Learning

Prime #

None

Origin domain

Psychology & Behavioral Sciences

Also from

Artificial Intelligence, Biology & Ecology, Organizational & Management Science

Aliases

Acquisition, Skill Acquisition

Core Idea

Learning is the process by which an agent acquires or modifies an internal capability — knowledge, skill, model, or behavior — as a durable result of experience or information, such that the agent's future performance or predictions change, an account that Skinner (1953) develops at length for behavioral capability change and that Bandura (1977) extends to observational acquisition. ^[1][2] It is the learner-side counterpart to pedagogy: where pedagogy is the deliberate teaching aimed at causing such change, learning is the agent's own update, a learner-side framing that runs through Piaget's (1952) developmental account of how the child constructs knowledge from interaction with the environment. ^[3] The essential commitment is a durable, experience-driven self-update of an agent's internal state that carries forward to alter later behavior, distinguishing it from a one-off response that leaves the agent unchanged — a durability requirement Thorndike (1898) established in his law-of-effect work and that Ebbinghaus (1885) made quantitatively precise through his retention curves. ^[4][5] The prime names a conjunction that no neighboring concept carries together: an agent with modifiable internal state, an experiential cause of change, and durability such that the change persists and shows up later in altered behavior or prediction.

The pattern recurs across human and animal cognition (conditioning, skill acquisition, observational learning), machine learning (parameter updates from data), adaptive immunity (B-cell affinity maturation and immune memory), and organizational routines (postmortem-driven revision). The substrate is irrelevant; what matters is that the agent's state is modifiable, the cause is experience or information rather than design or random drift, and the change persists.

How would you explain it like I'm…

Getting Smarter from Experience

Learning is when you change inside because of something you experienced, and the change sticks around. Like the first time you touch a hot stove — you remember, and you don't touch it again. Whatever happened to you taught you something that lasts.

Learning from Experience

Learning is when an agent — a person, an animal, a computer program, even an immune system — changes itself based on experience, and the change sticks so it affects what the agent does later. It's not just reacting once; it's updating yourself so the next time around, your behavior or your guesses are different. The thing being changed could be a skill, a memory, an idea, or a habit, but the key is: experience went in, the inside changed, and the change carries forward.

Learning

Learning is the process by which an agent acquires or modifies an internal capability — knowledge, skill, model, or behavior — as a durable result of experience or information, so that future performance or predictions change. It's the learner's side of teaching: where pedagogy is what a teacher does to cause change, learning is the agent's own update. The essential combination is an agent with a changeable internal state, experience as the cause of change, and durability so the change persists. This shows up in conditioning, skill practice, machine-learning models updating weights from data, immune systems remembering pathogens, and organizations revising routines after a postmortem. The substrate doesn't matter — what matters is modifiable state, experiential cause, and lasting effect.

 

Learning is the process by which an agent acquires or modifies an internal capability — knowledge, skill, model, or behavior — as a durable result of experience or information, such that the agent's future performance or predictions change. It is the learner-side counterpart to pedagogy: where pedagogy is the deliberate teaching aimed at causing such change, learning is the agent's own update. The essential commitment is a durable, experience-driven self-update of an agent's internal state that carries forward to alter later behavior, distinguishing it from a one-off response that leaves the agent unchanged. The prime names a conjunction that no neighboring concept carries together: an agent with modifiable internal state, an experiential cause of change (not design, not random drift), and durability so the change persists and shows up later. The pattern recurs across human and animal cognition (classical and operant conditioning, skill acquisition, observational learning), machine learning (gradient descent updating parameters from data), adaptive immunity (B-cell affinity maturation producing immune memory), and organizational routines (postmortem-driven revision of procedures). The substrate — neurons, weights, antibody repertoires, written policies — is irrelevant; what matters is that the state is modifiable, the cause is experience or information, and the change persists.

Structural Signature

Learning encodes a structural pattern: experiential input → agent with modifiable internal state → update mechanism → durable capability change — a four-role decomposition that Sutton and Barto (2018) formalize in the reinforcement-learning setting (state, signal, update rule, retained policy) and that Bishop (2006) develops in the supervised-learning setting (parameters, training data, loss-driven update, learned model). ^[6][7] It separates two states (the pre-experience agent and the post-experience agent) and names the transformation that the experience causes in the internal state, carried forward into altered future performance.

Recurring features:

• Experience-driven durable update of an agent's internal state
• Modifiable internal capability that persists past the triggering input
• Asymmetric coupling from experience to state to future behavior
• Acquisition that distinguishes a learner from a one-off responder
• Capability change retained by the agent that did the learning
• Update rule operating on state in light of input and performance signal

The structural insight is that the same four-role pattern recurs whether the modifiable state is synaptic weight, neural-network parameters, an immune-cell repertoire, a schema, or an organizational routine. A child learning to ride a bike, a neural network training on labeled examples, an immune system selecting antibody affinity to a new antigen, and an engineering team revising a runbook after an outage all instantiate the same skeleton: experience reaches a state that can change, a mechanism updates the state in light of a performance or error signal, and the change persists to alter future performance.

