Implicit Learning

Note: This material is derived in part from a chapter on "Explicit and Implicit Memory and Learning" by J.F. Kihlstrom, J. Dorfman, and L. Park, published in M. Velmans & S. Schneider (Eds.), Blackwell Companion to Consciousness (Malden, Ma.: Blackwell, 2007).

Link to Russian translation by MWMA Translations.
Link to Ukrainian translation by MWMA Translations.

 

Learning and memory are inextricably intertwined.  The capacity for learning presupposes an ability to retain the knowledge acquired through experience, while memory stores the background knowledge against which new learning takes place.  During the dark years of radical behaviorism, when the concept of memory was deemed too mentalistic to be a proper subject of scientific study, research on human memory took the form of research on verbal learning.

Explicit Learning  

Traditionally, learning has been defined as a relatively permanent change in behavior that occurs as a result of experience.  Early investigators -- Pavlov and Thorndike, Watson and Skinner -- construed learning as conditioning -- the formation of associations between environmental stimuli and organismal responses to them.  However, the cognitive revolution in psychology has led to a reconstrual of learning as a relatively permanent change in knowledge that occurs as a result of experience – declarative and procedural knowledge that the organism will subsequently use for its own purposes in predicting and controlling environmental events.  Thus, in classical conditioning the organism forms expectations concerning the likely consequences of events, and in instrumental conditioning the organism forms expectations concerning the likely consequences of its own behaviors.  How this knowledge translates into behavior is another matter.  In any case, the change in behavior is a publicly observable reflection of the change in knowledge.

In addition to classical and instrumental conditioning, researchers have studied a variety of other forms of learning -- including, but not limited to, the following.

At least in the case of humans (and certainly other primates, probably other mammals, perhaps other vertebrates, and maybe some invertebrates) this cognitive emphasis on individuals acquiring knowledge to help them predict and control events in the world implied that learning was a conscious activity.  This is also true of perceptual learning, which occurs more rapidly with active than passive movements of the observer.  And its emphasis on the role of observing, modeling, and teaching – not to mention the fact that civilization has created institutions to support these activities -- marks social learning, too, as a conscious act of mind.  Nevertheless, it is also true that some organisms, like the sea mollusk aplysia, can learn even though they probably do not have enough neurons, much less a cerebral cortex, to support consciousness.  Even in humans, who have a capacity for consciousness, it has long been evident that some learning can take place unconsciously (Adams, 1957; Razran, 1961).

Varieties of Implicit Learning

In point of fact, the concept of implicit learning was introduced into the psychological literature well before that of implicit memory.  

In a pioneering series of experiments, Reber asked subjects to memorize lists of letter strings, each of which had been generated by a Markov-process artificial grammar – a set of rules that specified what letters could appear in the string, and in what order.  Over trials, the subjects found it easier to memorize grammatical strings, compared to random strings, indicating that their learning was exploiting the grammatical structure.  Moreover, when presented with new strings, subjects were able to distinguish between grammatical and non-grammatical strings at levels significantly better than chance, indicating that they had acquired some knowledge of the grammar.  Yet when queried, the subjects were unable to specify the grammatical rule itself.  They had learned the grammar, and this knowledge had guided their behavior, but they were not aware that they had learned anything, and they were not aware of what they had learned (Reber, 1967, 1993).

At roughly the same time, neuropsychologists noticed that, over trials, amnesic patients improved their performance on such tasks as maze learning, pursuit-rotor learning, and mirror-reversed learning.  Clearly, then, amnesic patients had the capacity to acquire new skills, but they did not recognize the tasks, nor did they remember the learning experiences; moreover, they seemed to have no conscious awareness of their newly acquired knowledge.  Later studies showed that amnesic patients could learn artificial grammars, just as neurologically intact individuals do (Knowlton et al., 1992).

By analogy with memory, we can define explicit learning as a relatively permanent change in knowledge or behavior that is accompanied by conscious awareness of what has been learned.  Implicit learning, then, refers to a relatively permanent change in knowledge or behavior in the absence of conscious awareness of what has been learned.  Sometimes evidence for implicit learning is taken as evidence for implicit memory, but implicit memory is more narrowly restricted to the learning episode itself, while implicit learning covers the knowledge acquired in that episode.  In a famous case, Claparede described an amnesic patient who forgot an episode in which he pricked her hand with a pin while greeting her, but who was consciously aware that “Sometimes people hide pins in their hands” (Claparede, 1911/1951; see also Kihlstrom, 1995).  This patient was conscious of what she had learned, but displayed source amnesia -- also known as cryptomnesia, or unconscious plagiarism (Brown, 1989), a concept more closely related implicit memory (Schacter et al., 1984; Shimamura & Squire, 1987).  Implicit learning goes beyond the formation of simple associations, as in classical or instrumental conditioning, and involves the acquisition of knowledge of some complexity, at some level of abstraction.

Implicit learning has been studied in a wide variety of other experimental paradigms, in addition to artificial grammars and motor learning (for comprehensive reviews, see Berry & Dienes, 1993; Berry, 1994; Dienes & Berry, 1997; Frensch & Runger, 2003; Lewicki, 1992; Seger, 1994; Stadler, 1998).

In each of these cases, subjects demonstrate, by performance measures such as accuracy or response latency in judgment, or prediction or control of behavior, that they have acquired knowledge from experience; yet they are unable to provide an accurate account of the methods by which they achieve these results.  They have learned something new, but they do not know what they know.  

Implicit Learning and Source Amnesia

Implicit learning should not be confused with source amnesia, where the subject consciously recollects some piece of knowledge, but forgets the source of that knowledge.  

