implicit learning
TRANSCRIPT
Implicit learning
Zoltán Dienes Conscious and unconscious mental processes
Implicit learning People learn to make decisions on a task more accurately or more quickly without being able to justify their decisions adequately. OR:The learning process by which people come to acquire implicit (unconscious) knowledge. Consider:Acquisition of natural language, social skills, musical appreciation, many practical skills
Three common paradigms for investigating implicit learning:
1. Artificial grammar learning
2. Dynamic control tasks (complex systems)
3. Serial reaction time (SRT) task
MTTTTVMVRXVXRR
VXTVRXMTTVT
VXMMTVRX
MTVMVRXVTMVRXRR
1. VXTTTV2. MVRTR3. MVRXRM4. MTVT5. MTRVRX6. VXRM7. VRVXV8. MXRRM
1. Y
2. N
3. Y
4. Y
5. N
6. Y
7. N
8. N
Reber 1967: “implicit learning”
Art Reber
An example of a “finite state grammar” used for generating stimuli in artificial grammar learning experiments
People learn to classify test items though find it hard to describe relevant rules
Dynamic control tasks:
Subjects interact with a simulated system, e.g. the sugar production factory (Berry & Broadbent 1984)
On each trial, hire and fire workers to try to maintain the level of sugar production at a target value.
Underlying equation (unknown to subjects):
Pn = 2*W – P n-1 +N
Where Pn is the sugar production on trial n (in 1000 tonnes) and W is the number of workers (in 100s), and N is noise (randomly –1, 0, +1)
Donald Broadbent
1926-1993
Other dynamic control tasks include:
Interacting with a person, trying to make them friendly
Controlling a traffic system
Berry and Broadbent (1984)Training on the task improved ability to control the system but not ability to answer questions about how the system worked
Trying to consciously work out the rules impairs learning
Serial reaction time task
(Nissen & Bullemer, 1987)
On each trial a light goes on
Just press corresponding button
Unbeknownst to subject, sequence of lights is rule governed
Rule governed
Violates rules
Subjects are sensitive to the presence of the sequence even when they deny knowing that there was a sequence
To investigate implicit learning
First need to have a method for determining when knowledge is conscious and when unconscious
Then can investigate the properties of unconscious knowledge and how it might differ from conscious knowledge in terms of:
The conditions under which learning occurs and can express itself
What can be learned
etc
How can we tell if knowledge is unconscious?
1) Use free report.
2) Wordly discriminations – if people are objectively at chance on discriminating whether a feature is relevant to a task, they cannot have conscious knowledge about the relevance of that feature
3) Subjective measures – are people at chance at discriminating the mental state they are in?
Just having a first order representation means you know (and can make wordly discriminations) – but you are not aware of knowing (=> unconscious knowledge). Need to have a relevant higher order thought to have conscious knowledge.
1. Free report Mathews et al (1989)
- ‘original’ subjects exposed to grammatical strings (training phase) then classified new strings (test phase). After every 10 classification decisions they gave instructions on how to classify.
- ‘Yoked’ subjects followed the instructions and classified the same stimuli. They had no previous training phase.
Original subjects were always about 30% better than yoked subjects
Subjects acquire knowledge that could not be elicited in free report.
Is this because the knowledge is in a form fundamentally incompatible with the processes of free report?
BUTFree report gives the subject the option of not stating some knowledge if they choose not to (if they are not confident enough of it);
If the free report is requested some time after the decision, can be partially forgotten
Subject might only try to report the sort of information they think the experimenter wants (e.g. experimenter wants rules but the subject solved the task not with rules but by remembering whole items)
2. Objective threshold (wordly discrimination) Force the subject to respond regardless of confidence, and make sure there are appropriate cues present so that the test is sensitive (cf Shanks & St John, 1994).
Dulany et al (1984)
After training on on grammatical strings, S classified new strings.
After each classification decision, subjects underlined that part of the string that made it grammatical/ungrammatical.
e.g. if MTRXR is called ungrammatical, the subject might underline: MTRXR
This can be considered a consciously expressed rule: “TR cannot occur starting in the second position”
Rule validity of rule: Percentage of test strings that would be classified correctly if just this rule was applied. (If string has feature, call it nongrammatical; if string does not have feature, call it grammatical)
R=0.83 slope = .99 intercept = .01
=> Subjects conscious rules predicted classification performance without systematic error.
