Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Perceptual Learning and Transfer

Abstract. Perceptual learning describes improvements in sensory-guided judgments and actions produced by experience; transfer identifies the conditions under which these gains benefit unfamiliar stimuli, locations, decisions, or environments. This fundamental research monograph traces classical visual discrimination experiments, quantitative psychophysics, auditory training, computational accounts of sensory channel reweighting and task rules, neural evidence from visual and decision systems, and research on consolidation and intervention. Its methodological focus is the separation of sensory sensitivity, decision criterion, procedural familiarity, and learned response mapping. Experimental findings establish both narrowly specific learning and broader generalization made possible by training sequence, transfer-test precision, and shared task structure. Critical assessment addresses measurement validity, alternative neural interpretations, control conditions, replication, individual differences, and the limits of clinically relevant extrapolation. The clinical evidence section critically examines amblyopia training trials and a published randomized-trial meta-analysis. The synthesis treats perceptual improvement as a family of testable multilevel changes, with transfer assessed as an experimentally defined relation among training, novel tasks, and time. The review is a critical narrative scientific synthesis, not a newly registered systematic review or meta-analysis.

Keywords: perceptual learning; transfer of learning; stimulus specificity; perceptual discrimination; psychophysical thresholds; visual plasticity; auditory learning; external noise; channel reweighting; double training; retinotopic generalization; consolidation; perceptual decisions; amblyopia; artificial perception

The Scientific Object of Perceptual Learning and Transfer

Perceptual learning is an experience-dependent improvement in the extraction, discrimination, categorization, or use of sensory evidence. Transfer refers to a measurable benefit of that learning under conditions different from those directly practiced. A training benefit is defined operationally through performance on specified stimuli and decisions, rather than through a presumed change in sensory neurons. Researchers ask which information an observer learns to use, how practice changes the mapping from signals to responses, how stable the change becomes, and under which new conditions it remains useful. Orientation, texture, motion, spatial frequency, temporal interval, and contrast tasks provide controlled cases. The historical experiments of Fiorentini & Berardi (1980) and Karni & Sagi (1991) show why stimulus specificity became a central diagnostic of learning.

The phenomenon spans several levels of analysis. Sensory representations may change their precision, downstream decisions may combine stable sensory channels more efficiently, attention may change which information receives weight, and a decision criterion may move without an equivalent change in discriminability. Learning also unfolds over multiple time scales, so the same participant can show rapid within-session improvement, delayed gains, and long-term retention. These candidate processes make a behavioral learning curve an informative outcome but an incomplete explanation. Dosher & Lu (1998) established a quantitative strategy for separating improvements in filtering noisy inputs from other changes, while Sasaki et al. (2010) surveyed the evidence for mechanisms extending beyond any one visual area.

The central question of this dossier is how a trained perceptual improvement generalizes across unfamiliar stimulus features, retinal locations, response rules, environments, and modalities. A strict transfer test compares trained and untrained conditions with baseline equivalence, adequate controls, and explicit measures of precision. A failure to transfer can arise from changed stimulus statistics, a stricter untrained criterion, an unfamiliar response mapping, or interference, as well as from highly selective sensory plasticity. Conversely, apparent transfer can result from shared procedures or ordinary test repetition. The research therefore treats generalization as a property of a precisely described relation between training and evaluation, not as a global attribute assigned once to a person. Jeter et al. (2009) directly illustrates the importance of test demands.

From Early Discrimination Experiments to Modern Plasticity Research

Historical study of perceptual skill joined experimental psychophysics to theories of learning long before neural recordings could identify candidate circuits. The Italian experiments of Fiorentini & Berardi (1980) demonstrated improvements in distinguishing grating waveforms that depended on trained orientation and spatial frequency. Their importance lay in a testable specificity pattern: practice improved a narrow part of a sensory problem, suggesting involvement of representations selective for those stimulus properties. Subsequent work expanded this approach from orientation to texture segregation and other discriminations. Historical claims about an exact neural site, however, required separate evidence because several decision systems can depend on the same narrowly tuned signals.

The texture discrimination experiments of Karni & Sagi (1991) found long-term learning that was tied to retinal location, background orientation, and the trained eye. Such findings reshaped assumptions about the adult nervous system: a repeated visual judgment could change substantially after childhood, while retaining dependence on the precise sensory input. Follow-up work by Karni & Sagi (1993) examined temporal evolution and delayed improvement after training. The distinction between what is learned during exposure and when improved performance is expressed became central to consolidation research. Evidence for selectivity was strong within the tested paradigms, but it did not establish that all perceptual tasks must display identical transfer limits.

A second experimental tradition asked how much of observed improvement could be accounted for by more efficient use of information already present in a stimulus. Gold et al. (1999) assessed changes in signal and noise contributions to recognition, and Dosher & Lu (1998) introduced controlled external-noise manipulations with formal observer models. Later experiments challenged simple location-specific accounts through training protocols that promoted transfer. Xiao et al. (2008) showed that an additional training task at a second retinal position could permit transfer of an originally location-specific perceptual skill. The resulting historical trajectory is an accumulation of experimentally separable mechanisms, from local sensory selectivity through flexible readout and training-dependent generalization.

Conceptual Boundaries: Learning, Adaptation, Attention, and Familiarity

Learning is evaluated by relatively persistent improvement attributable to experience, whereas sensory adaptation often refers to changes in responsiveness produced by recent stimulation that may decay after removal of the input. The same session can contain adaptation and learning with opposing effects on measured accuracy. Repeated testing can also teach an observer the location of a display, the timing of a response interval, or the keyboard mapping. Those gains can be scientifically meaningful procedural skills while remaining distinct from a claimed improvement in orientation sensitivity. Censor & Sagi (2008) investigated how intense testing and adaptation interact with consolidation, demonstrating that short-term performance must be interpreted against the recent history of stimulation.

Attention modifies the selection, enhancement, or weighting of sensory evidence. An instructed participant may become better at knowing when and where to attend, producing improvements that transfer across stimuli even if early sensory coding remains stable. By contrast, some training effects occur when a feature is outside the participant's explicit task. Watanabe et al. (2001) reported learning for a task-irrelevant visual signal without reported awareness of that signal; subsequent analysis in Seitz & Watanabe (2009) elaborated the circumstances of task-irrelevant perceptual learning. Such results require a careful distinction between a feature being irrelevant to the overt judgment and its having absolutely no association with internal reinforcement or attentional states.

Familiarity denotes increased exposure to examples, but its effect can be either useful learning or memorization of incidental details. A participant who recognizes ten repeated images is not necessarily better at discriminating unfamiliar objects, and an observer who shifts a response threshold may report more targets without gaining sensitivity. Sound experimental inference separates identification memory, categorical rules, sensory discriminability, reaction time, and subjective confidence. The studies of Jeter et al. (2009) show that the precision required during evaluation can determine whether past practice seems stimulus-specific. For this reason, all claims about transfer need an explicit account of the probe task, its difficulty, and the information shared with training.

Operational Definitions and the Geometry of Transfer

Specificity is a statistical pattern: learning measured on one stimulus or task exceeds learning measured after a specified change to training conditions. Transfer is the complementary evidence of a benefit at an untrained condition relative to an appropriate control or counterfactual. Investigators can change one dimension at a time, such as orientation, retinal position, spatial frequency, eye, task rule, or response mapping, producing a transfer gradient rather than a binary judgment. The orientation and spatial-frequency effects in Fiorentini & Berardi (1980) supply a classic example of narrow generalization, while the position changes in Xiao et al. (2008) identify conditions where transfer became much broader.

