Shanks, D. R (2010). Learning: From Association to Cognition

Shanks, D. R (2010).  Learning: From Association to Cognition.  Annual Review of Psychology, 61, pp. 273-301.  Retrieved on June 27, 2010 from
http://arjournals.annualreviews.org.ezproxy.usq.edu.au/doi/pdf/10.1146/annurev.psych.093008.100519


Introduction
Shanks examines the dichotomous view that learning occurs unconsciously.  Stimulus-response behaviourism has progressed significantly since the early 19th century so much so that studies in experimental psychology of the time now fill current researchers "with amusement and relief" (pg. 2) as noted by Brewer in the 1970s.  Brewer observed that processes of association formation and reinforcement are by and large irrelevant to understanding human learning.  Learning occurs as a process of cognitive consciousness.


Shanks introduces associationism which has dominated the history of learning theory, achieving prominence through Edward Thorndike who understood reinforcement processes (Laws of Effect and Exercise) that strengthen S-R connections.  Shanks indicates that Thorndike was the pioneer of what is known as implicit learning.  Rewards (punishments) can have an effect on learning even when the individual is unaware that learning is taking place.  Thorndike's work used reinforcement process as a key role to support learning as independent to awareness.  Gestaltists and field theorists believe in an opposing position with regards to effect vs. information.  They (e.g. Kohler, Koffka, Tolman) suggest that events "set the context for the development of insight and problem solving" (pg. 2) rather than provide opportunities for gradual trial and error learning.  Shanks notes that Tolman in particular argued that reinforcement is not necessary.  Cognitive relationships (e.g. hypotheses, maps) are a result of stimuli or events that perpetuate the connections.


Brewer was convinced by evidence that showed that conditioning of autonomic, motor and complex responses only occurs in parallel with expectancies and awareness.  The significance of which indicates that all conditioning is a result of cognitive processing, and particularly so in the formation and testing of conscious hypothesis.  Shanks questions whether Brewer's position on learning as an intrinsic conscious process stands up against current research.  Shanks indicates that our scope for explaining the depth of learning would be seriously curtailed if excitement and reinforcement concepts were abandoned.


Role of cognition in learning
Shanks uses the terms CS and US where CS stands for conditioned stimulus and US stands for unconditioned stimulus.  Classically, conditioned stimulus predicts unconditioned stimulus.  As a result of learning, the normal response to unconditional stimulus is evoked by conditioned stimulus (e.g. Pavlovian context - dogs are conditioned to expect food or a shock on hearing a tone).


Shanks questions whether learning is a result of cue and outcome that becomes a mental link or if it is generated by propositional belief that represents the relationship.  If true conditioning results in an unconscious and automatic learning mechanism, how does verbal, cognitive and conscious instruction make contact?  Shanks indicates that "conditioning in fact gives rise to conscious, cognitive, propositional representations rather than to automatic, unconscious ones" (pg. 4).  A position that many current researchers align with.


Blocking:  associative accounts
Shanks refers to Kamin's work on blocking which is central to the history of learning theory.  Blocking is "the fundamental demonstration that learning about the relationship between two events depends on not just their frequency of pairing, but also on the extent to which one provides information about the other" (pg. 4).  Shanks uses the following example to demonstrate: consider an individual who has allergies to some food.  On the first day, the individual is fed tomatoes and suffers an allergic reaction.  On the second day, the individual is fed tomatoes with pasta and suffers an allergic reaction.  The individual is asked to determine how much the pasta is associated with the reaction.  Blocking refers to the tomato induced allergic reaction that weakens the pasta induced allergy connection.  Cues A and B (pasta and tomato) predicts an outcome.  The degree of learning about A outcomes is reduced if B has previously and individually predicted that outcome (e.g. we don't think about the pasta as being the culprit if the tomato has caused an allergic reaction previously).


