Mitigating age-related stereo decline with perceptual training

Summary

Aging is accompanied by visual decline, largely driven by central rather than peripheral changes. Encouragingly, however, the aging brain is neuroplastic. Here, we used perceptual learning to improve stereoscopic depth perception in older and younger observers, comparing their learning capacity and transferability across two depth tasks: signal-in-noise and fine discrimination. Ninety participants (45 younger and 45 older) were randomly assigned to train on one of the tasks or to a no-training control. All participants completed pre- and post-tests on both tasks, with training groups receiving training over three consecutive days. Despite lower baseline performance, older adults exhibited learning rates and magnitudes comparable to those of younger adults. In both age groups, fine discrimination training improved performance on both tasks, whereas signal-in-noise training yielded task-specific gains. Our findings demonstrate that the human binocular visual system retains substantial plasticity that is governed by mechanisms that remain effective with age.

Graphical abstract

Highlights

• •Perceptual learning enhances stereo depth responses in the aging visual system
• •Older adults showed comparable learning gains, rate, and transfer as younger adults
• •Aging brain retains binocular plasticity, potentially via age-preserved mechanisms

Health sciences; Medicine; Medical specialty; Neurology; Neuroscience; Clinical neuroscience

Introduction

The global population is undergoing rapid aging, with the percentage of adults aged 65 years and older predicted to rise from 8.0% in 1950 to 16.0% by 2050.¹ While this increased longevity reflects advances in medicine and public health, it is accompanied by a growing burden of age-related conditions that threaten global well-being. Visual decline is among the most consequential of these conditions, as it directly undermines autonomy and quality of life. Healthy aging is associated with declines across multiple domains of visual processing, including contrast sensitivity,² orientation discrimination,³ visual acuity,⁴ motion and form perception,^5,6 optic flow,⁷ color sensitivity,⁸ object recognition,⁹ and stereopsis.¹⁰ Critically, these declines are not restricted to peripheral mechanisms but also implicate central visual processing, where age-related changes are thought to relate in part from reductions in gray matter volume¹¹ and diminished GABAergic inhibitory signaling, resulting in a disruption of the excitation-inhibition balance within the visual cortex.^12,13

Within the context of age-related decline in vision, the deterioration of stereopsis is particularly impactful. Stereopsis refers to the perception of depth that arises from the brain’s integration of slightly different images projected onto each retina—known as binocular disparity. Intact binocular depth perception enables accurate judgments of distances and spatial relationships between objects, which are critical for everyday activities such as navigating complex environments, avoiding obstacles, and executing precise hand-eye coordination. Stereoscopic cues become especially critical in visually demanding conditions where monocular information is limited or unreliable, such as in low contrast, visual clutter, or reduced visibility. Impaired stereopsis, thus, significantly compromises functional independence and has been linked to an elevated risk of falls and motor vehicle collisions among older adults.^14,15,16,17 Notably, stereoacuity begins to decline after the age of 60, with more pronounced deterioration typically emerging beyond the age of 70—even in the absence of ocular pathology^18,19,20,21—likely reflecting age-related changes in cortical processing and the integrity of the binocular visual pathways.

At the neural level, stereoscopic depth perception is supported by disparity-selective mechanisms that emerge early in the visual hierarchy, including primary visual cortex (V1),^22,23 V2,²³ and V3A,^24,25 and extend across multiple extrastriate regions such as the middle temporal area,²⁶ inferotemporal cortex,²⁷ lateral occipital complex (LOC),²⁸ and regions within the posterior parietal cortex (PPC).²⁹ Functionally, this network is often described as supporting complementary roles, with dorsal-stream regions partaking in the processing of depth for spatial/motion determination, including visually guided action, and ventral stream regions establishing disparity-defined object structure and three-dimensional shape.³⁰ Despite the well-characterized loci and functional connections among these nodes for binocular disparity processing, relatively little is known about how healthy aging alters these neural circuits. Direct neurophysiological evidence as to the influence of aging on sensory systems remains limited. One relevant study used visually evoked potentials (VEPs) to compare cortical responses to both disparity-defined and motion-defined stimuli in younger and older adults.³¹ While early VEP components—typically linked to initial sensory encoding—were largely preserved with age, later components exhibited reduced amplitudes across stimulus types.³¹ This temporal pattern suggests that aging may spare early-stage visual processing but compromise later integrative or feedback-related mechanisms, possibly reflecting diminished neural synchrony³² or altered inhibitory dynamics.³³ These findings, though not stereopsis specific, offer preliminary insight into how aging may alter cortical processing of binocular disparity.

Encouragingly, a growing body of evidence suggests that the aging brain retains a greater degree of neuroplasticity than previously assumed,³⁴ offering a promising opportunity for mitigating age-related visual deterioration. Perceptual learning in older adults has been proven to be effective in improving contrast sensitivity,³⁵ orientation discrimination,^36,37 biological motion perception,³⁸ and motion direction discrimination.³⁹ Of immediate relevance to the present work, recent findings reveal that stereoacuity can be enhanced in individuals aged 79–96 years, with performance in a dynamic stereoscopic task improving after approximately 3,000 training trials.⁴⁰ While these results provide a compelling demonstration of preserved binocular visual plasticity in late life, the underlying mechanistic changes remain largely unknown. Do older adults retain the same degree of learning capacity and recruit the same learning mechanisms as younger individuals?