Near transfer concerns new conditions closely related to practice, while far transfer concerns outcomes with a more distant combination of stimuli, tasks, and environments. These labels should be backed by a transparent distance metric because conceptual similarity need not mirror the similarity seen by a sensory representation. Two displays may look different to a reader yet share the same diagnostic orientation signal; conversely, nearly identical images may demand a different discriminative rule. A rigorous design describes feature overlap, response mapping, feedback contingencies, temporal structure, and performance criteria separately. Zhang et al. (2010) argues that acquired rules and task structure can help explain both specificity and generalization.

Transfer testing also has a temporal dimension. Immediate post-training tests may detect procedural familiarity, whereas delayed tests can probe retention and consolidation. Serial testing of many new conditions can unintentionally train the observer on the test set, producing an apparently broad transfer matrix. Order counterbalancing, independent transfer cohorts, and a small number of diagnostic probes protect interpretation. The time-course findings of Karni & Sagi (1993) make it essential to separate training exposure, latent change, and expression of learning. The transfer surface should ultimately be represented as a multidimensional function over task demands and time, with uncertainty bounds estimated for each region actually sampled.

Psychophysical Measurement and Observer Models

A standard discrimination experiment measures accuracy across controlled stimulus differences. When responses are fit with a psychometric function, threshold is the difference yielding a prespecified performance level, and slope describes how sharply performance changes with stimulus strength. A threshold reduction is compelling when lapse rates, guessing, presentation timing, and response biases are estimated or constrained. For yes–no detection, sensitivity and criterion must be distinguished; for forced-choice identification, the number and balance of alternatives matter. Perceptual training may improve the observer's mapping from evidence to response without changing every component of the sensory distribution. The external-noise analysis of Dosher & Lu (1998) demonstrates how model-specific patterns can sharpen interpretations.

Equivalent input noise, perceptual templates, and decision-weight models generate different predictions about the relation between stimulus strength, added noise, and observed accuracy. External noise is useful because a learner might exclude irrelevant image fluctuations while remaining unchanged in a noiseless display, or might improve internal representation enough to benefit in both cases. Dosher & Lu (1998) reported performance patterns consistent with channel reweighting and improvements in external-noise exclusion and stimulus enhancement within its task. Such estimates depend on model assumptions and experimental calibration; they are explanatory constraints, not photographs of a neuronal change. Gold et al. (1999) offers a distinct signal-versus-noise analysis of learned recognition.

Psychophysical designs require explicit pretraining and posttraining tests, matched stimulus generation, stable viewing geometry, fatigue monitoring, and assessment of the amount of practice. Feedback should be documented because trial-by-trial correctness information can alter learning dynamics. Investigators should report individual curves rather than relying solely on group means when participants begin near ceiling or floor. Precision comparisons should use matched baselines or hierarchical models to limit regression-to-the-mean artifacts. The findings of Wright & Sabin (2007) show that required daily training can differ across auditory tasks, making dosage an experimental variable. Strong methods therefore combine calibrated measurement, transparent model fitting, and planned generalization probes.

Experimental Designs and Causal Identification

Random assignment to trained and active-control conditions is valuable when learning is being attributed to a specific intervention. A repeated-measures baseline supplies information about each participant's initial sensitivity but also introduces test exposure that can produce learning. Separate control groups can estimate improvement from simple repetition, feedback, changing motivation, or contact with the researcher. Cross-over designs are economical yet require scrutiny of carryover: perceptual skill can persist after a nominal washout period. Studies of sensory specificity, such as Karni & Sagi (1991), are informative because they compare performance across controlled stimulus transformations within the same research program.

The critical causal contrast is between the change that would have occurred with the trained perceptual task and the change that would have occurred under an otherwise comparable control. A transfer finding needs its own contrast, not merely a significant posttest difference in one group. The uncertainty of a training-by-test interaction should be reported, and null results should be interpreted in the context of sensitivity to detect a meaningful transfer effect. Some investigations use masked stimuli, reward pairing, or decoded neural feedback to isolate exposure from overt attention. Experiments such as Seitz et al. (2009) and Shibata et al. (2011) are valuable precisely because they manipulate aspects of the learning experience usually confounded in ordinary practice.

Mechanistic claims need converging designs. Psychophysical improvement alone cannot identify whether V1, parietal areas, or altered decision weights caused the gain, and fMRI activation alone cannot establish a corresponding change in perceptual sensitivity. Longitudinal recordings, adaptation tests, perturbations, computational fits, and carefully matched behavioral tasks provide partly independent constraints. The primate recordings of Law & Gold (2008) and visual-cortical findings of Schoups et al. (2001) illustrate that different tasks may recruit changes at different cortical loci. An adequately powered, preregistered, multi-session study can test a mechanistic prediction without assuming that one paradigm exhausts the phenomenon of perceptual learning.

Orientation and Spatial-Frequency Discrimination

Orientation discrimination offers an experimentally precise case: angular separation is manipulated while contrast, size, presentation time, and response mapping can be controlled. Fiorentini & Berardi (1980) found that improvement with grating practice was selective for trained orientation and spatial frequency. This finding motivates competing accounts involving tuned feature detectors, learned weighting of preexisting detectors, and task-specific decisions. A threshold decrease alone cannot decide which account is correct because all three may produce similar psychometric curves. Training data should consequently include an array of orientations and independent transfer probes rather than a single contrast between practiced and unpracticed stimuli.

An apparently narrow skill can have a broader usable range under different testing requirements. Jeter et al. (2009) found that task precision at transfer strongly influenced whether orientation discrimination learning generalized. Low-precision probes detected gains where demanding transfer tests revealed specificity. The result underscores a measurement issue: an observer may acquire a partially reusable improvement that is sufficient for easy comparisons but inadequate for a novel high-resolution judgment. Investigators should estimate transfer at several difficulty levels and examine the shape of the psychometric function, rather than inferring a fixed neural localization from one operating point.

Ahissar & Hochstein (1997) linked training difficulty to specificity in a reverse-hierarchy proposal, where more demanding conditions recruit finely selective visual representations. The resulting hypothesis can coexist with transfer-test effects if training determines available information and evaluation determines how much precision is needed. A factorial experiment manipulates training difficulty and transfer precision separately, holds feedback contingencies constant, and tests whether the interaction follows a defined computational prediction. Failure to control reward frequency or fatigue would permit alternative explanations because harder tasks change several psychological processes simultaneously. Specificity is therefore a measurable design-dependent gradient whose underlying mechanism must be inferred through additional tests.

Texture Learning and Retinotopic Organization

Texture discrimination requires the observer to detect local arrangements against a background of similar elements. Karni & Sagi (1991) reported durable learning that depended on retinal location, background orientation, and eye of presentation. These selective effects are compatible with activity in early retinotopic pathways, although later decision systems can themselves access highly localized feature information. The classical result demonstrates that even simple discrimination skills may preserve the particular conditions of practice. Whether the locus of change is sensory coding or selective readout remains an independent physiological and computational question.