3. Subjective measures (discrimination of mental states)
Does wordly discrimination really measure whether a mental state is conscious?Higher Order Thought theory – a mental state is conscious only if there is an appropriate higher order thought. Does the person know that she knows?Absence of wordly discrimination performance- good evidence for absence of conscious knowledge; but presence of wordly discrimination performance not good evidence for presence of conscious knowledge. Need a measure more tuned to conscious knowledge rather than any knowledge.
Two criteria for measuring implicit knowledge (Dienes & Berry, 1997): Guessing criterionWhen subjects believe they are literally guessing, is their performance above chance? Zero-correlation criterionIs their a lack of relationship between confidence and accuracy? Do subjects fail to discriminate between guessing and knowing?
The most direct way of testing for conscious knowledge is to test for higher order thoughts (Dienes, Altmann, Kwan, & Goode 1995)
The method has an advantage over free report in that
- low confidence is no longer a means by which relevant conscious knowledge is excluded from measurement; rather the confidence itself becomes the object of study
- can be quickly assessed on every trial
- it does not matter what knowledge the subject has, we don’t need to know how the task was solved
- groups of trials can be divided into those involving just unconscious knowledge (satisfy the guessing criterion) and those involving some conscious knowledge (as shown by the zero correlation criterion)
Artificial grammar learning
People asked to memorize a set of training strings
Then informed of the existence of rules and asked to classify new strings as following the rules or not.
Take confidence ratings in the test phase and apply guessing and zero correlation criteria.
Confidence ratings given on a 50-100 scale where
50 = literal guess, expected performance is 50%
100 = complete certainty, expected performance is 100%
Intercept above 0 => some unconscious knowledgeSlope > 0 => some conscious knowledge
Dienes, Altmann, Kwan, & Goode(1995). Advantage of trained group over untrained baseline:
Artificial grammar learning
Two problems with subjective measures:
(i) The problem of bias in reporting mental states
(ii) The problem of unconscious knowledge informing verbal reports of mental states
Problem 1: The bias problem:
When a person says guessing they may include cases where their higher order thought is that they know to some extent
If there is such a bias, guessing criterion may only indicate above chance performance because of the cases where HOTs are actually about knowing to some degree.
NB: If person is biased in this way, and HOTs accurately reflect mental states, zero correlation criterion should still indicate relationship between confidence and accuracy.
(One can measure a person’s ability to discriminate mental states independently of bias)
If there is no relationship between confidence and accuracy, subjects cannot discriminate when they are guessing from when they know: so there is no bias problem.
But often there is a relationship between confidence and accuracy. Should we discard the guessing criterion in those (most common) cases??
Forcing the subject to say “guess” less often should force them to choose more carefully which states are guess states, giving a higher confidence rating to the cases where they have a little bit of confidence.
=>Forcing subjects to say “guess” less often should, if the skeptics are right, reduce the percentage of correct answers when subjects say they are guessing.
Twyman & Dienes (submitted)
In an artificial grammar learning task, one group of subjects in the test phase were told after low confidence decisions that they were on average being under-confident (“warning group”). Another group were given no such warnings.
Confidence
1009080706050
Cla
ssifi
catio
n pe
rfor
man
ce
100
90
80
70
60
50
Overall: guessing criterion satisfied, but ZCC indicates conscious knowledge. Could guess responses be biased, i.e. include decisions for which the subject consciously felt a little bit of confidence, they knew that they knew to some extent?
6263N =
WARNING
warningno warning
Mea
n +-
1 S
E N
umbe
r of g
uess
resp
onse
s
14
12
10
8
6
4
2
=> The warnings DID reduce the number of guess responses subjects gave, so the manipulation worked.
The crucial question: Did the manipulation reduce the percentage correct when subjects believed they were guessing?
No significant difference between the two groups.
=> Subjects could not choose from amongst their guesses those cases in which they really knew a little bit rather than nothing. There was no problem of subjects being “biased”, i.e. including in their “guess” responses cases were they actually had some conscious knowledge.