Previous perceptual learning work on depth discrimination in young human adults has indicated that task demands—specifically, the presence or absence of external noise—play a critical role in determining the extent of learning and transfer. Specifically, training under noise-free conditions (i.e., discriminating the relative depth of two adjacent objects) not only enhances performance on the trained task but also generalizes effectively to noisy conditions. By contrast, training in the presence of external noise (i.e., judging the depth of a target object embedded in noise) yields limited transfer to noise-free contexts.^41,42 Importantly, these behavioral asymmetries are accompanied by distinct patterns of functional reweighting along dorsal and ventral visual pathways,⁴² highlighting task-dependent neural plasticity and functional reorganization within the visual system. This asymmetry in learning transfer aligns with established models of perceptual learning, most notably the perceptual template model⁴³ and related reweighting frameworks,⁴⁴ which propose that low-noise training enhances the stimulus gain and promotes broad generalization, whereas high-noise training selectively improves external noise exclusion, with benefits that are more task and context specific. Whether the aging visual system exhibits a similar profile of learning generality remains unclear. Notably, the aging brain appears to exhibit both reduced tolerance to external noise³⁹ and greater variability in neural responses,^39,45 both of which could plausibly limit the efficiency and generalizability of perceptual learning in late adulthood.³⁶

Here, we aimed to characterize the nature and extent of improvements in stereoscopic depth perception that could be attained through perceptual learning in the aging visual system. Specifically, we compared the capacity for depth discrimination learning in younger and older adults, and the generality of learning effects across two task types: (1) a signal-in-noise (SNR) depth task, which required judging the depth position of a target object embedded in visual noise, and (2) a fine depth (FD) discrimination task, which involved distinguishing between the relative depth of two closely positioned objects (Figure 1). The tasks were designed to differ not only in the presence or absence of external noise but also in the stereoscopic computations they engage, which have been shown to rely on partially dissociable neural pathways within the human visual system.⁴²

Figure 1.

Signal-in-noise depth and fine depth discrimination tasks

For the signal-in-noise depth task, signal dots in the central plane appeared either nearer or farther than the surrounding plane, which was fixed at zero disparity. For the fine depth discrimination task, the surrounding plane was held at a fixed disparity of +12 arcmin, while the disparity of the central plane varied finely relative to it. For both tasks, the participants judged whether the central target appeared in front of (“near”) or behind (“far”) relative to the surrounding. Example random-dot stereograms, rendered as red-cyan anaglyphs, are shown on the right for each task. The participants maintained fixation on the small white square at the center of the display.

To our knowledge, no prior study has directly compared stereoscopic perceptual learning between younger and older adults. However, research examining other visual features has shown that older adults are capable of substantial perceptual learning, with learning rates and magnitudes that sometimes approach those of younger adults.^{37,39,46,47,48} Importantly, some of these findings often emerge under conditions designed to accommodate age-related limitations, such as individually tailored task difficulty or extended training durations for older adults, and may not always adequately address baseline performance differences between age groups.^37,46,47 In the present study, by contrast, we employed identical stimuli (i.e., matched task difficulty) and equal training durations across age groups, enabling a more direct and controlled comparison of learning outcomes. Given the documented age-related changes in visual cortical processing, including reduced signal-to-noise ratios,³⁹ increased neuronal noise,^39,45 and changes in inhibitory function,⁴⁹ which may collectively constrain the behavioral expression of plasticity,⁵⁰ we predicted that older adults would retain the capacity for reliable perceptual learning (i.e., demonstrating significant performance gains on their respective training task), but the magnitude of these improvements might be lower than that observed in younger adults. Further, if the aging visual system engages learning mechanisms analogous to those of younger adults, we expected to observe an asymmetric pattern of transfer in both age groups: training on the FD discrimination task would generalize to the SNR depth task, but the reverse transfer would be limited.

Results

Learning-related changes and transfer

We first examined learning improvements for the SNR depth task. Signal-noise thresholds were analyzed with a 2 (age group – younger/older) × 3 (training condition – SNR training/FD training/no training) × 2 (session – pre-training/post-training) mixed ANOVA (Figure 2A). Prior to analysis, model assumptions were assessed using standardized residual diagnostics. Normality of residuals was evaluated via quantile-quantile plots (Q-Q plots) and Shapiro-Wilk tests, which indicated no significant deviations from normality in either age group. Levene’s tests confirmed the homogeneity of variance across between-subject factors. This analysis revealed a significant two-way training condition × session interaction (F_{(2, 84)} = 13.436, p < 0.001, and η²_p = 0.242) and a significant main effect of age group (F_{(1, 84)} = 5.435, p = 0.022, and η²_p = 0.061). The three-way interaction between age group, training condition, and session, however, was not significant (F_{(2, 84)} = 0.542, p = 0.583, and η²_p = 0.013), and there were no other interactions involving age group. Post-hoc comparisons on the main effect of age group revealed that the younger group exhibited significantly lower signal-noise thresholds (mean = 29.0, SD = 10.83) than the older group (mean = 34.40, SD = 11.53). To further investigate the two-way interaction, we collapsed the data across the two age groups and performed Bonferroni-corrected t tests, comparing data between sessions for each training group. In particular, we observed that signal-noise thresholds improved not only for the group who received dedicated SNR training (t₍₂₉₎ = 7.875, p < 0.001) but also for those who received FD training (t₍₂₉₎ = 7.622, p < 0.001). Thresholds for this task, however, did not improve for the no-training control group (t₍₂₉₎ = 1.522, p = 0.417). Individual threshold changes from pre-test to post-test are presented in Figure S1. It is noteworthy that the pre-test thresholds did not significantly differ across the three training conditions (F_{(2, 87)} = 0.176, p = 0.839), indicating that the observed learning effects are unlikely to be driven by baseline differences in performance. In an exploratory extension of this model, we added gender as a between-subject factor; however, no significant gender main effect or gender-related interactions were observed.

Figure 2.

Pre- and post-test depth discrimination thresholds for younger and older adults

(A) Signal-in-noise depth discrimination thresholds shown separately for younger (orange) and older (navy) groups, with pre-training thresholds in darker shades and post-training thresholds in lighter shades.

(B) Fine depth discrimination thresholds shown with the same color/shade conventions. Error bars represent ±1 SEM. ∗p < 0.001 by Bonferroni-adjusted paired-samples t test.