Retinal specificity must be measured with attention to fixation. The same monitor coordinate does not guarantee the same retinal input when gaze drifts; eye movements can alter apparent learning, especially with peripheral stimuli. A valid replication tracks eye position, calibrates viewing distance, balances stimulus exposure across quadrants, and checks whether performance changes are limited to a trained location. Interocular transfer adds a separate test because monocular and binocular visual channels overlap but are not identical. The empirical pattern in Karni & Sagi (1991) was a starting point for mechanistic inference, not a universal law covering every perceptual judgment.

A characteristic feature of texture learning is its time course. Karni & Sagi (1993) documented gains that appeared after a delay following training, distinguishing immediate acquisition conditions from delayed expression. This introduced consolidation into the scientific account of discrimination skill. Retest timing, sleep opportunity, recent adaptation, and the number of test trials all influence observed learning. If a delayed test itself contains intensive practice, its improvement is difficult to assign entirely to the intervening interval. A defensible longitudinal study separates exposure, rest, and assessment and measures retention after a prespecified period.

External Noise and Computational Mechanisms

External-noise experiments distinguish conditions in which performance is limited by fluctuating stimulus information from conditions in which the signal is comparatively clean. Dosher & Lu (1998) trained peripheral orientation discrimination while manipulating the amount of noise added to visual stimuli. Within a perceptual-template model, the resulting threshold patterns supported improvements in external-noise filtering and stimulus enhancement. This approach provides a richer mechanistic test than an overall average accuracy score because different latent mechanisms predict different threshold-by-noise functions. Its inference depends on model identification, reliable calibration, and correct treatment of lapse and criterion parameters.

The experiment of Gold et al. (1999) addressed signal and noise contributions to perceptual learning in a recognition task, reporting evidence emphasizing change in signal rather than noise. This result can be compared with perceptual-template accounts only after specifying how each model defines internal variability. Pixel noise, uncertainty about stimulus identity, variable sensory responses, and fluctuations in decision choice are distinct concepts even though each might be called noise. A scientific synthesis should identify which construct a measurement operationalizes and whether alternative models fit the same data equally well.

Transfer experiments extend computational model testing beyond trained stimuli. A learned filter might exclude noise efficiently only for practiced spatial frequencies, while a more general decision rule might apply to unfamiliar image exemplars. Training and testing at several noise strengths, orientations, and response mappings can identify these differences. Cross-validation on held-out stimulus conditions protects against models that simply fit an existing learning curve. The combination of Dosher & Lu (1998) and Gold et al. (1999) supports a program in which observed improvement constrains explicit mechanisms; the strongest theory is the one that predicts generalization before it is measured.

Motion Discrimination and Decision Readout

Random-dot motion tasks vary the proportion of coherently moving elements, allowing investigators to model how a learner aggregates weak motion evidence. The output is typically a categorical direction judgment with a measurable response, making the task suitable for analyzing sensory representation and decision readout separately. Law & Gold (2008) trained monkeys to discriminate motion while recording neuronal activity in the middle temporal area and the lateral intraparietal area. The design provides a link between psychophysical sensitivity and activity in regions associated with visual evidence and its transformation into action.

In the recordings reported by Law & Gold (2008), learning-related changes were associated with motion-driven responses in lateral intraparietal cortex, while the sampled middle temporal responses remained comparatively stable. This observation supports improvement in the use of sensory evidence by decision circuitry. It also places an important limit on inferences from selective psychophysical transfer: a task may rely on a precise stimulus feature without requiring all learning-related plasticity to occur in its earliest sensory representation. The evidence comes from a specified primate task and cannot establish a universal neural locus for human perceptual learning.

A motion-training experiment can test transfer after changing stimulus coherence, display position, or response target. The changes have different theoretical implications: coherence alters evidence strength, position changes sensory sampling, and action remapping tests a learned decision policy. Recording multiple regions and fitting trial-wise evidence-accumulation models would help distinguish those possibilities. Prediction must target the complete relation between stimulus evidence and choice, including response time and error patterns when available. The findings of Law & Gold (2008) make the learned sensory-to-choice mapping a central candidate mechanism rather than a peripheral methodological nuisance.

V1 Plasticity and Longitudinal Neural Evidence

Early visual cortex provides anatomically selective codes for spatial position and orientation, so it is a plausible contributor to finely tuned perceptual learning. The trained-monkey physiology of Schoups et al. (2001) identified changes in V1 orientation coding associated with improved discrimination. The result supports sensory cortical plasticity in a well-defined task and complements behavioral evidence of stimulus specificity. A cortical response difference still admits multiple explanations, including altered tuning, attention, normalization, and task state; it must be interpreted against the measurements used and the predictions of competing circuit models.

Human longitudinal imaging brings another level of evidence. Yotsumoto et al. (2008) followed behavioral performance and V1 blood-oxygen-level-dependent responses across visual practice. Behavioral improvement persisted after the initial period, even though the observed cortical activation pattern followed a different trajectory. The mismatch challenges treating a single activation measure as a permanent storage index of learned skill. Neural recruitment can become more efficient, be redistributed, or interact differently with attention after practice. The study highlights the importance of measuring the full acquisition and maintenance course rather than selecting only the session with the largest response.

A systems account must accommodate evidence for neural plasticity in sensory areas and altered readout in higher-level regions. Karmarkar & Dan (2006) surveyed experience-dependent changes in adult visual cortex, while Law & Gold (2008) supplied a detailed decision-area example. No single experimental method observes every relevant level. Electrophysiology samples selected units, fMRI estimates hemodynamic correlates, and psychophysics constrains performance. Their convergence is strongest when independent measures yield specific, mutually compatible predictions for transfer, intervention, and retention. A broader neural signal after training is evidence of change; identifying what changed requires additional causal or model-based analyses.

Task Difficulty and Reverse-Hierarchy Accounts

Ahissar & Hochstein (1997) proposed that perceptual learning can progress from broad higher-level representations toward precise lower-level information as task demands increase. This reverse-hierarchy account organizes observations of generalization under easy conditions and specificity during more demanding training. It predicts an interaction among stimulus complexity, presentation time, and the level of resolution required by the judgment. As a historical theoretical contribution, it moved the field beyond the expectation that one fixed sensory site would govern all improvement. Its empirical evaluation requires a distinction between processing level and covarying changes in motivation, attention, and correctness feedback.

The original framework also emphasized the possible importance of an enabling experience before difficult practice. This offers a research motivation for training sequences that begin with identifiable signals before progressively reducing stimulus differences. However, easy-first training may improve a task through familiarization, discovery of stimulus structure, or increased reward frequency without uniquely establishing a cortical reverse flow. Designs should match cumulative exposure and compare easy-to-hard, hard-only, and intermixed schedules. The distinction is important for applied training, where recommended schedules must follow direct evidence for durable and useful outcomes.

The precision account of Jeter et al. (2009) supplies a critical counterpart: learning specificity may be revealed or concealed by the evaluation task. A synthesis should place both training difficulty and transfer precision into the same experimental matrix. If transfer is broad on permissive tests but restricted on precise tests after identical practice, the outcome is informative about the capacities and limits of the acquired representation. Explanations should predict the interaction quantitatively rather than invoking either high or low visual processing after observing it. Such a synthesis preserves the historic value of reverse-hierarchy theory while allowing later evidence to refine its scope.

