Transfer of Learning

Abstract

Transfer of learning is the influence of prior learning on the acquisition or performance of a task in a new situation. It is classified along two axes: by direction, as positive when earlier learning helps and negative when it interferes, and by distance, as near when the training and transfer contexts share many features and far when they share few. The founding identical-elements theory held that transfer occurs only to the extent that two tasks share concrete components, a deflationary claim that overturned the doctrine of formal discipline. Later accounts added abstracted schemas, analogical mapping, and metacognitive control as vehicles that carry learning across dissimilar contexts. Whether substantial far transfer occurs at all remains the field's sharpest dispute, with recent meta-analyses of brain training reporting that trained gains rarely generalize.

Keywords: transfer of learning, near transfer, far transfer, analogical transfer, identical elements

Transfer of learning is what makes education worth the cost: the expectation that what is learned in one setting will be available in another. A student who studies proportions should reason about map scales; a pilot trained in a simulator should fly the aircraft. The construct names the carry-over itself, and a century of research has been devoted to measuring when it occurs, how far it reaches, and by what mechanism. It is closely tied to problem solving and analogical reasoning, because applying old learning to a novel problem is itself an act of transfer (Thorndike & Woodworth, 1901).

What Transfer Is

Transfer of learning is operationalized by a two-group comparison. A control group learns a target task with no relevant prior training; an experimental group learns the target task after some earlier training. The difference in how the experimental group acquires or performs the target task, relative to control, is the transfer produced by that training. When the experimental group does better, transfer is positive; when the earlier learning interferes and the group does worse, transfer is negative (Thorndike & Woodworth, 1901). A third, common outcome is zero transfer, in which prior training leaves target performance unchanged.

The second organizing axis is distance. Near transfer applies learning to a context that shares many surface and structural features with the original — a bookkeeper moving to a new ledger program. Far transfer applies it to a context that shares few — a chess player's training supposedly sharpening general reasoning. The taxonomy of far transfer specifies the dimensions along which a transfer context can differ from the training context, including knowledge domain, physical setting, social context, modality, and temporal distance, so that claims of far transfer can be graded rather than asserted (Barnett & Ceci, 2002). The greater the distance on these dimensions, the rarer and smaller transfer tends to be.

Figure 1

The Two Axes of Transfer: Direction and Distance

Theories of Transfer

The oldest and most durable account is the theory of identical elements. Against the nineteenth-century doctrine of formal discipline — the belief that studying Latin or geometry strengthened the mind as a whole — transfer was shown to occur only to the extent that two activities share specific stimulus-response components. Training on estimating the area of rectangles improved area estimates for shapes it resembled and barely touched estimates for shapes it did not (Thorndike & Woodworth, 1901). On this view there is no general faculty to be exercised; transfer is the reuse of concrete elements, and its narrowness is the default expectation.

A second tradition locates transfer in what is abstracted from training rather than in shared surface elements. A learning set — learning to learn — forms when repeated problems of the same type let a solver extract the common principle and apply it immediately to new instances, so that later problems are solved in a single trial (Harlow, 1949). The distinction between low-road and high-road transfer generalizes this: low-road transfer is the near-automatic triggering of a well-practiced routine in a similar context, while high-road transfer is the deliberate, mindful abstraction of a principle and its intentional application to a distant one (Salomon & Perkins, 1989). High-road transfer is the route by which far transfer, when it happens, is thought to occur.

A third strand recasts transfer as production-rule reuse. In this account the units of skill are condition-action rules, and two tasks transfer to the degree that they share the same underlying productions, a modern and computationally explicit descendant of identical elements that predicts the tight, componential transfer observed in skill acquisition (Anderson et al., 1996). Which task features count as identical is itself theory-laden — encoding conditions shape what is retrievable, so that a match between the processing at study and the processing at test can matter as much as a match between the nominal tasks (Morris et al., 1977).

Analogical Transfer

The most studied vehicle for far transfer is analogy. In the canonical demonstration, solvers who had read a story about an army capturing a fortress by converging on it from many roads rarely applied the same converging-forces idea to a structurally identical medical problem — using many weak rays to destroy a tumor — unless they were told the story was relevant (Gick & Holyoak, 1980). The failure is diagnostic: the two problems shared deep structure but no surface content, and spontaneous transfer is governed by surface similarity even though successful transfer requires mapping the structure.