3359N =
WARNING
WarningNo Warning
Mea
n +-
1 S
E pe
r ce
nt c
orre
ct w
hen
conf
=gue
ss
1.00
.95
.90
.85
.80
.75
.70
.65
.60
.55
.50
.45
Problem no. 2
The problem of unconscious knowledge informing verbal reports of mental states
“Even if confidence is related to accuracy, this may not indicate conscious knowledge, because surely confidence can be based on implicit knowledge?”
To interpret this as a legitimate point, we start by asking:
WHAT knowledge does the zero correlation or guessing criterion indicate is conscious or unconscious?
Training phase -> knowledge of structure of training items (structural knowledge)
Test phase -> knowledge that an item does or does not have that structure (judgment knowledge)
Presumably, conscious structural knowledge leads to conscious judgment knowledge
But if structural knowledge is unconscious, judgment knowledge could be conscious or unconscious.
Consider natural language: If shown a sentence one can know it is grammatical and consciously know that it is grammatical, but not know at all why it is grammatical
If both structural knowledge and judgment knowledge unconscious => phenomenology is of guessing
If structural knowledge unconscious but judgment knowledge conscious => phenomenology is of intuition (cf natural language)
In both cases, we have unconscious structural knowledge.
But in second case, zero correlation and guessing criteria might show all knowledge is conscious – because those criteria only assess judgment knowledge
Dienes and Scott (2005)
In test phase, subjects rated confidence in judgment and rated the basis of the judgment:
1. Guess – judgment has no basis whatsoever, may as well have flipped a coin
2. Intuition – have some confidence in judgment, but have no idea why it’s right
3. Pre-existing knowledge – judgment based on knowledge I had before the training phase
4. Rules – judgment based on rules acquired from the training phase I could state
5. Memory – judgment based on memory for training strings or parts of training strings
Independent variables:
1. In the training phase, urged to search for rules or just memorize exemplars.
Rule search should encourage the development of conscious structural knowledge.
2. In the test phase, classify with full attention or while performing a demanding secondary task (random number generation). Secondary task should interfere with the application of conscious structural knowledge.
Confidence
1009080706050
Perc
ent c
orre
ct
100
90
80
70
60
50
There is a slope (p < .0005) – Zero Correlation criterion indicates at least some conscious knowledge
Intercept at guessing above 50% accurate (p = .01) => guessing criterion satisfied for there being some unconscious knowledge
=> There is conscious and unconscious judgment knowledge; what about structural knowledge?
71597371N =
MemoryRulesIntuitionGuess
95%
CI p
erce
ntag
e co
rrec
t
1.0
.9
.8
.7
.6
.5
NB: proportion correct significantly above .50 for each basis
Confidence accuracy relation
Learning Condition
Rule searchMemorise
Cha
n di
ffere
nce
scor
e
.30
.20
.10
0.00
Attentional Conditio
No Distraction
Distraction
No effect of independent variables on confidence-accuracy relationship – no evidence that these variables affect amount of conscious judgment knowledge (contrast Chan 1992).
What about structural knowledge?
Implicit basis
(Guess plus intuition)
Learning Condition
Rule searchMemorise
Pro
portio
n co
rrec
t cla
ssifi
catio
n
1.00
.90
.80
.70
.60
.50
Attentional Conditio
No Distraction
Distraction
Explicit basis
(Rules + memory)
Learning Condition
Rule searchMemorisation
Prop
ortio
n co
rrec
t cla
ssifi
catio
n
1.00
.90
.80
.70
.60
.50
Attentional Conditio
No Distraction
Distraction
When there was an implicit basis: No effect of learning condition nor secondary task on percentage correct
When there was an explicit basis: A secondary task disrupted correct classification in the rule search condition
Conclusion for problem 2:
The conscious status of judgment knowledge, and its basis, structural knowledge, can be assessed independently.
Unconscious structural knowledge can produce conscious judgment knowledge.
The conscious/unconscious status of structural rather than judgment knowledge more sensitively reflects the effects of learning mode and secondary task.