Retinal Location Transfer and Double Training

Location specificity was once treated as especially persuasive evidence that a learned discrimination had altered the earliest retinotopic representation involved in its stimulus. The training arrangement matters, however, because location changes also require observers to select a new spatial region and establish useful connections between that region and task-relevant decisions. Xiao et al. (2008) used a double-training procedure in which participants trained on the main visual task at one location and performed an additional task at another location. The authors reported striking transfer of the original learning across retinal locations under that protocol, altering the field's theoretical interpretation of specificity.

Double training does not imply that retinally selective representations disappear after practice. Rather, it demonstrates that a protocol can enable an acquired improvement to be expressed with sensory input from an otherwise untrained position. Different models attribute this to learned location access, generalized decision rules, reweighting, or interactions among these factors. The rule-based explanation developed in Zhang et al. (2010) accounts for how learning a relevant task rule can become available at additional locations. To distinguish explanations, a new study should manipulate whether secondary training shares the sensory representation, motor response, feedback rule, or merely the location of stimulation.

The transfer outcome must be compared to both primary-only training and secondary-only training controls. Otherwise a posttest benefit at the new location may reflect ordinary learning acquired during the secondary task. Testing at another unexposed position can establish whether the effect extends beyond the two practiced locations. Retinal position also requires stable fixation and matched eccentricity, since changes in acuity and crowding can mimic training interactions. The experiments of Xiao et al. (2008) illustrate a larger methodological principle: specificity is conditional on a history of tasks and interventions, and the appropriate causal explanation is the one that predicts those conditions accurately.

Rule Learning and Perceptual Generalization

An observer can learn to interpret perceptual information through a task rule: which feature is diagnostic, which evidence supports a response, and which variations should be ignored. Such a rule may use stable sensory channels while improving discrimination across trials. Zhang et al. (2010) proposed that rule-based learning explains a range of visual learning and transfer phenomena that appear difficult to reconcile with strictly local feature-tuning accounts. This research helps clarify why perceptual learning is situated at the intersection of sensory processing and cognitive learning, with no requirement that every useful adaptation be assigned to one level of cortical representation.

Rule learning predicts asymmetric transfer when one task establishes information useful for a second task but the reverse direction does not. It also predicts differences between explicitly changing a stimulus feature and changing the operation the learner performs on that feature. Researchers can test these ideas by training orientation judgments and then asking for identification, categorization, or comparison under controlled remapping. The decisive outcome is not simply whether a posttraining score increases, but whether a model that learned a rule from the trained cases predicts novel examples. The experiments of Xiao et al. (2008) provide a relevant location-transfer test of this broader framework.

A rule model should specify what counts as shared structure. A simple verbal instruction may identify a feature without training an effective internal evidence mapping; conversely, extensive exposure can alter task policy without explicit insight. Separating explicit rule knowledge from implicit performance requires interviews, transfer to unfamiliar stimulus families, and ideally a manipulation that makes one rule valid while preserving overall stimulus statistics. Such tests would reveal how much of transfer arises from abstract task structure, sensory channel availability, and reinforcement history. Jeter et al. (2009) reminds us that any claimed rule must also succeed at the precision demanded by the new task.

Task-Irrelevant Learning and Implicit Reinforcement

Perceptual improvement can occur for signals outside the overt focus of a participant's task. In Watanabe et al. (2001), a visual motion signal repeatedly presented without task relevance was associated with later perceptual changes, challenging the necessity of directed conscious attention for every form of visual perceptual learning. The study opened a debate about which aspects of repeated exposure, reinforcement, and internal processing permit a signal to be learned. The inference concerns an experimentally specified relation between exposure and performance; it does not establish that all unattended stimulation is equally effective or that attention never contributes to learning.

Seitz & Watanabe (2003) argued that task-irrelevant learning can be linked to reinforcement generated by a separate successful activity. More direct evidence came from Seitz et al. (2009), which paired visual signals made imperceptible by a competing-eye display with liquid rewards and reported subsequent learned discrimination without conscious access to the signal–reward pairing. This research motivates models of plasticity regulated by reinforcement and exposure statistics. The evidence requires care with the precise masking conditions, motivational state, and transfer task. Reward-associated learning in a laboratory protocol should not automatically be generalized to arbitrary subliminal advertising or everyday persuasion.

A theoretical distinction is needed between an irrelevant feature's sensory representation, its temporal co-occurrence with reward, and the observer's report of awareness. Each can be operationalized independently and may influence learning. Seitz & Watanabe (2009) reviewed task-irrelevant perceptual learning and its boundary conditions. Researchers should test whether learning transfers when reward contingencies change, whether effects depend on stimulus strength and presentation timing, and whether awareness measures are adequately sensitive. The broader lesson is that purposive practice is one route to perceptual plasticity among several experimentally identified routes, with reinforcement and task structure changing the scope of generalization.

Reward, Feedback, and Reinforcement Contingencies

Feedback specifies whether a response is correct, while reward can provide a reinforcing event that changes the value of states or actions. During perceptual training, these signals may help select useful stimulus dimensions, strengthen a response rule, or alter attention. Seitz et al. (2009) demonstrated learning from reward pairing under conditions where the relevant visual feature was not consciously available. This does not mean all learning depends on explicit external prizes: correct performance, informative structure, and task engagement can supply distinct signals. Experimental designs must describe the timing, probability, and information carried by feedback rather than classifying every training protocol simply as rewarded or unrewarded.

Reward and feedback interact with difficulty. A task near chance performance provides limited reliable correctness information; a trivially easy task may offer little need for improved resolution. Studies motivated by Ahissar & Hochstein (1997) therefore should separate perceptual difficulty from the distribution of reinforcement. Yoking rewards across task conditions can help isolate the effect of stimulus precision, although it may change participants' experience of control. In a transfer test, the original response mapping should sometimes be reversed or withheld to see whether training established flexible sensitivity or a reward-linked response habit. Results must be evaluated using discriminability measures as well as accuracy.

A reinforcement-sensitive explanation can be integrated with sensory reweighting if reward changes which channels guide the final choice. Such integration demands a mechanistic prediction: the model should forecast how learning differs with reward timing, competing stimulus features, and transfer to a new response domain. Zhang et al. (2010) provides an account of rule-related generalization, while Dosher & Lu (1998) supplies noise-dependent behavioral constraints. Together they suggest experiments in which feedback contingencies and external noise are manipulated factorially. The objective is to identify when reinforcement teaches a sensory selection rule and when it yields improvement limited to the trained stimulus–response association.

Generalization Across Stimulus Families and Tasks

A narrow transfer test varies one perceptual dimension, whereas transfer across stimulus families also changes the statistical structure of examples. Recognition of unfamiliar faces, patterns, or objects may depend on discovering category-diagnostic information rather than refining a single orientation threshold. The experiment of Gold et al. (1999) is relevant because it analyzes learning in an identification context where learned signal use can be compared with noise-based alternatives. An adequate theory should explain both the relation between practiced and novel examples and why one type of task might generalize even though a fine discrimination task does not.