What lets structure win out is schema induction. Comparing two disparate analogues that share a solution principle leads solvers to abstract a general schema — the convergence schema — and it is that schema, not either story alone, that then transfers to a new problem (Gick & Holyoak, 1983). This is why a single worked example transfers poorly while two contrasting examples transfer well: the comparison is what strips away surface detail and leaves the portable structure. The finding reframes far transfer as a problem of representation — of getting the learner to encode the deep structure in the first place — rather than a problem of raw similarity between tasks. Gentner's structure-mapping theory gives this its formal footing: an analogy is a mapping of the system of relations from a source to a target that honors a systematicity principle, preferring higher-order relations that form an interconnected structure over isolated shared attributes, which is precisely why matched relational structure drives transfer while matched surface features do not (Gentner, 1983).

Measuring Transfer

Because transfer is a comparison between groups, it is quantified as a percentage relative to a control baseline. In a trials-to-criterion design, where fewer trials means faster learning, percent transfer is the control group's trials minus the experimental group's trials, divided by the control group's trials, times one hundred. A positive value means the prior training saved learning effort; a negative value means it cost effort. The measure is bounded above by one hundred percent — a case in which prior training eliminates the need for any target learning — and is unbounded below, since sufficiently disruptive prior learning can more than double the trials required.

The design's logic makes transfer a difference score, and difference scores inherit the reliability problems of their components, which is one reason transfer effects are often small and noisy. The measurement problem compounds the theoretical one: a study that reports far transfer must show that the gain is not simply near transfer to a criterion task that resembles the training more than the label suggests (Barnett & Ceci, 2002). Preparation for future learning is an alternative metric that credits training not by immediate transfer performance but by how readily the learner acquires the new material when later given resources to do so, capturing gains that a static transfer test misses (Bransford & Schwartz, 1999).

Table 1
Principal Types of Transfer and the Strength of the Evidence for Each
Note. The direction and distance categories are orthogonal; a given episode is classified on both axes at once (Barnett & Ceci, 2002; Salomon & Perkins, 1989).

Worked Example

Consider a proactive-transfer study in which a control group learns a target task from scratch while an experimental group learns it after related training. Learning is scored as trials to a fixed criterion, so lower is better. The control group reaches criterion in C = 40 trials.

Suppose the experimental group, having trained on a related task, reaches criterion in E = 28 trials. Percent transfer is 100 × (C − E) / C = 100 × (40 − 28) / 40 = 100 × 12 / 40 = +30%: the prior training saved roughly a third of the learning effort, a clear positive transfer. Now suppose instead that the prior training conflicted with the target task and the experimental group needed E = 48 trials. Then percent transfer is 100 × (40 − 48) / 40 = 100 × (−8) / 40 = −20%: negative transfer, because the earlier learning had to be overcome before the new task could be mastered. The same baseline of 40 trials thus yields a positive or a negative percentage depending only on whether prior learning helped or interfered, which is why the control group is indispensable — without it there is no way to know whether 28 or 48 trials is fast or slow. The interactive demonstration above reproduces these values as the control and experimental trial counts are adjusted.

Discussion

Transfer is the point at which learning theory has to make good on its promise. Every account of learning implicitly predicts a scope of transfer, and the century-long finding is that the scope is narrower than intuition expects. The identical-elements theory set the deflationary baseline, and its computational descendants in skill acquisition have largely confirmed it: transfer tracks shared components closely, and the felt generality of a well-learned skill is mostly an illusion produced by many overlapping near transfers (Anderson et al., 1996). The abstraction accounts do not overturn this so much as identify the narrow conditions — comparison of multiple cases, explicit principle extraction, mindful application — under which the reach of transfer can be extended (Salomon & Perkins, 1989; Gick & Holyoak, 1983).