We next analyzed data for the FD discrimination task with a similar 2 (age group) × 3 (training condition) × 2 (session) mixed ANOVA to examine learning gains (Figure 2B). As with the SNR depth task analysis, model assumptions were assessed using standardized residual diagnostics. Shapiro-Wilk tests revealed some deviations from normality in certain conditions, particularly in post-training scores, but Q-Q plots suggested these departures were mild and largely restricted to the distribution tails. Given the robustness of ANOVA to such minor violations, all data were retained for analysis. Levene’s tests confirmed the homogeneity of variance across between-subject factors. Consistent with the pattern observed in the SNR depth task, the analysis revealed a significant two-way interaction between training condition and session (F_{(2, 84)} = 8.991, p < 0.001, and η²_p = 0.176). There was also a significant main effect of age group (F_{(1, 84)} = 19.843, p < 0.001, and η²_p = 0.191). The three-way interaction between age group, training condition, and session was not significant (F_{(2, 84)} = 0.06, p = 0.945, and η²_p = 0.001), and no other interactions involving age group reached significance. Follow-up analyses on the main effect of age group confirmed that, overall, younger participants exhibited significantly lower (better) fine discrimination thresholds (mean = 27.95, SD = 14.84) than older adults (mean = 43.36, SD = 18.17). To further examine the significant training condition × session interaction, we concatenated the data from the two age groups and observed substantial improvements in discrimination thresholds exclusively within the group that received dedicated FD training (t₍₂₉₎ = 6.415, p < 0.001). No significant changes in thresholds were observed for the SNR training group (t₍₂₉₎ = 2.087, p = 0.138) or the no-training control group (t₍₂₉₎ = −0.074, p = 1). Consistent with the findings for the SNR depth task, a one-way ANOVA revealed no significant differences in pre-test thresholds across the training groups (F_{(2, 87)} = 0.305, p = 0.738). The learning effects observed here, therefore, cannot be due to pre-existing differences in baseline performances. Detailed individual threshold changes between pre-test and post-test are shown in Figure S1. Adding gender to the mixed ANOVA did not yield any significant main or interaction effects.

To further evaluate whether aging disproportionately affects performance on the two depth tasks, we quantified the baseline performance decrement for each older participant as a normalized difference from the younger group mean for each task, calculated as follows: ([performance − mean_young]/mean_young). Under this metric, positive values would indicate worse performance in older adults relative to younger adults, while negative values would reflect better performance in older adults. These normalized differences were then compared across the two tasks, using a paired-samples t test, which revealed a significant difference between the tasks (t₍₄₄₎ = −2.938, p = 0.005). Specifically, older adults exhibited a greater relative performance decrement on the FD discrimination task (mean = 0.417, SD = 0.604) than on the SNR depth task (mean = 0.146, SD = 0.387), indicating a larger age-related difference in pre-training performance on the FD discrimination task.

As the baseline performance on both test tasks differed between the two age groups, we computed a percentage change score to normalize individual improvement on the tasks relative to their respective baseline performance. The percentage change was calculated using the following formula: ([thresh_pre − thresh_post]/thresh_pre). This approach allowed us to quantify the magnitude of performance change from pre-training to post-training, while accounting for any inherent differences in baseline performance between the young and aging cohorts.

Subsequently, we conducted two complementary sets of analyses to examine learning-related improvement. First, for each age group and training condition, Bonferroni-corrected one-sample t tests were performed to determine whether percentage change scores on the SNR depth and FD discrimination tasks differed significantly from zero (i.e., no change). The results aligned with the mixed ANOVA findings reported earlier (Figure 3). Specifically, training on the FD depth task led to significant improvements in performance on both the SNR depth (younger: t₍₁₄₎ = 5.617, p < 0.001; older: t₍₁₄₎ = 6.311, p < 0.001) and FD discrimination tasks (younger: t₍₁₄₎ = 11.104, p < 0.001; older: t₍₁₄₎ = 5.720, p < 0.001) for both the younger and older adults. By contrast, training on the SNR depth task resulted in performance enhancements solely on the SNR depth task (younger: t₍₁₄₎ = 9.826, p < 0.001; older: t₍₁₄₎ = 6.698, p < 0.001), with no statistically reliable changes observed on the FD task (younger: t₍₁₄₎ = 2.363, p = 1; older: t₍₁₄₎ = 0.18, p = 1). The no-training control groups for both age groups showed no significant changes in thresholds for either the SNR depth task (younger: t₍₁₄₎ = 0.948, p = 1; older: t₍₁₄₎ = 0.083, p = 1) or the FD discrimination task (younger: t₍₁₄₎ = −0.198, p = 1; older: t₍₁₄₎ = −1.001, p = 1).

Figure 3.

Normalized differences in pre-test versus post-test depth discrimination thresholds in younger and older adults

(A) Normalized change scores for the signal-in-noise depth task.

(B) Normalized change scores for the fine depth discrimination task. For both panels, positive values indicate improvement (i.e., lower thresholds at post-test relative to pre-test). For both younger and older participants, training on the FD task led to significant improvements on both the FD and SNR tasks, whereas SNR training yielded improvements only on the SNR task, with no statistically reliable changes on the FD task (see main text for statistical details). Error bars represent ±1 SEM. ∗p < 0.001 by Bonferroni-corrected one-sample t test.