Action video game experiments provide a more complex test of cross-task learning. Green & Bavelier (2003) reported associations between game experience and multiple visual-attention skills and included training evidence in previously untrained players. This work suggests that coordinated, variable demands may produce benefits outside the exact practiced display. However, multiple components of gameplay change simultaneously, and generalization to unmeasured intellectual or real-world outcomes cannot be assumed. Randomization, active game controls, outcome blinding, and replication are needed before identifying which feature of a complex intervention causes a transferable improvement.

The useful comparison is between targeted training with a simple stimulus and variable training with many related tasks. A complex environment might teach flexible allocation of attention, while narrow psychophysics may achieve higher precision on a specific channel. Neither approach guarantees a universal transfer advantage because training dose, test similarity, and baseline differences can dominate outcome. Jeter et al. (2009) emphasizes the precision constraint on transfer, and Green & Bavelier (2003) illustrates the interest and complexity of wider behavioral generalization. A research program should compare standardized transfer batteries rather than selecting only favorable posttests for a particular intervention.

Near Transfer, Far Transfer, and Ecological Validity

Near and far transfer are operational relations among trained and tested situations. Distance can include a change in stimulus dimension, retinal position, sensory modality, environmental statistics, or response rule. Researchers frequently combine these differences into one vague designation, making it difficult to compare studies. A useful report presents a transfer matrix describing each change and includes the trained condition as an internal reference. Xiao et al. (2008) identifies a specific manipulation of retinal location, while Jeter et al. (2009) illustrates a different source of variation in transfer precision. Their results should not be merged into a single percentage for generalization without a common definition.

Ecological validity asks whether a laboratory improvement matters under everyday conditions where inputs vary and responses are embedded in ongoing behavior. Better recognition of a repeatedly trained noisy pattern may not improve identification of new objects in natural illumination. Conversely, training on varied examples can appear weaker at a fixed laboratory threshold but benefit more natural decisions. Ecological evaluation should use outcomes established before the intervention, multiple real-world contexts, and sufficient retention intervals. Findings from Green & Bavelier (2003) encourage such investigation but do not license untested claims about employment, education, or general intelligence.

A transfer index should also report the opportunity cost of additional training and whether a beneficial change persists without supervision. A large relative threshold gain at a narrow trained frequency may be practically less valuable than a smaller reliable gain across many relevant conditions. Reporting uncertainty and clinically or functionally meaningful benchmarks is essential. When the outcome is performance in a different sensory modality or cognitive domain, the hypothesized common mechanism must be specified. The scientific value of far-transfer research lies in identifying transferable operations and demonstrating them with controlled new tasks, not in attaching an expansive label to a promising training result.

Acquisition Curves and Within-Session Change

A perceptual learning curve is the relation between performance and experience, often indexed by practice blocks, trials, or days. Early improvement may be rapid because observers learn instructions, timing, and response conventions; subsequent gains may depend on more stable perceptual adjustments. Karni & Sagi (1993) identified temporal components in texture discrimination that could not be summarized by a uniform within-session progression. Modeling requires inspection of the data rather than imposing a universally exponential or power-law curve. Participants may approach a ceiling, fluctuate with fatigue, or begin at different baseline thresholds, so a single group-level fit can conceal heterogeneous acquisition pathways.

Learning rate is sensitive to task design and stimulus quality. Highly difficult trials may offer limited reliable information, whereas easy exemplars can reveal diagnostic features that support later discrimination. The reverse-hierarchy experiments of Ahissar & Hochstein (1997) motivate testing how difficulty sequences change acquisition. Within-session performance should be separated from across-session change, with rest, feedback, and stimulus order recorded. A plateau might mean stable skill, diminished motivation, rising adaptation, or insufficient sensitivity of the outcome measure. Differentiating these possibilities requires an independent probe or altered task condition rather than interpreting every flat curve as a hard biological limit.

A strong longitudinal model uses trial-level data and includes individual random effects, response bias, and time-varying lapse parameters where applicable. The model should be validated against held-out blocks and delayed assessments. This permits comparisons of whether training alters a threshold, slope, learning rate, or asymptote, which carry different scientific implications. When transfer is the question, acquisition on the trained task should not be mistaken for evidence that an untrained task improved. The design must measure both trajectories and the uncertainty around their difference. Wright & Sabin (2007) demonstrates how learning dynamics can vary appreciably across sensory tasks under comparable schedules.

Consolidation, Sleep, and Delayed Expression

Consolidation refers to processes by which acquired changes become stabilized, transformed, or better expressed after practice. Visual texture discrimination has provided important evidence for delayed gains. Karni & Sagi (1993) reported performance improvements emerging after intervals between training sessions, encouraging hypotheses about offline plasticity. Delayed behavioral gain, however, is an outcome measure: identifying sleep-specific physiology or a particular synaptic mechanism requires an experimental manipulation beyond elapsed time. An overnight comparison can confound sleep, circadian phase, passage of time, and relief from fatigue. Each alternative is relevant when interpreting a learning curve.

Later paradigms revealed variation in sleep dependence. Aberg et al. (2009) reported that learning with chevron stimuli could transfer to untrained directions and did not require sleep for the observed improvement. This finding cautions against generalizing the consolidation profile of one texture task to all perceptual training. Training dosage and stimulus configuration may determine whether offline gains are measurable. Censor et al. (2012) reviewed commonalities between perceptual and motor skill consolidation, while emphasizing the complexity of interactions among primary and higher-order systems. A plausible theory must predict which learning components are stable after waking rest, sleep, or additional practice.

Consolidation research benefits from randomized sleep and wake intervals when feasible, equal retest timing, and control of cumulative exposure. Researchers should record adaptation and tiredness before every session, because recovery from fatigue may mimic an offline gain. Several transfer tests can distinguish whether consolidation broadens a skill or merely protects the trained discrimination. Censor & Sagi (2008) found that training duration and subsequent intensive testing affected the resistance of acquired performance to adaptation. Thus consolidation is relevant both to the magnitude of a trained improvement and to its resilience under a changed testing context.

Interference, Adaptation, and the Spacing of Practice

More practice is not invariably better at every point in a learning schedule. Repetition can strengthen a discriminative rule while prolonged exposure creates adaptation or fatigue that temporarily lowers accuracy. Censor & Sagi (2008) investigated how compact training could promote long-term resistance to the disruptive effects of intensive assessment, illustrating the practical importance of distinguishing consolidation from immediately measured performance. A densely sampled learning curve can itself be an intervention that changes the phenomenon under study. Researchers therefore need planned assessment intervals and exposure accounting when comparing protocols.

Spacing manipulations alter the distribution of practice across sessions. They may influence retention by reducing interference, supporting stabilization, or permitting recovery from adaptation. Evidence cannot be imported without qualification from generic declarative memory tasks because perceptual discrimination has its own demands on stimulus representations and feedback. The auditory experiments of Wright & Sabin (2007) found that the training exposure required for improvement differed between frequency and temporal-interval discrimination. A universal minimum number of trials is therefore unjustified. Better designs compare dose–response relationships across tasks using comparable precision metrics and sufficiently long follow-up periods.

Transfer may respond differently to spacing than trained-task performance. A schedule maximizing immediate low-threshold precision might reinforce narrowly useful cues, whereas variable sessions could encourage flexible use of multiple features. This is a hypothesis that requires a randomized comparison with equivalent total practice and independent untrained tests. Carryover should be measured because a prior schedule can influence later learning even after the overt training changes. The combination of Aberg et al. (2009), Censor & Sagi (2008), and Wright & Sabin (2007) supports treating dose, rest, and testing order as essential explanatory variables rather than administrative details of an experiment.