The practical stakes are large because education is a bet on transfer. If far transfer were easy, a well-chosen curriculum of general skills would pay off everywhere; if it is hard, instruction must teach closer to the contexts of eventual use and must engineer the comparisons that expose deep structure. The evidence favors the second reading, which is why the analogical-transfer literature — with its insistence that structure must be made salient before it will carry — has had more influence on instructional design than any appeal to general mental training.

Current Directions

The liveliest current question is whether far transfer occurs at all in the one domain that most needs it: cognitive training. Meta-analyses of working-memory training find that trained tasks and very similar ones improve while the promised transfer to reasoning, fluid intelligence, and scholastic attainment does not reliably appear (Melby-Lervåg & Hulme, 2013; Melby-Lervåg et al., 2016). A comprehensive review of commercial brain-training programs reached the same verdict, finding evidence for improvement on trained tasks but little for the everyday-cognition benefits the programs advertise (Simons et al., 2016).

Extending the analysis across chess, music, and working-memory training, a broader synthesis argues that once study quality and publication bias are accounted for, the far-transfer effect approaches zero — that domain-general cognitive enhancement through training may simply not exist (Sala & Gobet, 2017; Sala & Gobet, 2019). Defenders reply that the training regimes tested were not designed to teach the abstractions that high-road transfer requires, and that preparation-for-future-learning measures would credit gains the standard transfer test cannot see (Bransford & Schwartz, 1999). The dispute is not settled, but its center of gravity has shifted from whether far transfer can be produced easily to whether it can be produced at all.

Common Misconceptions

Learning a hard subject strengthens the mind in general.

This is the doctrine of formal discipline, and it was the first casualty of transfer research. Training improves performance mainly on tasks that share concrete components with it, not on unrelated tasks, so studying Latin does not sharpen reasoning as such (Thorndike & Woodworth, 1901). The belief persists because a broadly educated person seems generally capable, an impression produced by many specific competencies rather than one strengthened faculty.

Seeing a worked example is enough to transfer its method.

A single example is bound to its surface features, and solvers usually fail to apply it to a structurally identical problem that looks different (Gick & Holyoak, 1980). Comparing two contrasting examples that share a principle is what induces the transferable schema, so transfer is a property of the comparison, not of exposure to one case (Gick & Holyoak, 1983).

Brain-training games make people generally smarter.

The trained tasks reliably improve, but the gains rarely reach the untrained abilities the products promise. Meta-analyses of working-memory and commercial brain training find robust near transfer and little or no far transfer to intelligence or everyday cognition (Melby-Lervåg et al., 2016; Simons et al., 2016). Improvement on the game is mistaken for improvement of the underlying faculty.

Glossary

Analogical transfer.

Transfer achieved by mapping the relational structure of a known source problem onto a novel target problem that shares that structure but not its surface content.

Far transfer.

Transfer to a context that shares few features with the training context, spanning differences of domain, setting, or modality; rare and much disputed.

Formal discipline.

The discredited doctrine that studying rigorous subjects strengthens general mental faculties transferable to any content.

High-road transfer.

Transfer achieved by deliberately abstracting a principle from one context and mindfully applying it to a distant one.

Identical elements.

Thorndike and Woodworth's theory that transfer occurs only to the degree that two tasks share specific stimulus-response components.

Learning set.

A learned readiness to solve a whole class of problems, formed across repeated instances so that new members are solved almost at once; learning to learn.

Low-road transfer.

The near-automatic triggering of a well-practiced routine in a context that closely resembles the one in which it was learned.

Near transfer.

Transfer to a context that shares many surface and structural features with the training context; common and comparatively easy to obtain.

Negative transfer.

The case in which prior learning interferes with a new task, so that the trained learner performs worse than an untrained control.

Percent transfer.

A quantitative measure of transfer expressing the experimental group's saving or loss relative to a control baseline as a percentage.

Positive transfer.

The case in which prior learning aids a new task, so that the trained learner acquires or performs it better than an untrained control.

Preparation for future learning.

An alternative transfer metric that credits training by how readily the learner acquires new material later, rather than by immediate performance on a transfer test.

Schema induction.

The abstraction of a general, transferable structure from the comparison of two or more analogous cases that share a solution principle.

Structural similarity.