Observations of preserved learning capacity in amnesic patients led Cohen and Squire (1980) initially to propose that amnesia impaired declarative memory (“knowing that”), but spared procedural memory (“knowing how”).  Perceptual and motor skills, such as covariation detection, sequence learning, and motor learning, can  certainly be represented as systems of if-then productions, but it is not at all clear that all forms of implicit learning are procedural in nature.  For example, there is evidence that amnesic patients can acquire new conceptual knowledge in amnesia – knowledge that, ordinarily, would be represented in the propositional format characteristic of declarative memory.  

In this respect, artificial grammar learning is an interesting case.  The structure of a Markov-process finite-state grammar lends itself easily to translation into a production system: If the first letter is a P, then the next letter must be a T or a V; If the first letter is a T, then the next letter must be an S or an X; and so on.  On the other hand, subjects memorizing grammatical strings might simply abstract what a “prototypical” grammatical string looks like.  When making grammaticality judgments, subjects could then compare test items to this stored prototype – or, perhaps, to the specific instances stored in memory during the memorization phase of the experiment.  In either case, the unconscious knowledge acquired through implicit learning would more closely resemble declarative than procedural knowledge – yet another reason not to use “declarative” instead of “explicit” to label conscious memory.  

In principle, both prototypes and rule systems are abstract knowledge representations that go beyond the specific instances encountered in the study set.  Accordingly, an important question concerns the degree to which implicit learning is generalizable beyond the specific.  That is, can subjects apply a grammar learned from strings of Ps, Vs, and Ts to test strings composed of Ls, Bs, and Ys?  Reber (1967) reported that this was the case, though subsequent research has often found that transfer is substantially degraded.  Studies of transfer in other domains have also yielded mixed results.  Of course, the degree of transfer will depend on the degree of initial learning.  In the artificial grammar experiments, classification performance typically ranges between 60-80% correct, with the more frequent outcomes at the lower end of this range.  Viewed against a baserate of 50% correct, a performance at 65% may be statistically significant, but may not leave a lot of room to show incomplete transfer.

Implicit learning is distinct from mere incidental learning, where knowledge is acquired in the absence of instructions or intention to learn, but the person is conscious of what he or she has learned (Eysenck, 1982).  The critical feature of implicit learning is that it is unconscious, in the sense that the subjects are unaware of what they have learned.  Documenting dissociations between explicit and implicit learning, then, is a somewhat tricky business.  Many studies do not go into a great deal of detail into the methods by which subjects’ conscious knowledge was assessed, but it is probably not enough merely to ask subjects in the artificial-grammar experiments to describe the rule that they governs the letter strings, and count them as “unconscious” when they fail to do so.  In the first place, unless the test stimuli are very carefully constructed, even partial awareness of the rule – that the first letter must be either a P or a T, for example, may be enough to permit subjects to discriminate between grammatical and ungrammatical strings at better than chance levels.  Although investigators of explicit and implicit memory have developed rigorous standards for matching explicit and implicit tasks, similar standards are generally lacking in studies of implicit learning.

The argument that implicit learning is really unconscious is sometimes bolstered by the fact that amnesic patients show preserved implicit learning.  Of course, amnesics also forget the learning episode as well, confusing implicit learning with implicit memory.  In this regard, it is somewhat disconcerting to note that subjects can show significant implicit “learning” even in the absence of any learning experience!  That is to say, in some experimental procedures involving classification performance, it is possible for subjects to intuit the structure of the target category from test instances, even when they were denied an opportunity to learn the category during a prior study phase (Palmeri & Flannery, 1999).

However, implicit learning is not always, or entirely, spared in amnesic patients.  For example, amnesics show normal levels of perceptual learning in a visual search task, but impaired learning when contextual cues were added to the procedure (Chun & Phelps, 1999).  Even intact implicit learning by amnesic patients does not mean that explicit and implicit learning must be mediated by different brain systems (Knowlton & Squire, 1996; Knowlton, 1999).  Dissociations between recognition and concept learning can be simulated in a computational model of exemplar memory that has only one system for storing memory (Nosofsky & Zaki, 1998, 1999), with different thresholds for recognition and classification.

It is sometimes claimed that implicit learning, precisely because it is automatic and unconscious, is a very powerful – as well as more primitive form of learning – more powerful than conscious forms of learning that emerged more recently in evolutionary history (Reber, 1993; Wilson, 2002).  While it does seem amazing that subjects can pick up knowledge of something as complex as an artificial grammar or a dynamic system automatically, and apply it unconsciously, claims for the superiority of unconscious processing sometimes seem to reflect a Romantic notion of the unconscious that goes back to von Hartmann (1868/1931, p. 40), who wrote that the unconscious “can really outdo all he performances of conscious reason”.  Unfortunately, enthusiasts of implicit learning have not always compared implicit learning to conscious, deliberate knowledge acquisition.  How well would subjects perform if we actually showed them the finite-state grammar, or if we gave them feedback about their classification performance?  What if we simply told subjects the sequence of quadrants in which the target would appear?

The Implicit and the Unconscious

Together with the concept of automaticity, research on implicit learning and memory constituted psychology’s first steps toward a revival of interest in unconscious mental life (Kihlstrom, 1987).  Although the psychological unconscious suffered much in the 20^th century from taint by Freudian psychoanalysis the concepts and methods employed to study implicit learning and memory have now been extended to other domains, such as perception (Kihlstrom et al., 1992) and even thinking (Kihlstrom, Shames, & Dorfman, 1996; Dorfman, Shames, & Kihlstrom, 1996) – and beyond cognition to emotion (Kihlstrom et al., 2000) and motivation McClelland & Weinberger, 1989).  In this way, the study of implicit learning and memory offer a new, non-Freudian perspective on unconscious mental life – and, in turn, on consciousness itself.

This page last modified 05/27/2014.