Auditory Perceptual Learning and Temporal Resolution

Auditory discrimination training allows perceptual learning to be evaluated outside the visual tasks that dominate historical accounts. Frequency discrimination measures how well listeners distinguish tones differing in pitch-related frequency, while temporal-interval discrimination asks for judgments about duration. Both demand calibrated sound presentation and methods sensitive to lapse, attention, and task procedure. Wright & Sabin (2007) compared daily training dose across these tasks and found that the amount needed for improvement was not the same. This finding challenges the expectation that one sensory-system-wide schedule governs acquisition across all auditory decisions.

Auditory transfer can be examined across frequencies, stimulus intensities, temporal patterns, ears, and response contingencies. Each dimension changes a different property of the tested information. Transfer between tone frequencies may reflect general comparison rules; transfer across ears can inform the participation of binaural or more central processing; transfer to speech outcomes demands a further account of category structure and language expertise. Results from visual orientation experiments should be used as methodological analogies, not as direct evidence for a particular auditory transfer result. The task-specific findings of Wright & Sabin (2007) illustrate the need for modality-respecting designs and matched sensory measures.

Training on auditory intervals can improve an explicit comparison while leaving unaided everyday listening unchanged. To establish a benefit beyond the trained stimuli, researchers should use new tokens, variable backgrounds, and retention tests. Active controls must receive comparable auditory exposure and researcher contact so that procedural familiarity is not mistaken for a change in sensory resolution. A computational analysis can separate improvements in temporal signal representation from a learned decision rule. The wider theoretical implication is that perceptual learning is a family of task-contingent adaptive changes whose generalization must be demonstrated anew across each sensory domain.

Individual Differences and the Development of Perceptual Skill

Participants differ in initial sensory acuity, attention, task comprehension, prior experience, fatigue, and neural history. These differences can change both observed learning rates and the amount of transfer measurable after training. A person near ceiling may have little room to improve even with meaningful representational change; a person at floor may show apparent nonlearning because the task provides uninformative feedback. The adult improvements established in Karni & Sagi (1991) and the clinical evidence of Levi & Polat (1996) show why the capacity for experience-dependent change should be tested at a range of initial performance levels rather than assumed absent after early development.

Developmental status matters because sensory encoding, response control, motivation, and environmental exposure change throughout life. Comparing children, adults, and older participants requires age-appropriate tasks that preserve the underlying construct. A higher threshold may reflect sensory loss or unfamiliarity with an experimental response procedure, and these produce different predictions for training. The review Sasaki et al. (2010) describes multiple forms of visual perceptual plasticity, but does not supply a universal age-by-training function. Longitudinal developmental work should measure the same participants across time, distinguish maturation from practice, and monitor whether transfer relates to changing sensory experience.

Individual-difference research needs more than reporting who improved most. Reliable predictors of transfer should be measured before training, tested prospectively, and evaluated in independent samples. Candidate predictors include baseline acuity, attentional selection, sleep regularity, motivation, and task strategy. Multiple correlated predictors invite overfitting, especially in small training studies. A hierarchical analysis can estimate heterogeneity while separating measurement error from stable variation, and repeated delayed probes can distinguish temporary fluctuations from durable differences in acquired skill. Such designs convert the general claim that learners differ into testable hypotheses about which conditions allow useful perceptual learning for which populations.

Clinical Vision Research and Amblyopia

The possibility of modifying adult visual performance has clinical relevance where ordinary visual development has produced persistent deficits. In amblyopia, reduced visual acuity can involve abnormal early visual experience and altered binocular interactions. The experimental study of Levi & Polat (1996) reported neural plasticity in adults with amblyopia, demonstrating the value of testing improvement after the conventional developmental window. Polat et al. (2004) subsequently studied perceptual learning as an intervention for adult amblyopia. These results connect laboratory training to measurable visual function, while the exact relationship among trained discrimination, acuity, and binocular performance must be assessed rather than presumed.

The 2025 systematic review and meta-analysis by Ming et al. (2025) assembled randomized trials of perceptual learning and video-game-based visual training for adults with monocular amblyopia. Its published analysis included 22 trials and 422 participants and reported improvement in visual acuity across the interventions analyzed. The conclusion is clinically relevant but requires attention to variation among tasks, control conditions, acuity measures, intervention durations, and risk of bias. Standardized average differences cannot be directly translated into the benefit expected by an individual patient. A research monograph should identify the design and measured endpoints before drawing broader claims about restoration of everyday vision.

The crucial applied question is whether training improves clinically meaningful function outside the practiced exercise, including retention and binocular outcomes. Better task thresholds can be prerequisites for useful gains, but they are not automatically equivalent to improved reading, navigation, or occupational performance. Independent clinical trials with masked outcome assessment, active controls, preregistered endpoints, and adequate follow-up can clarify this relationship. The research of Polat et al. (2004) and synthesis in Ming et al. (2025) support further investigation of targeted training while leaving individual assessment and treatment selection to qualified clinicians. The empirical domain remains adult amblyopia rather than an unrestricted promise of general visual enhancement.

Complex Training Environments and Action Video Games

Complex games combine rapid perceptual sampling, target selection, uncertain feedback, visuomotor coordination, and variable task contexts. Green & Bavelier (2003) showed that action video game experience was associated with multiple measures of visual attention and included an experimental training comparison in nonplayers. This makes game-based training an important case for testing transfer beyond a single laboratory discrimination. The intervention is nevertheless multifactorial. Participants may learn temporal allocation of attention, spatial tracking, response policy, or specific game conventions; a difference in one posttest cannot establish which component generalized.

Proper controls should match expectations, time, engagement, visual complexity, and prior gaming experience as far as feasible. An active alternative game provides a stronger comparison than no contact, although matching all motivational aspects is difficult. Outcome measures should be chosen independently of the games and repeated after a delay to separate familiarity from durable change. The meta-analysis of adult amblyopia interventions by Ming et al. (2025) evaluates a clinically defined visual-acuity outcome rather than the entire range of attentional skills measured in Green & Bavelier (2003). Combining their conclusions without acknowledging this endpoint difference would overstate what either study establishes.

A valuable future design would compare isolated perceptual discrimination, variable-task game training, and a carefully matched active control. The same independent battery could measure near transfer, feature transfer, response remapping, and everyday visual function. Predictions should specify whether broad variability teaches a reusable strategy or whether focused precision practice yields a larger improvement for a narrow feature. Data sharing, documented game versions, calibrated displays, and transparent exclusion criteria would help reproducibility. Strong transfer evidence comes from measured outcomes beyond training, not from branding an intervention as cognitively enriching or highly immersive.

Decoded Neurofeedback and the Causal Study of Representation

Neurofeedback paradigms attempt to influence learning by providing feedback contingent on measured neural activity. Shibata et al. (2011) investigated whether inducing selected visual cortical activity patterns through decoded fMRI neurofeedback could produce orientation-specific perceptual learning without presenting the usual physical orientation training stimulus. The study provides an experimental challenge to explanations requiring ordinary overt practice with the relevant sensory display. Its evidentiary force comes from the relation among decoded neural patterns, reinforcement, and subsequent behavioral measurement. The approach requires careful interpretation because a decoder is a statistical representation of measured signals rather than an isolated neural circuit.