Second, to compare the magnitude of learning-related improvement across age groups and training conditions, we analyzed the percentage change scores, using separate 2 (age group) × 3 (training condition) ANOVA for each depth task. Residual diagnostics indicated that the assumption of normality was reasonably satisfied, with only minor deviations observed in the tails of the older adult distribution for the SNR depth task. Levene’s tests confirmed that the assumption of homogeneity of variances was met for both tasks. The analyses indicated only a significant main effect of training condition for both test tasks (SNR: F_{(2, 84)} = 17.566, p < 0 .001, η²_p = 0.295; FD: F_{(2, 84)} = 13.826, p < 0 .001, η²_p = 0.248). Follow-up Bonferroni-corrected pairwise comparisons on the SNR depth task revealed significant differences in percentage change between the SNR training group and the no-training group (t₍₅₈₎ = 5.331, p < 0.001), as well as between the FD training group and the no-training group (t₍₅₈₎ = 3.656, p < 0.001). While the magnitude of learning on the SNR depth task appeared to be smaller for individuals who received FD training than for those who received SNR training, there was no statistically reliable difference between the two training groups (t₍₅₈₎ = 2.394, p = 0.06). For the FD discrimination task, follow-up analyses indicated that the FD training group showed a significantly greater percentage improvement than both the SNR training group (t₍₅₈₎ = −5.082, p < 0.001) and the no-training control group (t₍₅₈₎ = 4.697, p < 0.001). Conversely, the SNR training group did not show a significantly different percentage change on the FD discrimination task compared with the no-training control group (t₍₅₈₎ = 1.426, p = 0.477).

To further quantify evidence for the absence of age-related differences in the magnitude of learning-related improvement, we conducted Bayesian ANOVAs on the percentage change scores for both depth tasks, with age group and training condition as factors. Analyses were performed using JASP (version 0.95.4) with default JZS priors (r scale-fixed effects = 0.707). Bayes factors for inclusion (BF_incl) provided strong evidence against an age group × training condition interaction for both the SNR depth task (BF_incl = 0.11) and the FD discrimination task (BF_incl = 0.16), indicating that the magnitude of perceptual learning was comparable across younger and older adults for both depth tasks.

Taken together, these results indicated that while the younger adults showed an overall better performance on both SNR depth and FD discrimination tasks than their older counterparts, the magnitude of learning improvement and the pattern of learning transfer were comparable between the two age groups. Specifically, training to discriminate between FD differences enhanced subsequent performance on both the FD and SNR depth tasks. However, learning under signal-noise conditions led to significant improvements strictly within the SNR depth task, with limited transfer to the FD task.

Learning rate and progression

In addition to evaluating the overall magnitude of learning, we also compared the rate of learning over time for the two age groups. Figures 4A and 4B (left) present the group-averaged learning curves for the younger and older participants, plotted separately for each training condition. Visual inspection of these learning trajectories revealed that, although younger adults consistently demonstrated superior performance (i.e., lower thresholds) throughout training, the overall progression of learning was strikingly similar across age groups. Specifically, both the younger and older participants attained asymptotic performance around the 20th training block for both the SNR and FD training conditions. Individual learning curves are presented in Figure S2 (SNR training) and Figure S3 (FD training).

Figure 4.

Learning curves and model-derived learning rates for the training groups

(A) Signal-in-noise depth training group.

(B) Fine depth discrimination training group.

Left: Group-averaged learning curves for younger and older adults in each training condition. Each point represents a three-block moving average. Shaded regions denote ±1 SEM. Right: Individual learning rates (k) estimated from two-parameter logarithmic model fits. Negative k values indicate improved performance over time, with more negative values reflecting steeper learning rates.

To formally compare the learning rate, individual training data were fitted with a two-parameter logarithmic function: b(a) = kln(a) +a₀, where b(a) denotes performance thresholds (signal-noise/discrimination threshold) at training block a, and a₀ captures the subject-specific baseline performance at the onset of training to determine estimates of the learning rate (k). This model was elected to account for inter-individual variability in initial performance by decoupling the starting performance from the rate of improvement, thereby providing a more interpretable and unbiased estimate of learning dynamics. To minimize the influence of trial-level noise and enhance the robustness of parameter estimation, model fitting was performed on smoothed training data, using a 3-point moving average. Within this model framework, negative values of k indicate performance improvements over time, with larger negative values reflecting steeper learning rates. We then compared the obtained learning rate (k) between age groups and training conditions by running a 2 × 2 ANOVA (Figures 4A and 4B, right). The results indicated no significant main effect (age group: F_{(1, 56)} = 2.459, p = 0.122; training condition: F_{(1, 56)} = 2.222, p = 0.142) or interaction (F_{(1, 56)} = 0.353, p = 0.555). To further quantify the strength of evidence for the absence of age- or training-related effects on the learning rate, Bayesian ANOVAs were conducted on the fitted k parameter, using JASP (version 0.95.4) with default JZS priors (r scale-fixed effects = 0.707). The Bayes factors for inclusion provided anecdotal to moderate evidence supporting the null model, with age group (BF_incl = 0.60), training condition (BF_incl = 0.35), and their interaction (BF_incl = 0.19) all showing limited evidence for inclusion. That is, the rate of improvement over the course of training was broadly comparable between the young and aging participants, as well as between the two training conditions.

Correlational analyses of task performance and learning

Finally, to further characterize the relationship between performance across the two depth tasks and individual differences in learning outcomes, we conducted a series of exploratory correlational analyses. To control for inflated type 1 error across multiple comparisons, the Holm-Bonferroni step-down correction was applied (two-tailed; α = 0.05).

We first examined the association between baseline performance on the two depth tasks and observed a moderate positive correlation (r₍₈₈₎ = 0.352, p < 0.001; Figure 5A), indicating that individuals who were more accurate in judging depth position in noise also tended to perform better on fine depth discrimination.

Figure 5.

Relationships between task performance and transfer across depth tasks

(A) Relationship between baseline signal-in-noise (SNR) depth thresholds and fine depth (FD) discrimination thresholds across all participants. Individuals who were more accurate in judging depth position in noise also tended to perform better on FD discrimination.

(B and C) Normalized change scores [(pre-post)/pre] on the trained FD task plotted against normalized change scores on the untrained SNR task within the FD training group, shown separately for (B) younger adults and (C) older adults. In both age groups, the degree of improvement on the trained FD task did not predict the extent of transfer to the SNR task. Shaded regions indicate 95% confidence intervals; solid lines represent least-squares regression fits.