What It Is Not

Learning is not experience itself. Experience is the input — the stream of observations, feedback, practice trials, or instruction that supplies the cause. A system can be flooded with experience and learn nothing if it has no modifiable state to receive the change, or no update mechanism to convert input into capability — the point Hebb (1949) makes in arguing that durable behavioral change requires a physiological substrate ("neurons that fire together wire together") capable of being modified by coincident activity. ^[8] A camera records the visual scene without learning anything from it; a person who walks the same route every day without attending may retain little. Experience is necessary but not sufficient.

Learning is not memory. Memory is the retention and storage of representations; learning is the acquisition or update process that writes to memory and modifies what is stored. A perfect memory system that never updates its contents in light of new experience is not learning; an agent that learns must write to some store, but the writing — the update — is the learning, not the storage. The two are tightly coupled but conceptually distinct: memory is where the change lives, learning is the change itself.

Learning is not feedback. Feedback is the loop that routes output back to input as a signal. Learning may use feedback signals (especially error signals) as the experiential input that drives the update, but the loop itself is not the learning. A thermostat operates a feedback loop without learning anything; nothing in the thermostat changes durably from the trips it makes. Learning requires that the loop write a durable trace into modifiable state, which feedback alone does not guarantee.

Learning is not every change a system undergoes. A system can change for many reasons unrelated to experience: random drift, structural redesign by an external engineer, parameter resetting, decay. None of these are learning, because the cause is not experiential and the change is not the result of an update mechanism operating on the agent's own state in light of input. Calling all change "learning" empties the term.

Learning is not instantaneous reaction. A reflex, a one-off response, or a transient behavioral shift that leaves the agent unchanged the moment the trigger ends is not learning. The durability requirement is what separates learning from mere responding — the very distinction Pavlov (1927) drew between the unconditioned reflex (innate, transient with the trigger) and the conditioned reflex (acquired, persisting after pairing). ^[9] A startled jump is a reaction; conditioning that makes the same stimulus reliably evoke caution weeks later is learning.

Learning is not the goodness of what was learned. A system can learn maladaptive responses, superstitions, biased decision rules, or actively harmful associations. Learned helplessness, conditioned fear, and biased training data all count structurally as learning even though their content is harmful. The prime describes the mechanism of durable experience-driven self-update; whether the resulting capability is beneficial is a separate evaluative question that the structure does not settle.

Broad Use

Psychology and behavior: classical and operant conditioning, skill acquisition, practice effects, observational learning, habit formation, and the broader study of how organisms modify their behavior in light of consequences and exposure. Conditioning across species — pigeons, rats, primates, humans — shows the same experience-driven update pattern in agents with no language and no instruction, a cross-species regularity established by Thorndike's (1898) puzzle-box experiments with cats, Pavlov's (1927) salivary-conditioning work with dogs, and Skinner's (1953) operant programs across pigeons and rats. ^[4][9][1]

Machine learning and artificial intelligence: supervised, unsupervised, and reinforcement learning systems update model parameters from data so that future predictions improve on the target task. The structural skeleton — modifiable parameters, experiential input (training samples or environmental rewards), update rule (gradient descent or its analogues), durability across the training run — matches the biological skeleton exactly, despite the substrate being silicon rather than wetware.

Biology and immunology: adaptive immunity is the canonical substrate-furthest case. The immune memory cell repertoire is the modifiable internal state; pathogen exposure is the experiential input; clonal selection and affinity maturation are the update mechanism; the lifelong altered response is the durability. No cognition, no language, no consciousness — and yet every role in the structural pattern is cleanly present, as Burnet (1959) established in formulating the clonal selection theory that underwrites adaptive immune memory. ^[10]

Organizational and collective learning: routines, playbooks, runbooks, institutional knowledge, and incident response procedures get updated in light of accumulated experience. A team's deployment runbook revised after a postmortem; a hospital's checklist updated after a near-miss; an open-source project's contribution guidelines refined as new patterns of contribution emerge — all instantiate the same structure at a collective scale.

Animal behavior and ethology: classical conditioning in dogs, operant conditioning in rats, navigation learning in bees, song acquisition in birds, tool-use transmission in primates. These cases foreground the agent-internal aspect: no instruction is required, no language, no symbolic representation — only an animal with modifiable state, experiential exposure, and a mechanism that converts the exposure into durable capability change.

