1
IMPLICIT LEARNING
History and applications
Paul J. Reber, Laura J. Batterink, Kelsey R. Thompson, and Ben Reuveni
Introduction and history
The term âimplicit learningâ was first published 50 years ago in a report titled, âImplicit learning of artificial grammarsâ (A. S. Reber, 1967). This paper described studies with a novel paradigm aimed to create a laboratory analogue of language learning. The new approach was based on using mathematical formalisms for stimulus creation that were similar to ones being developed to help understand human language function. Surprisingly, participants exhibited an unusual behavioral pattern in their learning process. They appeared to be learning to be sensitive to the statistical structure of the underlying formalism, but seemingly without any awareness that there were any underlying rules. This report established the possibility of a dissociation between learning that could only be exhibited through performance and more traditional learning and memory that was available to conscious awareness. Over the next several decades a wide variety of additional studies and many more novel paradigms were constructed to drive research into understanding the phenomenon of implicit learning (A. S. Reber, 1989; P. J. Reber, 2013).
The historical context in which the original report was published provides some insight into why this finding had such a widespread, enduring impact and how the idea of implicit learning came to be foundational to the modern characterization of memory systems theory (Squire, 1992) and the cognitive neuroscience of memory. The story of the basis of the original studies (A. S. Reber, personal communication) starts with a chance meeting between the author, Arthur S. Reber, and George A. Miller in the early 1960s. Miller had fairly recently published the seminal paper on âthe magic number 7â and working memory (Miller, 1956) which is often cited as one of the core reports demarking the shift in the field of psychology away from behaviorism and to cognitive psychology, known as the Cognitive Revolution. Other notable publications also considered in the same vein include Broadbent (1958), Newell, Shaw, and Simon (1958) and a review written by Chomsky (1959) highly critical of a book by B. F. Skinner (1957) titled Verbal Behavior.
The Cognitive Revolution was effectively a movement against and away from the Behaviorist school that had attempted to put psychology on a robust scientific footing through the use of simple, well-characterized tasks with quantifiable measures that allowed for robust, reliable experimental paradigms. In practice, this meant using tasks from the tradition of physiologists (e.g., Pavlovâs conditioning research) that could also be studied in animal models. However, the extrapolation from animal cognition to human cognition has always posed some difficult questions, in particular when considering complex human cognition and especially the process of language, which is effectively unique to humans. Skinnerâs suggestion (1957) that language could be explained from reinforcement and conditioning studies was forcefully rejected by Chomsky (1959), implying that the study of human cognition needed a different approach.
The new approach favored by Chomsky led to his seminal work developing the field of computational linguistics. Early explorations of this work appeared in the Handbook of Mathematical Psychology (Luce, 1963) which includes three chapters authored or co-authored by Chomsky outlining how language production and comprehension might be modeled with formal grammars. Two of these chapters were co-authored with Miller, which provides some context for how Reber, as a graduate student at nearby Brown University, came into contact with these formalisms through Miller (at Harvard) via their occasional interactions.
While Chomskyâs research program can be seen as characterizing mathematical formalisms that would account for human language production and comprehension, Miller and Reber were considering a separate but related problem. If these grammars were how humans accomplished language, how does a human acquire them? The formalisms seemingly required to account for language use appeared to be exceedingly complex and possibly entirely unlearnable, especially considering the cognitive abilities of newborns. One approach was to assume they were not learned, necessitating the existence of a pre-wired âuniversal grammarâ embedded in the human brain (e.g., genetically endowed). Another approach was to try to capture this learning process in the laboratory using simplified âartificial grammars,â which then led to the seminal finding (A. S. Reber, 1967) and observation of a novel type of human learning that might solve this âunlearnabilityâ problem for language.
Researchers familiar with this history are aware that the idea of implicit learning did not immediately revolutionize the study of memory or language. In fact, for much of the next several decades, there followed a great deal of debate centered on the difficult problem of establishing the âimplicitâ part of this kind of learning. With a definition of implicit learning founded on ânot available to consciousness,â establishing even the existence of this phenomenon depends critically on proving a universal null, no awareness, which is an essentially intractable problem (Merikle, 1994). While experimental techniques and measurement approaches eventually began to provide guidelines for tackling this issue (Dienes and Berry, 1997), important support for the concept also emerged from ideas being developed separately and in parallel from research in neuropsychology and neuroscience.