Blocking: cognitive accounts
The key observation of Waldmann and Holyoak towards association is that it has no semantic properties (different meanings can be derived from words or symbols).  They suggest that it should make no difference whether the words cues and outcomes (described as cause and effect) become effect and cause.  Imagine a sample of patients' blood that indicates the patient is suffering from disease X.  To make the learning connections between substance (cue/conditioned stimulus) and disease (outcome/unconditioned stimulus), one must learn the association.  Waldmann and Holyoak indicate that the substance may be a cause or an effect of the disease, which would alter outcomes significantly.  Their argument underlines the fact that blocking works in a cause-effect scenario, and not the other way around.  Shanks' review of literature reveals that there are instances of association that provide an appropriate account of participant behaviour, although full understanding of the boundary conditions are not available.  Shanks uses related research by Matute, Vegas and De Marez on allergic reactions to medicine that questions:

• a) to what extent medicine causes reactions;
• b) the extent to which it predicts/indicates a reaction; and
• c) the extent to which medicine and allergies occur (e.g. not all individuals are allergic to anaesthetic).

An alternative to this theory can be found in De Houwer, Beckers and Glautier's work on cognitive influences of blocking, proposing that blocking arises from a sequence of thoughts (chain of reasoning).  Shanks quotes: "the outcome is as probable and intense after B as after AB ... therefore A is not the cause of the outcome" (pg. 6).  De Houwer, Beckers and Glautier note that this conclusion is only valid if the effect is not at maximum levels, and the outcome can occur with greater probability and intensity.  They speculated that if reasoning caused blocking rather than associative conditions, then participants would have trouble identifying the status of blocks from A cue.  Tests revealed:

• Stage 1 - B cue predicts outcome; and
• Stage 2 - A and B cue occur with the same outcome.

It is hard to distinguish whether A cue had an effect or if it was masked by B cue outcome:

• If B cue causes maximum intensity, there is no way an increase or alteration can be caused by A cue even if it does have a causal effect.

When contrasting maximum intensity outcomes to submaximal, the participants indicated that intensity was at 10/20.  De Houwer, Beckers and Glautier indicate that failure of A cue to change the outcome in Stage 2 is significant.  They observed much stronger blocking in submaximal conditions and argue that associative theories are inconsistent with their results.  A review of literature indicates that "[i]f participants believe cues are additive, then the failure of B cue to cause an increase in the outcome's magnitude indicates that it is ineffective, and blocking should be substantial" (pg. 7).  But if it is believed that cues are subadditive (where A cue and B cue cause maximum intensity - 1M), then together they should cause 2M.  Failure of B cue to increase the magnitude of M does not imply lack of causal power in B.  Evidence of blocking should be weaker in subadditive conditions, and common observation will provide detail to naturally occurring patterns.


Shanks indicates a counter argument exists.  Wheeler, Beckers and Miller's research shows subadditive pretreatment is sensitive to manipulation which compromises generalisations.  They hypothesise that the effect emerges through reflection of associative structures in context that lead to blocking (e.g. I ate a tomato, I suffered an allergic reaction; I ate tomato and pasta, I suffered an allergic reaction; I think I'm allergic to tomatoes; I will avoid tomatoes from now on).  Haselgrove's simple experiment introduced an extra cue at trial stage and found that a pretreatment stage reduces blocking.  Schmajuk and Larrauri propose a model where associative connections are updated by reinforcement processes and connections have no symbolic reference.  While different from Haselgrove's model, the additivity effect is reproduced.  Schmajuk and Larrauri suggest that it is not necessary to interpret additivity effects by means of propositional reasoning. Other research provides the possibility that effects from maximality and additivity influence the shifts between elemental and configural processing.  Shanks refers to literature that suggests both human and nonhuman organisms represent stimuli in flexible ways, and indicates that as there are no set weights for each element there are no set responses across trial types.  Configural representations are required that are distinct from constituent parts.