Cognitive science and education research: schema acquisition, mental model development, expertise development, transfer of learning, spaced repetition, retrieval practice, and the empirical study of which experiential regimes produce the most durable and most transferable capability changes. Piagetian accommodation — the schema-revision mechanism by which an existing schema is restructured to fit a violating input — sits inside this family as one specific update mechanism rather than as the whole of learning, as Piaget (1952) develops in his account of the assimilation/accommodation dialectic in early cognitive development. ^[3]

Clarity

Learning names a specific, narrow thing inside the broader cluster of change: a durable, experience-driven update to an agent's internal capability that carries forward into altered future behavior or prediction — the conjunction Bandura (1977) emphasizes in distinguishing learning (acquired capability) from performance (its expression on a given trial). ^[2] It is not the input (experience), not the storage (memory), not the loop that supplies error signals (feedback), not the teacher's effort to cause it (pedagogy), and not every kind of system-level change under pressure (adaptation). What learning adds is the conjunction of three commitments that no neighboring prime carries together: an agent with modifiable internal state, an experiential cause of the change, and durability such that the change persists and shows up later.

Naming that conjunction prevents the common slippage where "the system changed" gets reported as learning when in fact the change was random drift, a one-off reaction, or a structural redesign with no role for experience. A change-management consultant who labels every organizational shift "organizational learning" loses the diagnostic value of the term; the structure asks specifically whether the change was driven by experience reaching a modifiable internal state through some update mechanism, with durability checked against later performance. If those roles are not present, the change may still be real and important, but it is not learning.

The clarity also distinguishes between performance change and capability change. A system may perform better today because conditions are favorable or because it is being driven harder; it has not necessarily learned. Learning is the underlying capability shift that would produce the better performance even under independent conditions. Tests that distinguish learning from performance — generalization tests, transfer tests, delayed retention tests — are clarifying precisely because they separate the durable internal change from the moment-to-moment behavioral output.

Manages Complexity

Learning decomposes an opaque "the agent got better" into four named roles whose presence can be checked one at a time: (1) an agent with a modifiable internal state — memory, weights, schema, routine, or whatever holds the change; (2) experiential input — observations, feedback, practice, instruction, or exposure that supplies the cause; (3) an update mechanism that modifies the internal state in light of that input together with a performance or error signal; and (4) durability — the change persists and manifests as altered future behavior or prediction — a decomposition Sutton and Barto (2018) make explicit as the canonical agent/environment/update/value formulation underlying reinforcement learning. ^[6]

Once those four roles are named, an analyst can convert a vague "we learned this" into a structured problem: which state is modifiable, what experiential signal is reaching it, what update rule is operating, and does the change actually persist past the moment? This makes the same diagnostic apply equally to a child learning to ride a bike, a neural network training on examples, and an organization adjusting its routines after a postmortem. When a learning system fails, the decomposition tells you where to look: is the state insufficiently modifiable (too rigid, too narrow, capacity-limited)? Is the experiential input absent, sparse, or scrambled? Is the update mechanism mis-specified? Or is the change occurring but failing to persist (catastrophic forgetting, interference, decay)?

The decomposition also makes diagnostic questions sharp for organizational learning, where the failure modes are notoriously elusive. "The team isn't learning from incidents" becomes: are postmortems happening (input)? Is there a routine or playbook that can be revised (state)? Is there a process that converts incident lessons into routine revisions (mechanism)? Are the revisions actually retained and used (durability)? Each of those four questions has a different remediation, and lumping them under "we need to be a learning organization" obscures which is broken.

Abstract Reasoning

Learning supports a sharp counterfactual: if the experiential input were withheld, removed, or scrambled, the change would not happen — and if the agent had no modifiable state, the input would pass through without leaving a trace — a counterfactual logic Bjork and Bjork (1992) leverage in their "new theory of disuse," where retrieval strength and storage strength are independently manipulable so that the durable-capability signature can be cleanly separated from transient performance. ^[11] That move lets the analyst predict where learning will and will not occur in any new domain: it requires an asymmetric, directed coupling from experience to internal state to future behavior, and the absence of any of those three breaks the pattern.

The reasoning also enables clean cross-cutting comparisons across substrates — supervised versus unsupervised updates, single-trial versus multi-trial regimes, sample-efficiency, transfer between contexts, forgetting and decay, the gap between performance gain and underlying capability change. These operations are substrate-neutral because they ride on the structural topology (agent → experience → update → durable capability) rather than on the substrate's specific machinery. A question like "how sample-efficient is this learner?" applies equally to a child acquiring a second language, a neural network on a new benchmark, and an immune system encountering a novel pathogen.

The counterfactual also helps disconfirm spurious learning claims. If a behavior would have arisen identically without the supposed experiential input, the input did not cause learning; the behavior emerged from some other source (innate program, structural change, environmental scaffolding). If the supposed learning vanishes the moment the input ends and leaves no trace, durability fails and the change was reactive rather than learned. The counterfactual gives the analyst a clean test for ruling out look-alikes.