Correspondence between two problems in their underlying relations rather than their concrete content; the basis on which successful analogical transfer depends.

Surface similarity.

Resemblance between two problems in their concrete objects and features; what drives spontaneous retrieval even though it is a poor guide to transferable structure.

Transfer-appropriate processing.

The principle that a training operation benefits a later task to the degree that the cognitive processing it engages matches the processing the later task requires.

Key Researchers

John R. Anderson (b. 1947). R. K. Mellon University Professor at Carnegie Mellon University; his production-rule account of skill recast identical elements in computational form and predicted componential transfer. Faculty Page

Stephen J. Ceci (b. 1951). Helen L. Carr Professor of Developmental Psychology at Cornell University; with Susan Barnett he built the taxonomy that grades claims of far transfer along explicit dimensions. Faculty Page - ORCID

Fernand Gobet (b. 1962). Professorial Research Fellow at the London School of Economics; with Giovanni Sala he synthesized the evidence across chess, music, and cognitive training arguing that far transfer approaches zero. Wikipedia - ORCID

Harry F. Harlow (1905-1981). Psychologist at the University of Wisconsin-Madison; his work on learning sets showed that transfer accrues across repeated problems as a solver learns how to learn. Wikipedia

Keith J. Holyoak (b. 1950). Distinguished Professor of Psychology at UCLA; with Mary Gick he established analogical transfer and schema induction as the mechanism of structure-based generalization. Faculty Page - ORCID

Charles Hulme (b. 1953). Professor of Psychology and Education at the University of Oxford; with Monica Melby-Lervåg he produced the meta-analyses showing working-memory training yields near but not far transfer. Faculty Page - ORCID

David N. Perkins (b. 1942). Senior Professor of Education, Emeritus, at the Harvard Graduate School of Education; with Gavriel Salomon he drew the low-road/high-road distinction in transfer mechanisms. Faculty Page

Edward L. Thorndike (1874-1949). Educational psychologist at Columbia University's Teachers College; with Robert Woodworth he founded the experimental study of transfer and its identical-elements theory, demolishing the doctrine of formal discipline. Wikipedia

Frequently Asked Questions

What is transfer of learning?
It is the influence of prior learning on the acquisition or performance of a task in a new situation; when the earlier learning helps it is positive transfer, and when it interferes it is negative transfer (Thorndike & Woodworth, 1901).

What is the difference between near and far transfer?
Near transfer applies learning to a context that shares many features with the original, while far transfer applies it to a context that shares few; the taxonomy of far transfer grades this distance along dimensions such as domain, setting, and modality (Barnett & Ceci, 2002).

What is the identical-elements theory?
It is Thorndike and Woodworth's proposal that transfer occurs only to the extent that two tasks share specific stimulus-response components, which overturned the belief that studying difficult subjects strengthens the mind in general (Thorndike & Woodworth, 1901).

Why do people fail to transfer a solution to a similar problem?
Spontaneous transfer is driven by surface similarity, so solvers often miss a structurally identical problem that looks different unless the connection is pointed out or a second analogue exposes the shared structure (Gick & Holyoak, 1980).

What is the difference between low-road and high-road transfer?
Low-road transfer is the automatic triggering of a well-practiced routine in a similar context, while high-road transfer is the deliberate abstraction of a principle and its mindful application to a distant context (Salomon & Perkins, 1989).

How is transfer measured?
It is measured as a percentage of a control baseline: percent transfer is the control group's trials to criterion minus the experimental group's, divided by the control group's, times one hundred, so positive values mean prior training saved effort.

Do brain-training programs improve general intelligence?
Meta-analyses find that trained and closely related tasks improve while transfer to reasoning, intelligence, and everyday cognition does not reliably occur, so the advertised far-transfer benefits are largely unsupported (Melby-Lervåg et al., 2016; Simons et al., 2016).

Does far transfer exist at all?
It is the field's sharpest dispute; syntheses across chess, music, and cognitive training argue that after correcting for study quality and publication bias the far-transfer effect approaches zero, though critics say the training was never designed to teach transferable abstractions (Sala & Gobet, 2019).

References