Melchers, Lachnit and Shanks assigned new discrimination in EX+ and FX- with very different results.  The group favoured E cue to F cue as evident in pretraining on control discrimination but not in feature-neutral contexts.  This pattern of evidence is consistent with group training in control discrimination transferring elemental strategy to a new problem by breaking down components and assigning positive weighting to E cue and zero weighting to F and X cues.  Groups trained in feature-neutral contexts transferred configural strategy where compounds and elements are treated as independent.


A reported range of manipulations that may enhance elemental processing also increases blocking (similar to additivity instructions).  Manipulations that (are assumed to) enhance configural processing decrease blocking (subadditivity instructions).  When cues are presented in a configural manner, blocking can be eliminated even when an additivity pretreatment stage was included.  Hence the proposition that maximality and additivity effects influence by shifting the balance between elemental and configurational process.  However, Shanks indicates that it is difficult to profile the locus of additivity effects.


Retrospective Revaluation
Backward blocking is the reduced judgement made between A cue and its result.  Similar to forward blocking where B and AB represent the same outcome, A is not an independent cause.  Shanks indicates that the main findings are compatible with associationism.  Ghirlanda assigned 50 units each with coded elemental stimulus that revealed:

• reduction in judgments for A after AB+, B+ training (e.g. pasta is considered okay after putting pasta and tomato together, with initial positive tomato training) - yields backward blocking;
• increase in judgments for A after AB+, B- training; and
• negative judgments for A after AB-, B+ training.

Coding schemes allow cues to activate overlapping sets of elements and define the principle of stimulus generalisation.  Kruschke assumes that knowledge is not acquired by the strength of a single association but through a distribution of a range of beliefs.  These hypotheses are adjusted according to feedback.  Dickinson and Burke propose that in backward blocking A cue and B cue form a key association (within-compound) retrospectively.  Using the following example, they propose a revised associative theory:

• Stage 1 - AB cue results in an allergic reaction, associations are made to cues, outcome and the relationship between cues; and
• Stage 2 - B cue occurs independently and results in an allergic reaction.

if

• B cue activates the representation of A cue in Stage 1 using the association between cues; then
• A cue is reduced in associative strength.

Shanks reiterates that associative processes and reinforcement govern learning.  Memory retrieval processes propose an alternative account.  Participants in Melchers, Lachnit and Shanks' research accessed trial types via memory recall and replay (connectionist theory).  Replayed trials function the same way experienced trials do if memory can be recalled accurately enough to reduce A cue strength.  Backwards blocking occurs reflectively.  Significantly there was no correlation between forward blocking and participant memory in their compound trials.


Blocking: memory for the blocked cue
Shanks notes that impaired memory can make recalling judgments on associations weak as outcomes that occurred as a result of A cue involvement will not activate the same response during replay.  Inferential accounts provoke opposing predictions where participants are assumed to follow a chain of reasoning: "If A predicts the outcome, then the effect of AB would have been greater than that of B alone.  B and AB predicted the same outcome, thus A is not predictive" (pg. 11).  Studies carried out in 2006 by Mitchell, Lovibond, Minard and Lavis reveal that subadditivity pretraining leaves the participant ambiguous about B cue status.  Forward blocking on predictive and recall judgments strongly supports associative analysis.  Later work by Griffiths and Mitchell describe their model as a bridge between studies in associative learning and research on recognition and (other forms of) memory to provide "a middle way between the extreme cognitivism of inferential accounts and the equally extreme reductionism of S-R theory" (pg. 13).


Awareness and learning
Lovibond and Shanks argued in 2002 that there was no strong reason to revise the view that a necessary condition for learning is awareness.  Shanks presents some recent studies on eyeblink, fear and evaluative conditioning.  Eyeblink conditioning determines that there is a delay between the end of conditioned stimulus and the start of unconditioned stimulus.  In studies of fear conditioning where conditioned response (CR) is a change in skin conductance, participants were observed to have changes in skin conductance even when the conditioned sitmulus (a tone) was not audible and no cognitive association could be made.  Evaluative conditioning is the transfer of affect or positive feeling (liking) from one stimulus to another by pairing them (e.g. putting a favourite painting next to a painting of neutral status will increase positive feelings for the painting with neutral status; the converse is also true).