Knowledge Transfer

The four-role structure travels intact across substrates with no metaphorical stretching. A machine-learning researcher reading about adaptive immunity recognizes a parameter-update loop driven by exposure samples; an immunologist reading about supervised learning recognizes antigen-driven memory cell selection; an organizational-learning theorist recognizes the same structure in postmortem-driven routine revision; a psychologist recognizes it in operant conditioning — a cross-substrate recognition that LeCun, Bengio, and Hinton (2015) make explicit when they describe deep learning's loss-driven parameter update as the same broad pattern of "experience changing internal representation" that runs through biological learning. ^[12]

The biological case — adaptive immunity — is especially clean because there is no language, no instruction, no consciousness, and yet the four roles are all clearly present: modifiable internal state (the immune memory cell repertoire), experiential input (pathogen exposure), an update mechanism (clonal selection and affinity maturation), and durability (life-long altered response). That single case rules out the suspicion that "learning" is a specialty of cognitive psychology and forces the prime to live at the structural level where the same pattern crosses cleanly into machine learning and organizational routines, because Burnet's (1959) clonal selection account already supplied a substrate-neutral selection-and-retention mechanism that is structurally indistinguishable from the loss-driven parameter-update loop LeCun, Bengio, and Hinton (2015) describe for deep networks. ^[10][12]

Transfer of the frame itself — taking the diagnostic questions ("which state is modifiable, what is the experiential signal, what is the update rule, is the change durable?") into a new domain — is correspondingly clean. A roboticist debugging why an agent fails to improve in simulation can ask the same four questions a teacher asks about a struggling student or an immunologist asks about a non-responder. The mechanisms differ wildly across substrates; the diagnostic skeleton does not.

Examples

Formal/abstract

Adaptive immunity (B-cell affinity maturation): A naive B cell carries a particular B-cell receptor that binds antigens with some baseline affinity. Upon encountering a pathogen, B cells whose receptors bind the antigen are activated; in the germinal center they undergo somatic hypermutation, with daughter cells carrying small variants of the receptor. Those whose variants bind the antigen with higher affinity receive stronger survival signals; the population shifts over rounds of mutation and selection toward higher-affinity receptors. After clearance, a fraction persist as memory cells with the matured receptor. The next exposure to the same antigen meets a population already enriched for high-affinity binders, producing a faster and stronger response. Mapped back: Walk the four roles. The modifiable internal state is the B-cell receptor repertoire (variable regions of the immunoglobulin genes). The experiential input is pathogen exposure providing antigen and the co-stimulatory signals that mark a binding event as worth committing to. The update mechanism is somatic hypermutation plus antigen-driven selection, structurally a parameter-update loop driven by an error signal (affinity gap). The durability is the persistence of memory cells across years or decades. No cognition, no instruction, no consciousness — and every role of learning is cleanly present. The case is decisive against the suspicion that learning is a cognitive-psychology specialty.

Supervised learning of a neural network: A network with randomly initialized parameters is shown labeled examples. For each example, the network produces a prediction; a loss function compares prediction to label; gradient descent computes how each parameter should change to reduce the loss, and the parameters are updated by a small step in that direction. Over many examples, the parameters move into a configuration that produces accurate predictions on held-out data. Mapped back: The state is the parameter vector; the experiential input is the stream of (input, label) pairs; the update mechanism is gradient descent on the loss; the durability is the persistence of the trained parameters across the run and into deployment. The skeleton is identical to the immune case; the substrate is silicon and the update rule is differentiable optimization rather than mutation-and-selection. The structural identity is what justifies calling both "learning" without metaphor.

Applied/industry

Engineering team postmortem and runbook revision: A production service has an outage. The team conducts a blameless postmortem, identifies the contributing factors (a poorly handled timeout, an alerting gap, an under-tested code path), and updates the deployment runbook and on-call playbook to require additional checks, add new alerts, and route a particular class of incident differently. Six months later a similar fault pattern emerges and the new runbook prevents the second outage. Mapped back: The state is the runbook and playbook (and the team's tacit routines around them); the experiential input is the outage and the postmortem analysis; the update mechanism is the postmortem process that converts incident details into routine revisions; the durability is the persistence of the runbook changes across personnel turnover and the demonstrated success at preventing the same failure mode. This is organizational learning in the strict structural sense — a modifiable collective state, an experiential cause, an update mechanism, and durability checked against later performance — the same four-role pattern Bandura (1977) traces from individual observational learning up into collective behavioral codification. ^[2] If any of those four roles is missing — the postmortem happens but the runbook is not updated, or it is updated but the change does not stick — the team has not learned, even if individuals feel they understand the incident better.