NB: Problem 2 is not actually a problem for whether zero correlation criterion and guessing criterion measure conscious knowledge – they do, of judgment knowledge.
Conclusions for problems with subjective measures:
1. Bias is an issue but not a reason for avoiding subjective measures like the guessing criterion: For some commonly used materials in artificial grammar learning research, it is not a problem.
2. It is useful to distinguish structural and judgment knowledge and the conscious and unconscious status of each. There are plausible methods for doing this.
Application of subjective measures to other paradigms?
Destrebecqz and Cleeremans (2001)
People trained on SRT task.
Rule governed
Violates rules
Then told that there were some rules; asked to generate a sequence of responses that did NOT follow the rules (exclusion task).
Still generated responses that followed the rules! (p < .01) (exclusion significantly above baseline)
Why might this happen?
Subjects think of a response – their implicit knowledge will tend to make them think of legal responses
If they feel they are sure the response is legal they withhold that response;
If they feel they are guessing about whether it is legal or not (they don’t know that they know) they press the key for that response.
Hence this task relies on a subject’s assessment of his mental state – it is a disguised subjective measure!
Summary of methods for determining the conscious or unconscious status of knowledge:
1. Free report: Good face validity but can be very insensitive
2. Wordly discriminations: Allows only asymmetric inferences: Null discrimination good evidence for lack of conscious knowledge; but good discrimination uninformative about the conscious status of the knowledge. Further, when there is null discrimination, not only is there no conscious knowledge, but unconscious knowledge might be low, degraded or non-existent.
3. Subjective measures: Good face validity. Remember to deal with bias problem.
We have an approach for determining whether knowledge is unconscious.
What properties does unconscious knowledge have?
1. You may be able to use it even when consciously distracted
2. It may have preferred constrained contents
1. You may be able to use it even when consciously distracted
Unconscious knowledge does not need working memory/executive system for its application – so should apply even when distracted.
Many people have tried to test this, with conflicting results (cf Jimenez, Shanks). Conflicting results may be because few tasks are pure measures of unconscious knowledge .
Few people have used subjective measures to separate conscious and unconscious knowledge.
Dienes, Atmann , Kwan & Goode (1995):
Training phase: People memorized some grammatical strings
Test phase: Classified strings.
Half the subjects simultaneously generated random numbers: every couple of seconds they said a digit, the sequence of digits had to be random (each digit follow every digit equally often etc).
Half the subjects devoted all their attention to classification.
Responses divided into those the subject gave a guess response to and those the subject had some confidence in.
Guessing knowledge not affected by secondary task; confident knowledge is affected.(NB: such results are also arguments against the bias problem)
Performance
2. What are the contents of unconscious knowledge?
In implicit learning paradigms
a. People store items and classify a test string depending on how similar it is to one or more training strings (Brooks, Vokey, Higham)
b. A stored item may include details of what the subject was doing (Whittlesea: “episodic processing”)
c. People learn fragments of strings, mainly bigrams (e.g. “MT”), but some trigrams (e.g. “XTV”) (Perruchet)
In principle, the above knowledge contents could be conscious or unconscious
d. People learn by adjusting weights in a neural network, acquiring knowledge like the above, but knowledge that can grade into more abstract knowledge as well (Cleeremans)
Input units: pattern of activation codes e.g. which light is currently on in an SRT task
Activation flows along the weights according to their value (synaptic strength); the value is changed with learning so that the output better matches reality
Output units: prediction of which light will be on next
Pattern of weights codes knowledge of sequential regularities
Neural network models:
Cleeremans (1993)
Used the Sequential Recurrent Network (SRN) of Elman to model sequence learning (SRT).
Hidden units come to encode a variable graded window into the past. Weights encode knowledge of sequential statistical regularities in an analog, graded non-conceptual way. The ideal bearers of unconscious knowledge?
Axel Cleeremans
Summary:
Implicit learning is the process of acquiring knowledge you are not aware of
Implicit learning can be measured in the artificial grammar, dynamic control and serial reaction time tasks.
When we measure the extent to which subjects are aware of their knowledge we find:
1. Unconscious knowledge is resistant to secondary tasks
2. In tasks where people often acquire knowledge they are not aware of, the knowledge may be largely statistical