The Perruchet effect "is potentially convincing evidence of unaware learning" (pg. 17).  It was observed that participants involved in conditioned stimulus-only trials faced a decline in conditioned response.  Participants involved in conditioned stimulus-unconditioned stimulus trials were observed to increase conditioned response.  The Perruchet effect indicates that participants are more likely to have awareness when they expect a twist.  In later studies the Perruchet effect was induced through backward pairings, with the same behavioural outcomes.  Shanks notes that his review of literature plainly indicates further work is necessary to establish secure validity.  Devising tests to accurately predict unconscious learning has not been easy.


Other experimental procedures in implicit learning include artificial grammar learning, the learning of sequential dependencies in quickened reaction time settings, and the learning of contextual cues in visual search.


Sequence learning
In trials by Nissin and Bullemer a visual target was displayed in four locations, and the participants were asked to locate the target and press a button as fast as possible.  The target was programmed to move in an allocated sequence.  Participants were not aware of the sequence but learning was evident in response time.  Shanks questions that if participants can learn a target sequence, are they also aware of the sequence?  Several tests have been devised to assess awareness (e.g. reproducing the sequence post trial).  A large body of evidence reports "clear associations between learning and awareness" (pg.18).


Visual search and context cuing
Chun and Jiang devised a test that has familiar symbols in an unfamiliar context (e.g. the letter T in horizontal form) surrounded by a patterned sequence of distractors.  If participants learned the sequence, they would be able to accurately predict the target cue's appearance as evident in response time.  It was noted that participants may be unaware of the repetition of displays.  Later studies (2008) show that participants are in general aware of the repetition if prompted and assessed appropriately.  Smyth and Shanks ran trials to measure awareness and found that with 12 trials (sequence of displays) no evidence of awareness was obtained, but on trials with 48 displays reliability increased so above-chance performance was observed.  Shanks indicates that a common theme amongst this research is that awareness and implicit performance is largely uncorrelated.


Conclusions and future directions
Although the current fashion for learning appears to be heavily cognitive based unconscious (implicit) learning has become a substantial and distinct research topic.  Shanks states that the mind has two systems: one is automatic, implicit and nonrational while the other is slow, effortful, rational, explicit and conscious.  The first system presents itself as unconscious learning and the second uses inferential reasoning.  Shanks argues that it is questionable whether significant forms of learning can be transferred through automatic or unconscious processes.  Nor should it be assumed that only rational forms of learning can be reasoned out as is the case with retrospective revaluation that can be explained associatively or inferentially.  He comments "[i]t is plain that future research with more demanding tests of awareness will allow much firmer conclusions to be drawn about the limits of unconscious processing" (pg. 22).


Blocking has received considerable attention in research, and it can be observed from humans to molluscs although reasons for occurrence may be different.  In humans and rats, blocking is sensitive to manipulations of additivity.  Future work should be able to support dual-process theories that demonstrate qualitative differences more thoroughly.  Many typical experimental procedures involve meaningless stimuli (tones, lights) and Shanks suggests that studies over prolonged periods could establish expertise (e.g. table tennis experiments by Koedijker, Oudejans and Beek).


An alternative to dual-process perspectives comes from associative learning theory and connectionism in human cognition that provide key roles in associative processes.  Retrospective revaluation (the readjustment of cue weighting) "may result from activation and incremental learning within a simple network of neuron-like processing units" and "[a]lthough connectionist models are in principle consistent with unconscious learning, it is often assumed that stable states of activation within the brain, subject to selective attention, are precisely those states of which we are conscious" (pg. 22).


Shanks proposes that a priority of the research agenda should focus on gaining a more comprehensive understanding on the limits of associative principles when explaining additivity and its related effects on blocking.  Pretreatment designs have significant impact on demonstrating additivity and nonadditivity in cues although precise locus is unknown and will potentially feature prominently in future research.