A causal claim requires ruling out learning from visible feedback cues, expectancy, spontaneous rehearsal, task exposure during testing, and selection artifacts. Suitable control conditions can vary the decoded target, contingency, and amount of feedback while maintaining participant engagement. The findings of Shibata et al. (2011) are therefore best understood as evidence that a specific closed-loop neural-training arrangement can alter subsequent perceptual performance under its tested conditions. They do not demonstrate that arbitrary complex knowledge can be installed into a brain, nor that all useful perceptual skill can be acquired without sensory experience.

Neurofeedback is also informative for transfer mechanisms. If a trained pattern improves performance only for a predicted orientation, it constrains the representational specificity of the intervention. If it transfers across stimuli or tasks, computational analysis is needed to identify shared features. Because fMRI measures indirect hemodynamic signals, combining it with psychophysics, control decoders, and delayed testing is essential. Related longitudinal evidence from Yotsumoto et al. (2008) further shows that the relation between cortical activation and behavioral improvement can change across training. A complete neural theory must therefore predict both immediate acquisition and durable functional expression.

Costs of Learning, Biases, and Perceptual Errors

Perceptual learning is usually evaluated through improved accuracy, yet training can also alter what observers are likely to report when sensory evidence is weak. Seitz et al. (2005) reported a training-associated increase in reports of motion in the trained direction even on trials without the corresponding motion stimulus. Such findings make the costs of learned expectations a distinct research problem. An observer can become more efficient in a practiced environment while acquiring biases that are misleading under different stimulus distributions. Sensitivity, false-alarm rate, response criterion, and calibration should therefore be reported alongside positive training outcomes.

Reward-based and task-irrelevant learning may likewise change performance in ways that depend strongly on context. Seitz et al. (2009) provides evidence for training effects linked to imperceptible reward-paired stimuli, while Watanabe et al. (2001) demonstrates that overt task relevance is not always necessary for a later perceptual effect. These phenomena are theoretically important because they involve sensory evidence that an observer has not chosen as the explicit subject of learning. Their practical implications require testing: heightened sensitivity to one pattern may help under a stable distribution but create false positives or distraction under another.

The relevant principle is to evaluate transfer for both gains and costs. A training program should include untrained nontarget distributions, catch trials, shifting base rates, and delayed reassessment. A high trained-task accuracy score can mask a worse decision threshold when the environment changes. Mechanistic models that include criterion shifts and prior expectations can predict when learned specialization becomes maladaptive. Evidence about perceptual errors should be analyzed with the same rigor as evidence about improvement, providing a fuller scientific meaning to the idea of adaptation and its consequences for behavior.

Artificial Perception and the Comparative Problem of Transfer

In the Artificial Era, perceptual transfer can also be examined in trained computational systems that classify sensory signals, fuse modalities, or act on estimated environmental states. A machine-vision system may improve on a specific image distribution while failing under changes in illumination, camera position, or object appearance. This supplies a comparative research problem rather than a license to assume that human and artificial learners possess identical mechanisms. Artificial perception is operationally described through architecture, training data, sensor characteristics, objective functions, and held-out performance. Human perceptual learning is described through experimental task, organism, experience, and behavior. Each requires its own valid measurement design.

Cross-system comparison can ask whether training produces specificity along comparable dimensions, whether improved feature weighting predicts generalization, and whether reinforcement history affects transfer. The concept of channel reweighting developed in human psychophysics by Dosher & Lu (1998) provides a useful mathematical point of comparison, as do the rule-learning experiments of Zhang et al. (2010). A responsible computational analysis states precisely which elements are structurally similar and which are different. An algorithm's performance under distribution shift is measurable; interpreting it as a human-like subjective percept or conscious experience would require a separate theory and evidence.

Research on human–artificial partnerships adds another direction: an observer may learn to perceive through an assistive interface, and interface feedback can change the perceptual strategies used in daily decisions. Measuring this phenomenon requires separating user learning from model adaptation. A jointly improving performance score cannot determine which participant in the coupled system changed. Controlled tests can hold the artificial classifier fixed while training a human operator, and conversely hold human instructions constant while updating model parameters. This distinction connects perceptual learning to broader science of distributed cognition and the psychology of human–artificial relations without collapsing independently measurable processes into one narrative.

Evidence Quality, Replication, and Research Priorities

The evidence base combines small controlled psychophysical experiments, specialized animal physiology, longitudinal human imaging, computational modeling, complex training interventions, and emerging clinical synthesis. Each method answers a different question. Behavioral specificity studies such as Karni & Sagi (1991) identify measured transfer boundaries; primate recordings in Law & Gold (2008) constrain neural readout mechanisms; the meta-analysis of Ming et al. (2025) summarizes selected randomized clinical outcomes. A narrative integration cannot legitimately be called a new systematic review unless it uses a documented exhaustive search, prespecified eligibility criteria, extraction procedures, and quantitative synthesis where warranted. The present monograph is a critical narrative scientific synthesis grounded in directly identified studies and reviews.

Reproducibility concerns include stimulus calibration, eye tracking, choice of precision level, laboratory-specific training schedules, statistical flexibility, small samples, and selective reporting of favorable untrained outcomes. A generalization claim should be tested with a prespecified matrix of stimulus transformations and independent active controls. Authors should report the untrained tests that fail alongside the ones that succeed, use uncertainty intervals around interaction effects, and distinguish confirmatory results from exploratory post hoc patterns. The contrast between Xiao et al. (2008) and Jeter et al. (2009) makes clear that what looks like contradictory evidence can arise because the training protocol and transfer criterion are different.

Priority research questions include how sensory coding and decision readout interact during learning, which sleep or waking intervals support durable expression, how generalization depends on training diversity, how age and baseline acuity modify outcomes, and which clinical improvements remain useful after formal training ends. Future multisite experiments should preregister transfer gradients, share reproducible stimulus code, and collect neural and behavioral data across acquisition and retention. Models should be evaluated through prospective prediction rather than retrospective accommodation. The research value of perceptual learning lies in its ability to connect carefully measured experience-dependent change to explicit hypotheses about adaptive information processing.

Conclusion

Perceptual learning is a measurable change in the efficiency, precision, or organization of sensory-guided performance after experience. The evidence reviewed here establishes durable improvements in multiple discrimination tasks and shows that these improvements can be highly specific to stimulus orientation, spatial frequency, retinal location, task demand, and training history. The foundational observations of Fiorentini & Berardi (1980) and Karni & Sagi (1991) remain scientifically important because they defined reproducible constraints on change. Later experiments demonstrated that specificity depends on more than the spatial scale of a candidate sensory representation: testing precision, learned rules, and additional training can alter the range over which improvement is expressed.

The central empirical synthesis joins several partially compatible mechanisms. Dosher & Lu (1998) supports changes in the effective selection and enhancement of visual signals; Schoups et al. (2001) identifies neural correlates of sensory cortical tuning; Law & Gold (2008) reveals a prominent role for sensorimotor decision readout; and Xiao et al. (2008) demonstrates training-enabled transfer between locations. Karni & Sagi (1993) and Aberg et al. (2009) show why temporal consolidation must be studied as a task-dependent phenomenon. These results collectively support a multilevel framework for explaining learning trajectories and transfer, with the precise mechanism established by the predictions and controls of each experiment rather than inferred from a generic improvement score.