Next, we investigated whether individual differences in learning-related outcomes were associated with participant characteristics. Performance change was quantified as the change in threshold from pre- to post-training (Δ = pre − post), with positive values indicating improvement. We first tested for associations between the age and performance change on the trained task, separately within each training condition and age group (Figure 6A). For the SNR depth training, no significant correlation was observed between learning-related changes and age in either the younger (r₍₁₃₎ = −0.100, p = 0.722) or older participants (r₍₁₃₎ = −0.285, p = 0.304). A similar pattern was observed in the FD discrimination training condition, where learning-related changes were also not significantly associated with age in either age group (younger: r₍₁₃₎ = −0.415, p = 0.124; older: r₍₁₃₎ = 0.230, p = 0.409). Second, we examined whether the baseline task performance predicted the extent of learning by correlating initial (pre-training) thresholds with performance change across training groups (Figure 6B). Among younger adults, baseline thresholds were positively correlated with training-related improvements for both tasks (SNR: r₍₁₃₎ = 0.976, p < 0.001; FD: r₍₁₃₎ = 0.892, p < 0.001). A similar trend was observed among older adults, with positive associations between baseline thresholds and learning-related changes in both training conditions. This relationship, however, reached statistical significance only in the FD training group (r₍₁₃₎ = 0.763, p = 0.001) but not in the SNR training group after correction for multiple comparisons (r₍₁₃₎ = 0.653, p = 0.008). It is important to note that correlations between baseline performance and change scores should be interpreted with caution, as such analyses are susceptible to mathematical coupling and regression-to-the-mean artifacts.

Figure 6.

Individual differences in training-related performance changes

(A) Relationship between age (years) and performance change (Δ = pre − post) on the trained task, shown separately for each age group and training condition. No significant associations were observed between age and learning magnitude in any group.

(B) Relationship between baseline threshold and performance change (Δ = pre − post) on the trained task, shown separately for each age group and training condition. In all groups, individuals with higher baseline thresholds (poorer initial performance) tended to show larger absolute threshold reductions following training. However, these baseline-change associations should be interpreted cautiously, as correlating baseline with change scores is susceptible to mathematical coupling and regression-to-the-mean artifacts. Shaded regions indicate 95% confidence intervals; solid lines represent least-squares regression fits.

Finally, within the FD training group, we explored whether changes in the trained task were related to changes in the untrained SNR depth task. To allow for meaningful comparison between tasks that use different measurement metrics (present signal for the SNR task and arcseconds for the FD task), normalized change scores ([thresh_pre - thresh_post]/thresh_pre) were used in the correlation analysis. The analyses revealed no significant correlations in either the younger participants (r₍₁₃₎ = 0.282, p = 0.308; Figure 5B) or the older participants (r₍₁₃₎ = −0.099, p = 0.725; Figure 5C).

Discussion

We examined the capacity of the aging binocular visual system for learning-induced stereoscopic depth improvement by comparing learning outcomes between younger and older adults across two depth discrimination tasks differing in perceptual demands, one requiring the extraction of depth signals embedded in noise, and the other involving the discrimination of relative FD differences under noise-free conditions. While older adults initially performed worse on both depth tasks at baseline, the magnitude and trajectory of learning were highly similar between the younger and older adults. Critically, learning effects were largely task dependent in both age groups: training on the FD discrimination task yielded improvements not only on the trained task but also generalized to enhance performance on the SNR depth task. By contrast, training on the SNR task yielded task-specific gains, with minimal transfer to the FD discrimination task. Notably, the degree of improvement on both trained and untrained tasks did not differ significantly between the younger and older participants, and both age groups exhibited comparable learning rates over the course of training.

We first consider the baseline performance differences between the two age groups. Consistent with previous findings that stereoscopic function declines with healthy aging,^{10,18,19,20,21} older adults demonstrated significantly poorer performance than younger adults on both the SNR and FD discrimination tasks. Multiple factors likely contribute to the age-related decline observed across both depth discrimination tasks. Healthy aging has been associated with elevated monocular contrast thresholds, reflecting reduced effective input gain, as well as alterations in interocular suppression processes,⁵¹ both of which may influence the fidelity of binocular integration required for stereopsis.^52,53 Beyond these alterations, aging is also characterized by more global alterations in visual cortical processing, most notably as: (1) elevated internal noise, reflected in increased intrinsic response variability,^39,45,54,55 and (2) reduced efficiency in filtering external noise.^38,39 While these mechanisms have not been directly assessed in the context of stereoscopic tasks, either in previous work or in the present study, they nonetheless provide a plausible account of the performance deficits observed in both depth discrimination tasks among older adults.

Somewhat unexpectedly, older adults in our study demonstrated greater impairment on the FD discrimination task than on the SNR depth task, a pattern that, at first glance, appears inconsistent with previous findings in motion and biological motion perceptions, where age-related deficits are often amplified under conditions of added visual noise.^38,39 Importantly, however, “noise” manipulations are not directly comparable across visual domains. The two depth tasks employed in the present study differ not only in the presence or absence of external noise but also in the type of disparity cues they emphasize and the nature of the perceptual judgments they require. The observed pattern, therefore, does not necessarily contradict prior findings in motion processing, but instead suggests the importance of task-specific computational demands in shaping how aging impacts visual performance. In line with this interpretation, one plausible account is that the FD discrimination task places greater demands on relative disparity processing, requiring precise comparisons of subtle depth differences between adjacent elements. Such computations are likely to be particularly sensitive to age-related reductions in the fidelity of binocular encoding and integration.^51,56,57 Even modest increases in internal neural noise or imprecision in disparity representations may disproportionately impair performance when the task hinges on detecting minimal depth differences. By contrast, the SNR task is more detection-like in nature and may allow observers to rely more heavily on coarser disparity information (e.g., disparity relative to a zero/planar reference) and/or on global, feature-independent mechanisms of signal-noise segregation, potentially mediated by parietal regions.⁴¹ Accordingly, the SNR task could be comparatively less dependent on the finest disparity mechanisms that constrain fine discrimination. It is important to acknowledge, however, that this interpretation remains speculative and should not be considered as the sole explanation for the observed performance pattern.