Chess player adapting to a counter-system: A player repeatedly opens with a particular system, loses several games to a specific counter, studies those losses with engine analysis, and over the next month stops falling for the counter. Mapped back: The modifiable internal state is the player's repertoire of move-evaluations and pattern recognition; the experiential input is the sequence of losses plus the post-game review; the update mechanism is whatever cognitive process re-weights the response to the position the counter creates, combining the error signal (losing) with the corrective information (engine variations); the durability is the persisting altered behavior — the next time the position appears in a tournament, the player plays differently. This is learning, not adaptation in general: the change is experience-driven (no losses, no review, no change), agent-internal (the chess world did not change), and capability-altering (future performance is different in the same external conditions). The same skeleton runs through every domain where an agent updates its capability from experience.

Structural Tensions

T1: Performance change versus underlying capability change. Short-term improvements in behavior can come from many sources unrelated to durable capability shifts: favorable conditions, increased effort, motivational state, environmental scaffolding, or transient priming. Identifying learning requires distinguishing these from a real internal update. Standard tests — generalization to novel inputs, transfer to related tasks, retention after a delay, performance under independent conditions — were designed to make this separation, but in real deployments the separation is often blurred. An analyst who treats every performance gain as learning will misattribute improvements that are actually transient, and an analyst who demands strict capability tests in every domain will under-credit real learning that has not yet stabilized.

T2: Durability versus plasticity. Durability is a defining commitment of learning, but a system that retains every change is brittle and over-fit; a system that updates freely is responsive but loses what it knew. Every learning system faces the stability-plasticity dilemma: how to incorporate new experience without erasing prior capability (catastrophic forgetting) and how to retain prior capability without freezing out new learning. Neural networks trained sequentially on new tasks famously lose performance on old ones; humans face the same tension in skill maintenance and unlearning. The tension is fundamental to the prime — there is no general "more learning is better" because more updating can erode previously acquired capability.

T3: Experiential cause versus structural redesign. Learning specifically requires that the change arise from the agent's own update process in response to experience, not from an external redesigner reaching in and rewriting state. But the line is porous: a teacher restructuring a student's mental model through direct instruction is exerting more design pressure than a child figuring out a counterfactual on their own; a machine-learning engineer setting hyperparameters and curating data is partly designing the trained system, not merely letting it learn. Distinguishing learning from external authorship matters for explanation and credit, but the boundary is rarely clean — most real learning happens under some scaffolding, and scaffolded learning is still learning if the agent's own update mechanism does the work.

T4: Same skeleton, very different sample efficiency. The four-role pattern is substrate-neutral, but the amount of experience required to update state varies by orders of magnitude across substrates. A human can learn a face from one exposure; a typical supervised classifier needs thousands of labeled examples. A child acquires grammar from far less linguistic data than current language models require. The skeleton tells you the same roles are present; it does not tell you why some learners are dramatically more sample-efficient than others. Attributing the difference is a deep open problem (inductive biases, structured priors, active sampling, transfer from prior learning), and the structural prime does not resolve it.

T5: Learning what to learn. Most learning systems do not autonomously decide what to attend to or what counts as a reward signal; those are supplied by the environment, by an external designer, or by prior learning. When the wrong things get reinforced, the system reliably learns the wrong thing — superstitious conditioning, reward hacking in reinforcement learning, learned helplessness in animals, biased patterns learned from biased data. The prime tells us the mechanism worked exactly as it should; the failure is at the boundary of what experience reached the update process. Without explicit attention to that boundary, learning systems will competently acquire content their designers did not intend.

T6: Individual learning versus collective learning. The four-role structure scales from individual agents to teams, organizations, professions, and societies, but the mechanisms and timescales change in ways that the structural identity can obscure. Individual learning operates on a single nervous system with one update process; collective learning depends on transmission, codification, and institutional memory, any of which can fail independently of individual learning. A team in which every member individually learned from an incident may collectively learn nothing if there is no mechanism to encode the lesson into shared routines. Conversely, an organization may codify a lesson into a runbook that persists past every individual who experienced the originating event. The prime applies at both scales, but the failure modes are scale-specific.

Structural–Framed Character

Learning sits at the structural end of the structural–framed spectrum, with one small framed-side caveat: the prime presupposes an agent-with-internal-state, which is what holds the aggregate fractionally off zero. Strip that one assumption away and what remains is a substrate-neutral pattern — a durable, experience-driven update of an internal capability that carries forward to alter later behavior or prediction.