The enduring scientific challenge is to determine what experience changes, when the change stabilizes, and where it remains beneficial under new conditions. Progress depends on calibrated psychophysics, causal controls, validated neural measurement, explicit computational models, and independent transfer and retention tests. Applied evidence, including the clinical synthesis of Ming et al. (2025), extends the question to functional outcomes while requiring endpoint-specific interpretation. Comparative research in the Artificial Era can further sharpen operational definitions by studying the transfer behavior of distinct learning systems. A complete account of perceptual learning is therefore an account of acquired information use, its biological or computational implementation, and the experimentally demonstrated limits of generalization.

References

  • Aberg, K. C., Tartaglia, E. M., & Herzog, M. H. (2009). Perceptual learning with Chevrons requires a minimal number of trials, transfers to untrained directions, but does not require sleep. Vision Research, 49(16), 2087–2094. https://doi.org/10.1016/j.visres.2009.05.020
  • Ahissar, M., & Hochstein, S. (1997). Task difficulty and the specificity of perceptual learning. Nature, 387, 401–406. https://doi.org/10.1038/387401a0
  • Censor, N., & Sagi, D. (2008). Benefits of efficient consolidation: Short training enables long-term resistance to perceptual adaptation induced by intensive testing. Vision Research, 48(7), 970–977. https://doi.org/10.1016/j.visres.2008.01.016
  • Censor, N., Sagi, D., & Cohen, L. G. (2012). Common mechanisms of human perceptual and motor learning. Nature Reviews Neuroscience, 13, 658–664. https://doi.org/10.1038/nrn3315
  • Dosher, B. A., & Lu, Z.-L. (1998). Perceptual learning reflects external noise filtering and internal noise reduction through channel reweighting. Proceedings of the National Academy of Sciences, 95(23), 13988–13993. https://doi.org/10.1073/pnas.95.23.13988
  • Fiorentini, A., & Berardi, N. (1980). Perceptual learning specific for orientation and spatial frequency. Nature, 287, 43–44. https://doi.org/10.1038/287043a0
  • Gold, J., Bennett, P. J., & Sekuler, A. B. (1999). Signal but not noise changes with perceptual learning. Nature, 402, 176–178. https://doi.org/10.1038/46027
  • Green, C. S., & Bavelier, D. (2003). Action video game modifies visual selective attention. Nature, 423, 534–537. https://doi.org/10.1038/nature01647
  • Jeter, P. E., Dosher, B. A., Petrov, A., & Lu, Z.-L. (2009). Task precision at transfer determines specificity of perceptual learning. Journal of Vision, 9(3), Article 1, 1–13. https://doi.org/10.1167/9.3.1
  • Karmarkar, U. R., & Dan, Y. (2006). Experience-dependent plasticity in adult visual cortex. Neuron, 52(4), 577–585. https://doi.org/10.1016/j.neuron.2006.11.001
  • Karni, A., & Sagi, D. (1991). Where practice makes perfect in texture discrimination: Evidence for primary visual cortex plasticity. Proceedings of the National Academy of Sciences, 88(11), 4966–4970. https://doi.org/10.1073/pnas.88.11.4966
  • Karni, A., & Sagi, D. (1993). The time course of learning a visual skill. Nature, 365, 250–252. https://doi.org/10.1038/365250a0
  • Law, C.-T., & Gold, J. I. (2008). Neural correlates of perceptual learning in a sensory-motor, but not a sensory, cortical area. Nature Neuroscience, 11, 505–513. https://doi.org/10.1038/nn2070
  • Levi, D. M., & Polat, U. (1996). Neural plasticity in adults with amblyopia. Proceedings of the National Academy of Sciences, 93(13), 6830–6834. https://doi.org/10.1073/pnas.93.13.6830
  • Ming, X., Huang, G., Chen, X., Liao, M., & Liu, L. (2025). A systematic review and meta-analysis of perceptual learning and video game training for adults with monocular amblyopia. Ophthalmology and Therapy, 14, 857–881. https://doi.org/10.1007/s40123-025-01128-9
  • Polat, U., Ma-Naim, T., Belkin, M., & Sagi, D. (2004). Improving vision in adult amblyopia by perceptual learning. Proceedings of the National Academy of Sciences, 101(17), 6692–6697. https://doi.org/10.1073/pnas.0401200101
  • Sasaki, Y., Náñez, J. E., & Watanabe, T. (2010). Advances in visual perceptual learning and plasticity. Nature Reviews Neuroscience, 11, 53–60. https://doi.org/10.1038/nrn2737
  • Schoups, A., Vogels, R., Qian, N., & Orban, G. (2001). Practising orientation identification improves orientation coding in V1 neurons. Nature, 412, 549–553. https://doi.org/10.1038/35087601
  • Seitz, A. R., & Watanabe, T. (2003). Is subliminal learning really passive? Nature, 422, 36. https://doi.org/10.1038/422036a
  • Seitz, A. R., Náñez, J. E., Holloway, S. R., Koyama, S., & Watanabe, T. (2005). Seeing what is not there shows the costs of perceptual learning. Proceedings of the National Academy of Sciences, 102(25), 9080–9085. https://doi.org/10.1073/pnas.0501026102
  • Seitz, A. R., Kim, D., & Watanabe, T. (2009). Rewards evoke learning of unconsciously processed visual stimuli in adult humans. Neuron, 61(5), 700–707. https://doi.org/10.1016/j.neuron.2009.01.016
  • Seitz, A. R., & Watanabe, T. (2009). The phenomenon of task-irrelevant perceptual learning. Vision Research, 49(21), 2604–2610. https://doi.org/10.1016/j.visres.2009.08.003
  • Shibata, K., Watanabe, T., Sasaki, Y., & Kawato, M. (2011). Perceptual learning incepted by decoded fMRI neurofeedback without stimulus presentation. Science, 334(6061), 1413–1415. https://doi.org/10.1126/science.1212003
  • Watanabe, T., Náñez, J. E., & Sasaki, Y. (2001). Perceptual learning without perception. Nature, 413, 844–848. https://doi.org/10.1038/35101601
  • Wright, B. A., & Sabin, A. T. (2007). Perceptual learning: How much daily training is enough? Experimental Brain Research, 180(4), 727–736. https://doi.org/10.1007/s00221-007-0898-z
  • Xiao, L.-Q., Zhang, J.-Y., Wang, R., Klein, S. A., Levi, D. M., & Yu, C. (2008). Complete transfer of perceptual learning across retinal locations enabled by double training. Current Biology, 18(24), 1922–1926. https://doi.org/10.1016/j.cub.2008.10.030
  • Yotsumoto, Y., Watanabe, T., & Sasaki, Y. (2008). Different dynamics of performance and brain activation in the time course of perceptual learning. Neuron, 57(6), 827–833. https://doi.org/10.1016/j.neuron.2008.02.034
  • Zhang, J.-Y., Zhang, G.-L., Xiao, L.-Q., Klein, S. A., Levi, D. M., & Yu, C. (2010). Rule-based learning explains visual perceptual learning and its specificity and transfer. Journal of Neuroscience, 30(37), 12323–12328. https://doi.org/10.1523/JNEUROSCI.0704-10.2010