We next turn to the changes observed following perceptual learning. Our results revealed that both younger and older adults exhibited significant improvements in the respective tasks they were trained, with statistically comparable learning rates and magnitudes across age groups. This pattern of results indicates that, despite poorer baseline performance in older adults, the efficiency and extent of performance gains over time were equivalent to those observed in younger adults. Such age-invariant learning suggests that, within the domain of stereoscopic depth perception, the aging visual system retains a substantial capacity for experience-dependent functional enhancement—one that can approach the plasticity typically observed in younger individuals. Importantly, these learning gains were achieved following only three days of training, with each session lasting approximately 45 min. This rapid improvement highlights the potential of brief, targeted perceptual learning protocols to drive meaningful and efficient functional changes in the aging visual cortex.

In addition to characterizing learning magnitudes, a key interest of the present work was to examine the pattern of learning transfer between depth discrimination training performed under clear versus noisy visual conditions in older adults. While asymmetric transfer of learning between these two conditions has been well documented in younger adults,^41,42,44,58 the current findings demonstrate that older adults exhibit a similar pattern of learning transfer: training on fine (noise-free) disparity discrimination not only improved performance on the trained task but also generalized to enhanced performance on the SNR depth task, whereas training specifically to extract depth signals within noise yielded only minimal improvements on the FD discrimination task. This asymmetric learning transfer is consistent with reweighting-based accounts of perceptual learning,⁴⁴ in which training with clear stimuli strengthens the effective readout of task-relevant sensory information, thereby supporting broader generalization. Conversely, training under noisy conditions may preferentially strengthen the processes involved in excluding or suppressing task-irrelevant inputs, potentially yielding more context-specific improvements that do not generalize to tasks that do not place the same demands on noise exclusion.

That older adults exhibit a comparable pattern of asymmetric transfer is particularly noteworthy in light of prior evidence that healthy aging is associated with reduced neural selectivity, along with increases in internal noise and alterations in gain control across visual domains, as demonstrated through both psychophysical and neurophysiological measures.^38,39,54,59 The broadly similar learning profile across age groups is consistent with the idea that fundamental mechanisms supporting perceptual learning, such as reweighting of task-relevant sensory signals, remain at least partly preserved in later life. While the present behavioral data do not directly speak to the underlying neural mechanisms, causal evidence from younger adults shows that depth discrimination training under distinct task demands can induce targeted functional reorganization along the dorsal and ventral visual pathways. Specifically, before training, repetitive transcranial magnetic stimulation (rTMS) over the PPC selectively impairs performance on the SNR depth task, whereas stimulation over the lateral occipital complex (LOC) disrupts fine depth discrimination. Remarkably, following training on the fine disparity task, SNR task performance becomes susceptible to disruption at LO rather than PPC, indicating a shift in the cortical locus of task-relevant computation. This finding provides a plausible neural-level instantiation of training-induced functional reweighting in younger adults. Whether older adults achieve comparable behavioral transfer via the same circuit-level reweighting, however, remains an open question.

Speculating about the neural basis of perceptual learning in older adults based solely on behavioral data is likely to be more complex and warrants caution. The comparable magnitude, rate, and pattern of learning observed across age groups do not necessarily imply the engagement of identical neural substrates; similar behavioral outcomes may instead arise through partially distinct neural pathways in older adults. One possibility is greater reliance on broader or alternative networks with age,^60,61,62 consistent with evidence for a posterior-anterior shift in aging, whereby reduced occipital engagement is accompanied by increased activation in frontal (including prefrontal) regions during visual perceptual task performance.^62,63,64,65 Extending this compensatory account, age-related neural dedifferentiation may yield less selective and more distributed activation across visual systems,⁶⁶ potentially altering how training influences sensory readout and transfer. Supporting the idea that similar behavioral improvements may be mediated by different neural changes, prior work has demonstrated significant increases in white matter fractional anisotropy beneath the early visual cortex following perceptual training in older, but not younger, adults, despite comparable behavioral improvements.⁶⁷ Future work combining the present stereoscopic depth learning paradigms with neuroimaging and/or brain stimulation across age groups will be critical to determining whether older adults achieve asymmetric learning transfer through similar mechanisms of sensory reweighting, broader compensatory recruitment, or a combination of both.

Beyond advancing our understanding of binocular plasticity in the aging visual system, the present findings also carry potential functional significance for aging populations. As noted earlier, stereopsis plays a critical role in near-to-intermediate distance judgments that support everyday visuomotor behaviors, including obstacle avoidance, negotiating steps and curbs, and precise hand-object coordination. Even modest enhancements in stereoscopic sensitivity could, in principle, benefit these activities—domains in which stereo deficits have been associated with increased fall risk and motor vehicle collisions in older adults.^14,15,16,17 However, it is important to note that the present study assessed depth perception using controlled laboratory tasks, and as such, we cannot assume that training-induced gains necessarily generalize to real-world functional outcomes. A critical next step will be to pair stereoscopic training with ecologically valid performance measures, such as obstacle navigation,⁶⁸ step descent,⁶⁹ and mobility and balance assessments,⁷⁰ to determine whether improvements in laboratory-based stereopsis predict meaningful enhancements in everyday visuomotor functioning among older adults.