Domain vocabulary does not travel; psychology, machine learning, immunology, and organizational theory each describe the update in their own native terms, and the cross-substrate signature is the structure of the update rather than any shared lexicon. The prime carries no evaluative weight — learning is descriptive of a state-change pattern, not normatively loaded. Institutional origin reads zero: no school, curriculum, or community is required for the pattern to obtain. The half-step toward framed comes from the human-practice-bound criterion: every instance requires some agent with modifiable internal state, and most worked examples are biological or cognitive systems, though B-cell affinity maturation and ML parameter updates are squarely covered by the same structural commitments. Import-vs-recognize is recognition: noticing that an adaptive immune system durably updates from antigen exposure is finding learning already present in the substrate, not borrowing a cognitive framing. On the spectrum, the verdict is structural with a mild agent-binding tint.

Substrate Independence

Learning is about as substrate-independent as a prime can be — composite 5 / 5 on the substrate-independence scale. The pattern is one substrate-neutral commitment: an agent with modifiable internal state updates that state from experience or information so that future behavior or prediction durably changes. Domain breadth is at the ceiling, since the identical structural arrangement recurs across human and animal learning (conditioning, skill acquisition), machine learning (parameter updates from data), adaptive immunity (immune memory), and organizational learning. Transfer evidence is similarly heavy: the experience-driven, durable, capability-changing self-update has been deliberately ported between psychology, neuroscience, statistics, and computer science, and a single formal apparatus (loss, signal, update rule, retention) is now routinely shared between human-learning research and machine-learning research without translation friction. Structural abstraction sits one rung below maximum because the pattern presumes an agent with modifiable internal state — slightly more committal than a purely relational signature — but the agent itself can be biological, formal, organizational, or computational without changing the shape of the prime. The verdict is that learning is one of the catalog's canonical universal primes, recognized in any system that retains experience-driven changes to its own future capability.

• Composite substrate independence — 5 / 5
• Domain breadth — 5 / 5
• Structural abstraction — 4 / 5
• Transfer evidence — 5 / 5

Relationships to Other Primes

Parents (1) — more general patterns this builds on

• Learning is a kind of Adaptation

  Learning is a kind of adaptation specialized to an agent's internal cognitive or behavioral capability. Both share the adaptation pattern of modifying internal structure in response to sustained environmental input so that continued functioning under new conditions improves. Learning narrows the substrate to knowledge, skill, model, or behavior held by an agent, and narrows the trigger to experience or information. The broader adaptation prime spans physiological, developmental, and evolutionary modification; learning is the particular case where the modified system is an information-processing agent whose future predictions or performance change as the durable trace of acquisition.

Children (11) — more specific cases that build on this

• Absorptive Capacity is a kind of Learning

  Absorptive capacity is a specialization of learning. The general pattern is the durable, experience-driven self-update of an agent's internal capability such that future performance changes. Absorptive capacity instantiates this with the experience being exposure to external knowledge and the update being internal processes that recognize, assimilate, and apply it. Prior related knowledge determines what can be acquired, which is the schema-formation mechanism learning depends on. It is learning specifically organized around uptake from outside sources rather than from direct experience or instruction.

• Collective Systemic Learning is a kind of Learning

  Collective systemic learning is a kind of learning specialized by the locus of the durable update: the learning agent is an organization, team, consortium, or community rather than an individual mind. It inherits learning's general commitment to durable, experience-driven self-update of an agent's internal capability, and adds the specific machinery of system-level acquisition — knowledge capture, integration, embedded processes, structures, documentation, and culture — that makes the update organizational rather than personal and that lets the system as a whole adapt while individual members come and go.

• Conditioning (Behavioral) is a kind of Learning

  Learning is the process by which an agent durably updates an internal capability — knowledge, skill, model, or behavior — as a result of experience. Behavioral conditioning is the specific family of learning mechanisms that detect statistical contingencies between environmental events (stimulus-response pairings, contingent reinforcement) and adjust behavior accordingly, with generalization, discrimination, and extinction as characteristic features. It inherits learning's durable-experience-driven-self-update structure and adds the specific mechanism — contingency detection through pairing — that produces the update. A specialization of learning keyed to associative contingency.

▸ Show 8 more

Path to root: Learning → Adaptation

Neighborhood in Abstraction Space

Learning sits among the more crowded primes in the catalog (7^th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.

Family — Learning & Foresight Capacity (14 primes)

Nearest neighbors

• Pedagogy — 0.86
• Learning Curve Effects — 0.84
• Transformation — 0.83
• Observational Learning (Social Learning) — 0.83
• Foresight — 0.82

Computed from structural-signature embeddings · 2026-05-29

Not to Be Confused With

Learning must be distinguished from pedagogy, its closest neighbor and the teaching-side pair in the structural-framed build. Pedagogy is the teacher-side practice — the deliberate design and delivery of experience aimed at producing learning in another agent. Pedagogy presupposes learning (it is intelligible only as a way to cause it) but is not learning itself: the pedagogue need not change at all, and the structural roles of pedagogy live on the teaching side (instructional intent, designed experience, scaffolding, assessment of the learner) rather than on the agent-side update. Many human learning episodes involve no pedagogy (animal conditioning, immune affinity maturation, machine learning from web-scraped data, organizational learning from outage), and pedagogy can fail to produce learning despite being well executed. The pair captures the production-and-consumption structure of the broader teach/learn cluster: pedagogy produces opportunities for learning; learning consumes experience (including pedagogically designed experience) and converts it into durable capability. The two are tightly coupled in education but structurally distinct, and treating "learning" as a synonym for "good teaching" loses the agent-side update that is the whole point of the learner-side prime.