Cognitive neuropsychology and systems neuroscience
At around the same time as the several famous publications in cognitive psychology that launched the Cognitive Revolution were published, a landmark paper in human cognitive neuropsychology was also reported. Scoville and Milner (1957) described the famous case of the patient H. M., who exhibited severe and selective impairment to his ability to acquire new conscious memories after bilateral medial temporal lobe (MTL) removal to treat otherwise intractable epilepsy. While there had been a few prior reports of selective cognitive loss following localized brain regions (e.g., the patients described by Paul Broca and Karl Wernicke in the 19th century) the theoretical model of the time was dominated by Lashleyâs (1929) theory of equipotentiality that hypothesized that any region of the brain could support high-level cognitive function. The case of patient H. M. established that memory was dependent on a specific neural region and did not arise from mass action of neural changes across the entire brain.
Research over the next 35 years characterized the structure and function of the memory circuitry within the MTL (hippocampus and adjacent cortical areas) and established that this system was critical for the acquisition and consolidation of memories for facts and events (Squire, 1992). Patients with damage similar to H. M. are unable to acquire new explicit memories, but are able to retrieve remote episodic memories of events that occurred prior to the damage to the MTL. More recent memories are partially affected by a temporal gradient of retrograde amnesia (see Lechner et al., 1999, for a review and history), leading to the development of a theory of memory consolidation dependent on a gradual process of memory strengthening and reorganization that depends on the MTL after initial learning.
However, detailed neuropsychological assessment of H. M.âs memory capabilities subsequently indicated that not all learning processes in his brain were entirely disrupted. Corkin (1968) and Milner, Corkin, and Teuber (1968) documented improvements in performance in procedural tasks (mirror tracing), maze learning, and picture identification from fragments. Shortly after, Weiskrantz and Warrington (1970) described a broader phenomenon of intact memory from fragmentary information in amnesic patients (priming) that would come to be known as âimplicit memoryâ and very widely studied (Schacter, 1987). Together these findings indicated that another type of memory existed that did not operate in the same manner as memory for new facts and events that depended on the MTL memory system.
These findings were foundational to the development of a âmemory systems frameworkâ that aimed to connect these observations about human memory to research going on in parallel on the neuroscience of memory. The field of neuroscience also progressed remarkably over the course of the 20th century (c.f. Gross, 1999 for a highly readable overview) with a notable moment in this progression being the founding of the Society for Neuroscience in 1969. With respect to specifically the neurobiology of learning and memory, an important early paper was the work of Kandel and Spencer (1968), who began to characterize the underlying biology of synaptic change in the nervous system. It is of note that all three of these then independent lines of research on learning and memory saw significant results in a similar time frame in the second half of the 1960s. However, integration of the related ideas across these research areas did not emerge until somewhat later during the development of the interdisciplinary field of cognitive neuroscience.
A great deal of neuroscientific memory research through the subsequent years was focused on establishing and characterizing the role of the MTL in explicit, declarative memory (facts and events). While observations from patients such as H. M. were fascinating, it was understood that it would require the establishment of a model system to be able to characterize how MTL damage affected memory with experimental control. The roles of the hippocampal formation, the adjacent cortical areas (entorhinal, perirhinal, parahippocampal), and the amygdala were all studied in detail (Squire, 1992). Systems-level analysis eventually converged on the key importance of the hippocampus and the adjacent cortical areas with the amygdala playing largely a modulatory role related to emotional memory. In addition, examination of the phenomenon of retrograde amnesia following MTL damage led to the characterization of memory consolidation processes as a key feature for how the MTL operates to store information.
