a significant interaction between flanker configuration (radial vs. tangential) and training effect. Also, paired-sample
t-tests revealed significant differences in the pre- and post-training anisotropy index (A
BOLD). Spearman's rank correlation coefficient test revealed a significant correlation between the standardized training-induced changes in A
BOLD and A
psy, indicating that the greater the training-induced change in A
psy the larger the changes in A
BOLD across participants
. To the best of our knowledge, our study demonstrates for the first time that training-induced changes in the anisotropic shape of the crowding zone are also reflected at the level of the BOLD signal in early visual cortex.
An interesting tendency in the BOLD signal changes that we observed was an increase of BOLD signal in the post-training radial–target-present condition compared to pre-training measurements, whereas the radial–target-absent condition showed the opposite trend. Three effects could contribute to the results of
Figures 5A and
5B: (1) radial versus tangential bias in the fMRI signal strength, which was also reported in previous studies by Sasaki et al.
52 and Kwon et al.
25; (2) the repetition suppression phenomena,
53,54 where the neural responses to repeated stimuli are reduced relative to novel stimuli when stimuli are presented multiple times; and (3) BOLD response suppression due to crowding in the pre-training radial–target-present condition and an increase in BOLD signal intensity in the post-training radial–target-present condition. This last effect presumably is induced by the training effect that leads to weakening crowding-related BOLD signal suppression. We also noticed an unexpectedly strong reduction of BOLD signal in the post-training radial–target-absent condition compared to the pre-training radial–target-absent condition, for which we currently do not have an explanation.
Multiple models attempted to explain the asymmetry of the anisotropy features of crowding. Dayan and Solomon
55 proposed a quantitative model to interpret several paradoxical properties of crowding. In particular, their model tries to describe the inward–outward asymmetry of crowding by optimal (Bayesian) inference operating over spatially extended receptive fields. Another approach, developed by van den Berg and colleagues,
26 is based on the principles of population coding. Those models predict several properties of crowding but still cannot explain the radial–tangential anisotropy.
An alternative explanation for the elliptical shape of the crowding zone was proposed by Nandy and Tjan,
56 who implicated physiological and anatomical properties specific to V1 and saccade-confounded image statistics. They have suggested that normal saccadic eye movements that are radial with respect to the fovea will lead to the acquisition of inaccurate image statistics in peripheral vision. However, this theory does not predict whether crowding will lead to a decrease or an increase in neural activity.
What might be the possible neural mechanisms of training-induced changes in the anisotropic feature of crowding that we observed in our experiments? Allegedly, perceptual learning affects almost every level of processing in the brain, from synaptic connections to global patterns of blood flow.
43 However, the neural mechanisms underlying the training effect remain highly controversial. Based on the suggestion that there is a link between BOLD signal and synaptically generated local field potentials,
57–59 one can hypothesize that an increase of BOLD signal after training might be induced by an increase in the number or strength of synaptic connections. A new study by Contemori et al.
60 investigating the effect of transcranial random noise stimulation (tRNS) on perceptual learning demonstrated a stronger reduction of crowding in the tRNS group compared to the sham stimulation group. They proposed that the boosting effect of tRNS could be evoked from a general increase in cortical excitability, as well as by an improved signal-to-noise ratio. Earlier studies on monkeys have shown a so-called push–pull response pattern in V1 neurons over the course of training on a contour detection task, such that responses of neurons with receptive fields lying on the contour are progressively enhanced, and those on the noisy background are suppressed.
61,62
More recently, based on fMRI data, it has been proposed that visual perceptual learning could be associated with the two types of plasticity processes, feature based or task based.
63 Feature-based plasticity implies that there are changes in tuning properties of the neuronal representations of a trained visual feature.
45,64–66 Task-based plasticity refers to improvement in task-related processing due to training on a task and is characterized by connectivity changes between feature representations and decision units.
67,68 Neural correlates of feature-based plasticity can be observed during passive viewing condition, when subjects are passively exposed to the trained stimuli. Task-based plasticity, on the other hand, is assessed during active-task conditions, when subjects are actively performing the task.
Given all of these considerations, we suggest that reduced crowding after training can be associated with an enhanced signal-to-noise ratio of neural responses to the stimuli by refinement of neural population codes in early visual cortex that represent the trained stimulus features. The reduction of the anisotropy effect in the radial direction could be explained by the presence of different lateral interactions in the periphery that exhibit training-dependent modulation.
69 This suggestion is supported by an earlier fMRI study on orientation discrimination, the results of which indicate that training enhances the discriminability of individual neural populations (represented by individual voxels) without altering the overall activation strength, indicating that the learning-induced modulatory effects could be facilitatory for some neural populations but inhibitory for the others in response to the trained stimulus.
70
Another interesting observation we made in our study is related to the loci of the BOLD activation during crowding. We found that most ROIs fall between visual areas V1 and V2 (
Fig. 2). There are debates regarding the loci where crowding first emerges. Whitney and Levi
71 promoted a multiple-level hypothesis of visual processing occurring as early as V1. Kwon et al.
25 showed that anisotropy in BOLD response occurs as early as V1. On the other hand, there are studies showing evidence for learning in visual areas V2
9 and V4,
72 as well as high-level visual areas such as the lateral occipital cortex,
33 and these could be the candidate loci for visual crowding. In our study, we did not perform individual retinotopic mapping; therefore, it is not possible to precisely define the exact borders between early visual areas involved in crowding.
In summary, the results of our study indicate that training reduces visual crowding and its anisotropy feature that is reflected in BOLD responses in early visual cortex. Our results support the effectiveness of plasticity-based approaches to improve visual function. We suggest that further studies, similar to those that have investigated the size and shape of population receptive fields, are required to reveal the processes that underlie training effects on crowding.