For these potential functional benefits to be meaningful, however, it is essential that training effects not only generalize beyond the laboratory but also persist over time. Evidence for long-term retention in stereoscopic learning is limited in normally sighted adults, although follow-up studies in younger adults suggest that improvements can persist for months.^42,71 In older adults, lasting benefits of perceptual training have been documented for several visual functions, including spatial contrast sensitivity,⁷² contour integration,⁴⁸ and texture discrimination,⁴⁶ with improvements persisting for at least three months; however, long-term retention of stereoscopic learning has been examined only rarely, with one study reporting maintained gains in stereoscopic acuity in eight of eleven older adults up to six months.⁴⁰ To explore the long-term effects of our training paradigms, we conducted an exploratory follow-up in which eight participants (n = 8; 3 younger adults and 5 older adults) from the original sample were retested on both depth tasks at an average of 492 days (∼16 months) after the initial post-test. Descriptively, the participants performed comparably to, or slightly below, their post-training levels (Figure S4), while remaining substantially better than their pre-training baseline, suggesting that training-related benefits may persist over extended intervals. Although the small sample size and potential self-selection of returners preclude formal statistical analysis, these preliminary observations provide suggestive evidence for long-term retention in both younger and older adults. Moving forward, future longitudinal studies with larger samples and systematically planned retention intervals will be essential for identifying the optimal training protocols that maximize and sustain perceptual learning effects on depth perception across the adult lifespan.

By directly comparing depth discrimination learning under noisy versus noise-free conditions in younger and older adults, we show that despite a four-decade age gap, older individuals demonstrate learning that is comparable to that of younger adults in terms of both magnitude and rate. Importantly, both age groups showed a similar asymmetric pattern of transfer: training on FD discrimination tasks in noise-free displays not only enhanced performance on the trained task but also generalized to improved extraction of depth signals embedded in noise, whereas training with noisy depth stimuli did not yield reciprocal benefits to FD discrimination. Our findings indicate that the aging binocular visual system retains a substantial degree of functional plasticity, potentially comparable to that of younger adults. While many important questions remain, such as the neural mechanisms supporting this preserved plasticity and the long-term retention of learning effects, the current results offer a promising foundation for the development of targeted perceptual learning protocols aimed at mitigating age-related declines in binocular visual functions.

Limitations of the study

We note a methodological limitation related to the fixed near-viewing distance (50 cm) used throughout training and testing. In the absence of standardized near-refraction or near-add correction beyond participants’ habitual prescriptions, age-related declines in accommodative function may have introduced individual variability in retinal image quality and visual effort.⁷³ This, in turn, could have subtly increased measurement noise and reduced sensitivity to detect training-related changes. While sustained near-viewing demands were mitigated through brief task blocks and frequent rest intervals, and none of the participants reported overt blur or discomfort, the potential influence of subtle, unreported optical limitations cannot be definitively ruled out. To more effectively investigate perceptual learning in aging populations while minimizing potential optical confound, future studies should incorporate distance-appropriate refractive assessments in conjunction with standardized measures of visual comfort and fatigue.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Dorita H. F. Chang (changd@hku.hk).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All behavioral data have been deposited at HKU DataHub research repository and are publicly available as of the date of publication at https://doi.org/10.25442/hku.31449905.
• •All original code has been deposited at Zenodo and is publicly available at https://doi.org/10.5281/zenodo.18414906 as of the date of publication.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported by General Research Fund [17612621] from the University Grants Committee (Hong Kong) to D.H.F.C.

Author contributions

Conceptualization, D.H.F.C.; methodology, D.H.F.C.; data curation, L.C.H.N. and K.Y.K.; formal analysis, K.Y.K.; writing – original draft, K.Y.K.; writing – review & editing, K.Y.K. and D.H.F.C.; funding acquisition, D.H.F.C.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Raw behavioural data	This paper, HKU DataHub research repository	https://doi.org/10.25442/hku.31449905
Original MATLAB code	This paper, Zenodo	https://doi.org/10.5281/zenodo.18414906
Software and algorithms
MATLAB (R2019a)	MathWorks	https://www.mathworks.com
Psychtoolbox	Psychtoolbox-3	http://psychtoolbox.org
SPSS (version 25)	IBM	https://www.ibm.com/spss
JASP (version 0.95.4)	JASP Team	https://jasp-stats.org

Experimental model and study participant details

Participants

We recruited a total of 90 participants for this study [45 younger adults (the younger group), aged 18 to 36 years (mean age = 24.11 years; 27 females); 45 older adults (the older group), aged 60 to 77 years (mean age = 68.4 years; 27 females)]. All participants had normal or corrected-to-normal visual acuity as confirmed by a LogMAR visual acuity of ≤ 0.0. Binocular fusion was verified with the Worth 4-Dot test, and stereoscopic depth perception was assessed using the Butterfly Stereo Acuity test, with all participants exhibiting stereoacuity of 40 arcseconds or better. Visual field integrity was evaluated using standard automated perimetry, with all participants demonstrating normal visual fields. Participants were screened based on self-reported history to exclude any neurological, neuropsychological, cognitive, or significant visual impairments. Cognitive functioning was further evaluated using the Montreal Cognitive Assessment (MoCA),^74,75 with all participants meeting established criteria for unimpaired cognitive status. Written informed consent was obtained from all participants, in accordance with the procedures approved by the Human Research Ethics Committee (HREC) at The University of Hong Kong (Reference number: EA200298). Within each age group, participants were randomly assigned to one of three training conditions: signal-in-noise (SNR) depth training, fine depth (FD) discrimination training, or a no-training control group, with 15 participants per condition. Age and sex distribution for each training group is presented in Table S1.

Method details

Procedures

All participants were tested on the signal-in-noise depth task and fine depth discrimination task before and after training. The order of the tasks was counterbalanced across participants. Participants assigned to the training conditions received task-specific visual training—either on the signal-in-noise depth task or the fine discrimination depth task—over three consecutive days. Each daily training session lasted approximately 45 minutes and comprised 8 blocks, yielding a total of 2,496 trials over the training period. The post-training test was conducted on the next day, following the final training session. Participants in the no-training control group did not receive any training but completed the same tests as the training groups, at matched pre- and post-test intervals.