Evidence for consolidation theory was also accumulating in parallel in research on the neurobiology of synaptic change (McGaugh, 2000). Synergy across these areas demonstrated how cellular and systems neuroscience could inform each other in building a theory of memory (Milner, Squire, and Kandel, 1998). Connections to research on psychological phenomena directed at studies of complex cognition were not immediately evident. Animal models do not allow for research on processes related to language or subjective measures of consciousness. Instead, many of the paradigms used to characterize and quantify learning and memory processes in these animal model systems were closely related to the tasks developed by the Behaviorist researchers (e.g., conditioning models of learning) which were very well suited to neuroscientific study of learning and memory.
Implicit learning and the problem of assessing awareness
Studies of implicit learning through two decades following the original description of the AGL task aimed to better characterize this kind of learning (A. S. Reber, 1989) but struggled with the question of how to firmly establish when learning was outside awareness. Assessing a lack of awareness depends on an accurate model of the information learned by participants to guide assessments of conscious knowledge. Dulany, Carlson, and Dewey (1984) and Perruchet and Pacteau (1990) found that asking participants about the letter strings used in the AGL paradigm specifically elicited some additional knowledge related to determining whether the strings followed the grammar rules or did not. This raised the possibility that participants were inferring another type of representation that allowed them to make âgrammaticalityâ judgments without being aware of the specifics of the formal grammar. However, it was also possible that these assessments were not of the awareness of the knowledge that drove the grammaticality judgment, but reflected concomitant explicit memory for the study stimuli (which would naturally be acquired by cognitively healthy participants but might not contribute to AGL performance).
Similar questions were being raised about studies of implicit memory (e.g., Roediger, 1990). To show that this type of memory did not depend on explicit memory for previously seen stimuli, it would be necessary to show robust priming in the absence of conscious memory. In cognitively healthy participants, this proved to be extremely difficult as a participant with an intact MTL memory system will always have some explicit memory of the study items. The inability to show a strong dissociation made it impossible to rule out the hypothesis that implicit memory phenomena simply reflected a weaker form of explicit memory (similar to familiarity) rather than a separate form of memory entirely.
A new paradigm for studying implicit learning was described by Nissen and Bullemer (1987), the Serial Reaction Time (SRT) task that became quite widely popular. This task embedded a covert repeating sequence into a simple choice reaction time task. Participants were found to increase their speed of responding to a practiced sequence compared with unpracticed sequences without seemingly being aware of the repetitions. In addition to the dissociation with awareness, this paradigm was also shown to exhibit intact learning in memory-impaired patients (Korsakoffâs) in the original report. Like with the AGL paradigm, concerns emerged over the content of the representation (Reed and Johnson, 1994) which led to protocol improvements without changing the basic character of the finding. However, the development of increasingly sensitive measures of explicit sequence knowledge (Perruchet and Amorim, 1992; Willingham, Greeley, and Bardone, 1993) started to show the same pattern observed in other tasks used in implicit memory research. Participants with intact explicit memory tended to have at least some memory for the covertly embedded (implicit) information, even if it was not clear that it contributed to task performance.
Memory systems theory
The emergence of an integrated memory systems theory that used an interdisciplinary cognitive neuroscience approach eventually showed how the neural basis of memory function in the brain could be used to help understand the type of learning observed in implicit learning paradigms. Squire (1992) described a taxonomy of memory types within a single major subdivision based on the importance of the MTL memory system. Declarative memory referred to information that required the MTL memory system to store (and consolidate) and produced representations that were generally available to awareness and verbal report. Nondeclarative memory described a collection of other phenomena that did not depend on the MTL memory system but were instead supported by synaptic change in other circuits.
Applying this framework to phenomena of implicit learning, Knowlton, Ramus, and Squire (1992), and Knowlton and Squire (1996) showed that as predicted, AGL was intact in patients with severely impaired memory due to MTL damage. P. J. Reber and Squire (1994; 1998) established the same parallel finding for the SRT task with techniques in protocol design and awareness assessment that had been advanced since Nissen and Bullemer (1987). Research on implicit memory with particularly severely memory-impaired patients indicated that it was possible to observe intact priming in the complete absence of explicit (declarative) memory for stimuli (Hamann and S...