Learning: Is There a Single Definition That Fits All?

A long-standing puzzle in learning science is deciding which kinds of system change should count as “learning.” The literature notes that behavior can shift without learning: injury, hormonal effects, maturation, and senescence can all expand or shrink an organism’s performance range. Because these changes are usually treated as distinct from learning, older definitions often tried to exclude them using negative qualifiers—an approach researchers say is conceptually awkward.

A proposed fix is to use a positive qualifier: learning should be triggered by specific input rather than by any internal change. Many definitions therefore narrow learning to processes that extract information from experience. That framing is attractive because it can apply to organisms and machines, but it raises practical questions: what exactly is “experience,” and must a learner be aware? Computers can learn without any connotation of feeling or awareness, making “experience” a slippery term.

Alternatives—like “activation” or “sensed environment”—also struggle with operational clarity. In theory, an environment can be defined for any system, but in practice it is difficult to specify the boundary between internal processes and what counts as external input. Even for nested systems, where one subsystem’s internal signals could feed another, critics argue that the ultimate causal driver is still environmental.

Information theory offers a cross-disciplinary anchor: learning relies on processing information. Yet this idea risks becoming too broad. Processes such as gene expression during maturation and hormonal modulation do involve information-handling at some level, meaning that a mere “information processing” criterion would overinclude effects that most researchers would not call learning.

The emerging compromise is to require that a system has (1) a learning-capable mechanism (such as synaptic updating or neural network weight changes) and (2) the ability to detect information from its environment. This would exclude purely genetically programmed maturation, while still allowing learning-relevant plasticity during development—such as experience-expectant designs that prepare sensory systems to anticipate incoming signals.

Crucially, the definition also addresses a common intuition: learning should change how a system responds to information it has processed before. A person who hears “Paris is the capital of France” reacts differently on repetition if the statement has been learned. Rather than defining learning as any behavior change toward a stimulus, the argument emphasizes that the system must alter its reaction pattern to processed environmental input.

To avoid the vagueness of “novel” information, the discussion reframes learning as a change in the system’s response to repeated information—so learning can occur even when the input is identical. In that view, learning is not just a structural change; it is an updated mapping from information processed from the environment to future system reactions.

Lövdén and colleagues present this as a definition “to rule them all,” aiming for a consistent, system-general concept that captures both biological plasticity and machine learning, while keeping the boundary between learning and other adaptive shifts.