Learning is also distinct from adaptation, which it parents in the DAG. Adaptation is the broader pattern of a system changing structure or behavior under sustained pressure to maintain fit with its environment; it includes evolutionary change at the species level, physiological acclimation in organisms, ecological succession in communities, and structural reorganization in firms — many of which involve no agent, no modifiable internal state in the learning sense, and no experiential update process. Natural selection adapts populations without any individual organism updating its internal state from experience; physiological acclimation reshapes an organism through homeostatic mechanisms that are not learning. Learning is the experience-driven, information-driven, memoried kind of adaptation: an agent with state, experience writes to state, the change persists. The subsumption edge (learning → adaptation) holds in this strict sense: every instance of learning is an instance of adaptation (the agent has shifted to better fit its environment in light of pressure), but most adaptation is not learning (the change can occur without an agent updating its own state from experience). Confusing the two collapses an important distinction — between selection acting on populations and update acting on individuals, between structural redesign by environment and capability acquisition by agent.

Learning is distinct from memorization, a narrower process that lives inside the broader learning family. Memorization is the durable acquisition of specific representations — facts, items, sequences — without necessarily acquiring the structural understanding or transfer capability that broader learning produces. A student who memorizes a list of dates has learned the dates in the strict sense (durable, experience-driven update of internal state) but may have learned nothing about the underlying historical structure. Learning includes memorization but extends to skill acquisition, model formation, pattern recognition, and conceptual restructuring, which memorization alone does not capture. Treating learning as synonymous with memorization underestimates the prime; treating memorization as not-learning overestimates the gap. Memorization is one mechanism — a particular kind of update producing a particular kind of capability — within the structural pattern that learning names.

Learning is distinct from Conditioning (Behavioral) (classical and operant conditioning), which is a specific mechanism of learning identified by behaviorist psychology. Conditioning describes durable behavioral updates driven by stimulus-stimulus associations (classical) or stimulus-response-outcome contingencies (operant). It is a real and important mechanism — recurring in animals across species, in humans, and in some machine-learning architectures — but it is one mechanism among several. Learning by observation, by inference, by structured instruction, by deliberate practice, by symbolic problem-solving, and by neural-network gradient descent are all learning, none of which reduce neatly to conditioning. The relationship is hierarchical: conditioning is a particular update mechanism within the broader pattern that learning names, just as memorization is a particular capability outcome. Calling all learning "conditioning" imports a specific theoretical commitment from behaviorism; treating conditioning as not-learning misses that it cleanly instantiates every role of the structural skeleton.

Learning is distinct from accommodation in the Piagetian sense, which is another specific learning mechanism, this time identified by developmental psychology. Accommodation is the schema-revision mechanism: when a child encounters an experience that violates an existing schema, the schema itself is restructured to fit the new input (as distinct from assimilation, where the input is interpreted to fit the existing schema). Accommodation is one route by which experience updates internal state durably — it cleanly instantiates the four-role learning skeleton — but it is a specific cognitive-developmental mechanism, not the whole of learning. Conditioning, gradient descent on a neural network, affinity maturation in B cells, and reweighting of pattern-recognition heuristics in a chess player are all learning without being accommodation in the Piagetian sense. The structural relationship is the same as with conditioning: accommodation is a particular mechanism inside the broader pattern that learning names, and the prime sits above any single mechanism so that the same diagnostic carries across cognitive, biological, computational, and organizational substrates.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.

Notes

Learning is the learner-side sibling of pedagogy. Learning parents the acquisition-side primes — observational learning, constructivist learning (which has constructivism as its second parent), mastery learning, transfer of learning (which presupposes it), spaced repetition (which presupposes it), learned helplessness, and collective systemic learning. Pedagogy parents the teaching-side methods. A key re-home is collective_systemic_learning, originally classified under feedback in R14; it sits more naturally under learning since the structural pattern it instantiates is collective experience-driven update of routines, not feedback per se.

The learning → adaptation subsumption edge and the pedagogy → learning composition/presupposes edge are proposed for the R16 re-home round. Both are structurally clean: every instance of learning is an instance of adaptation (the memoried/informational kind), and pedagogy is intelligible only as a deliberate causal upstream of learning.