Apparatus

Stimuli were generated using custom scripts programmed in MATLAB, with extensions from Psychtoolbox. Stimuli were presented via a shutter-presentation setup that consisted of an ASUS 3D-vision-ready LCD display (resolution: 1920 × 1080 pixels; refresh rate: 120 Hz) coupled with NVIDIA 3D Vision 2 shutter glasses. Viewing distance was fixed at 50 cm and stabilized using a chin rest to ensure consistent visual alignment and minimize head movement throughout the experiment.

Stimuli and tasks

Stimuli were random dot stereograms (RDSs) presented against a uniform gray background and surrounded by a binocularly presented grid-like frame comprised of solid black and white squares, each measuring 1.5 degrees in size (Figure 1). This frame functioned as an unambiguous background reference to promote stable binocular fusion. The RDS depicted a central circular target plane with a diameter of 4.5 degrees and a surrounding plane 9 degrees in diameter. Dots within the RDS were randomly assigned as either black or white, positioned at a density of 12 dots/deg², with each dot subtending 0.2 degrees.

For the signal-in-noise depth task, the annular surround was presented at zero disparity, while the central target plane was assigned a disparity of ±6 arcmin (either crossed or uncrossed). Task difficulty was manipulated by varying the percentage of signal dots (i.e., dots with coherent disparity) defining the central target plane relative to noise dots. At 100% signal, all the dots comprising the central target region had the same disparity (i.e. ±6 arcmin), producing a fully coherent depth-defined plane. Task difficulty escalated as fewer signal dots were presented relative to noise dots, whose disparities were randomly drawn from a uniform distribution between ±12 arcmin. At 0% signal, the target plane consisted solely of randomly-positioned noise dots, with no coherent disparity information. The initial test value was set at 80% signal-to-noise ratio. On each subsequent trial, the signal-to-noise ratio of the stimulus was adjusted using the QUEST staircase procedure, estimating the signal threshold corresponding to an 82% accuracy level.

For the fine depth discrimination task, the disparity of the surrounding annular plane was fixed at a disparity of +12 arcmin (a non-zero pedestal disparity), while the disparity of the target plane was varied finely from 1 to 240 arcseconds relative to the surround. This nonzero reference was used to ensure that both the center and surround contained clearly defined depth signals, thereby enforcing a relative disparity judgment rather than a flat-versus-depth detection, which could lead to ceiling-level performance. This approach is grounded in prior work demonstrating that disparity discrimination thresholds are systematically modulated by the presence of a pedestal disparity,^76,77 and it aligns with established paradigms used in both human and animal studies of fine disparity processing.^42,78 The initial test value of the target plane was set to the maximum difference (i.e. 240 arcs relative to the surround). Task difficulty was manipulated by adjusting the disparity difference between the central target and the surrounding plane using the QUEST staircase procedure to estimate the threshold required for 82% correct performance.

For both tasks, participants were required to determine the position of the central target —whether it appeared in front (“near”) or behind (“far”) relative to the surround. Responses were collected using a two-button response box. Each trial lasted a maximum of 3 seconds. Stimuli were presented for 500 ms, during which participants could respond; responses were accepted throughout the entire trial interval. Trials terminated immediately upon response or after the full 3-second duration, whichever occurred first. Trials were separated by a 500-ms interval. Each block (test or training) consisted of two interleaved QUEST staircases of 52 trials. In both pre- and post-training tests, participants completed one practice block and one test block per task. During the training phase, auditory feedback was provided following each response; no feedback was given during the pre- or post-training test sessions.

Quantification and statistical analysis

Depth performance was quantified as threshold estimates for the signal-in-noise depth task and fine depth discrimination tasks at pre- and post-training (lower thresholds indicate better performance). Learning-related changes on each test task were evaluated using separate mixed-design ANOVAs with Age Group (younger, older) and Training Condition (SNR training, FD training, no training) as between-subjects factors and Session (pre-training, post-training) as a within-subjects factor. Model assumptions were assessed using standardized residual diagnostics that included Q-Q plots and Shapiro-Wilk tests for residual normality and Levene's tests for homogeneity of variance. Significant effects were followed by planned Bonferroni-adjusted paired-samples t-tests comparing pre-training and post-training thresholds within each training condition. Baseline equivalence across training conditions was assessed with one-way ANOVAs on pre-training thresholds. In figures depicting learning-related changes (Figures 2 and 3), asterisks denote p < .001 for the specific statistical test specified in the corresponding figure legend.

Learning magnitude was additionally quantified using percentage change scores computed as [(thresh_pre - thresh_post) / thresh_pre]. Percentage change scores were tested against zero using Bonferroni-corrected one-sample t-tests within each Age Group × Training Condition cell and compared across groups using separate 2 (Age Group) × 3 (Training Condition) ANOVAs for each test task. Complementary Bayesian ANOVAs on percentage change scores were conducted in JASP (version 0.95.4) using default JZS priors (fixed-effects r = 0.707). Bayes factors for inclusion (BF_incl) were used to quantify evidence for or against Age Group effects and interactions.

Learning rate during training was estimated for each participant by fitting blockwise thresholds that were smoothed using a 3-point moving average with a two-parameter logarithmic function b(a)=kln(a) +a₀. In this model, b(a) denotes thresholds at training block a, a₀ captures the subject-specific baseline performance, and k indexes the rate of improvement, with more negative values indicating faster learning. The fitted k values were compared using a 2 (Age Group) × 2 (Training Condition) ANOVA and a matched Bayesian ANOVA.

Exploratory associations were assessed with Pearson correlations. Family-wise errors across correlation tests were controlled using the Holm-Bonferroni step-down procedure.

Published: March 26, 2026