The structural-framed read is that learning is a structural prime — the four-role pattern is substrate-neutral, the same skeleton runs through cognitive, biological, computational, and organizational instances, and the vocabulary travels without metaphorical stretching. The adaptive-immunity case is load-bearing: no cognition, no language, no consciousness, every role present. It forces the prime to the structural level where cross-substrate pattern is the value case.

The substrate-furthest cases are machine learning and the immune system. Both involve no cognition in the everyday sense, both cleanly instantiate every role, and both are described by their own practitioners in language that maps directly onto the four-role decomposition. The convergence is the strongest empirical evidence that the prime sits at the substrate-neutral level, not at the level of any one domain's mechanisms.

References

[1] Skinner, B. F. (1953). Science and Human Behavior. Macmillan. Systematic operant-conditioning framework: behavior is selected and durably modified by its consequences in agents from pigeons through humans. Establishes the experimental program in which experience-driven, capability-changing self-update is the central explanandum, across species and without requiring language or instruction. ↩

[2] Bandura, A. (1977). Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review, 84(2), 191–215. Cross-DP-17 cite: self-efficacy (task-specific capability belief) as the construct self-handicapping is designed to protect. (No paired FACT-D18 anchor in current draft — deferred to B-resolution) ↩

[3] Piaget, J. (1952). The Origins of Intelligence in Children (M. Cook, Trans.). International Universities Press. Foundational constructivist account of cognitive development; introduces the assimilation/accommodation dialectic in which the child constructs knowledge from interaction with the environment, supplying a specific schema-revision update mechanism inside the broader learning pattern. ↩

[4] Thorndike, E. L. (1898). Animal intelligence: An experimental study of the associative processes in animals. Psychological Review Monograph Supplements, 2(4), 1–109. Founding experimental study of trial-and-error learning in cats; the law of effect formalizes durable experience-driven behavioral updates as a function of consequence — the earliest quantitative grounding for the durability commitment that separates learning from one-off responding. ↩

[5] Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie [Memory: A Contribution to Experimental Psychology] (H. A. Ruger & C. E. Bussenius, Trans., 1913). Teachers College, Columbia University. Founding quantitative study of retention and forgetting; the forgetting curve makes durability of learned material measurable and establishes that durability is an empirically tractable property of an experience-driven internal update. ↩

[6] Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2^nd ed.). MIT Press. Standard reference on the temporal credit-assignment problem: discounting and eligibility traces back-project credit for a delayed reward across the actions that produced it (850), the same backward propagation that, applied to incident review, resists stopping at the proximate actor (855). ↩

[7] Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer. Standard textbook treatment of supervised and unsupervised machine learning; develops parameter-update mechanisms (likelihood, loss, gradient methods) that instantiate the four-role learning pattern with silicon substrate, training data, differentiable update, and retained model weights. ↩

[8] Hebb, D. O. (1949). The Organization of Behavior: A Neuropsychological Theory. Wiley. Introduces the Hebbian learning rule ("cells that fire together, wire together") as a synaptic update mechanism, grounding the requirement that learning needs a physically modifiable internal substrate; experience without such a substrate cannot produce durable capability change. ↩

[9] Pavlov, I. P. (1927). Conditioned Reflexes: An Investigation of the Physiological Activity of the Cerebral Cortex (G. V. Anrep, Trans.). Oxford University Press. Canonical demonstration of classical conditioning in dogs; rigorously distinguishes the innate unconditioned reflex (transient with the trigger) from the conditioned reflex (acquired through pairing and persisting after acquisition), establishing the reflex/learning boundary at the heart of the prime. ↩

[10] Burnet, F. M. (1959). The Clonal Selection Theory of Acquired Immunity. Vanderbilt University Press. Foundational formulation of clonal selection: antigen exposure selects from a pre-existing diverse lymphocyte repertoire, and selected clones persist as immune memory. Establishes adaptive immunity as a substrate-furthest case of the learning skeleton — modifiable repertoire, experiential input, selection-and-mutation update, lifelong durability — entirely without cognition or instruction. ↩

[11] Bjork, R. A., & Bjork, E. L. (1992). A new theory of disuse and an old theory of stimulus fluctuation. In A. F. Healy, S. M. Kosslyn, & R. M. Shiffrin (Eds.), From Learning Processes to Cognitive Processes: Essays in Honor of William K. Estes (Vol. 2, pp. 35–67). Erlbaum. Develops the distinction between retrieval strength and storage strength as independently manipulable; grounds the "desirable difficulties" finding and provides the counterfactual machinery for separating durable capability change from transient performance. ↩

[12] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. Authoritative review of deep learning: frames loss-driven parameter updates in neural networks as a substrate-neutral instance of "experience changing internal representation," explicitly connecting biological and computational learning under a shared four-role structure. ↩