Subconscious Learning:

Subconscious Memory Affects Performance

in the Absence of Conscious Strategy

W. Williams

2001

Abstract

Subconscious learning was measured using a variation of the invariant digit paradigm as developed by McGeorge and Burton (1990). Ninety participants performed an arithmetic distraction task involving 10, 30 or 50 four-digit number strings, all of which conformed to a simple hidden rule, based on a recurring invariant digit. They were then given a 2-second forced-choice memory test in which they were presented with pairs of strings and were led to believe that one string of each pair had appeared in the previous study phase. Whilst both strings were however new, one of the pair conformed to the hidden rule. Test and study strings were manipulated to prevent similarity and repetition confounding performance. When 30 or 50 strings were presented at study, participants responded significantly more often to the hidden rule. Participants were unable to verbally report explicit awareness of the hidden rule, even when it was suggested. This was interpreted as performance based on implicit acquisition of the invariant rule, and accounted for by the increased exposure to the rule and the inhibition of explicit processes through the 2-second response deadline. Further congruence with previous literature and implications were discussed.

Introduction

Subconscious learning is typically defined as �information acquired without intention, in such a way that the resulting knowledge is difficult to verbalise� (Berry and Dienes, 1993), although it has gathered many definitions during its development as a popular field of investigation. These are largely based on awkward distinctions between incidental and intentional learning, conscious and unconscious knowledge, and performance and awareness. It is then, not surprising that a difficulty in conceptualising subconscious learning at the outset is concurrent with its elusive nature as a focus of study. The existence of subconscious learning has been inconclusively argued for over thirty years, its presence largely inferred through a series of studies demonstrating a dissociation of task performance from verbalisation of knowledge. That verbal awareness is not an accurate reflection of task performance and the involved mental processes is clearly problematic in any situation upon which it is relied, a factor which has influenced the demand for conclusive evidence regarding implicit acquisition of knowledge and the processes involved.

Experimental paradigms to investigate subconscious learning have typically developed to involve three components: exposure to some complex rule-governed environment under incidental learning conditions; a measure to test the subject�s acquisition of knowledge about this environment through performance on the same or a different task; and a measure of the extent to which subjects are conscious of the knowledge they have acquired (Cleeremans, Destrebecqz & Boyer, 1998). To minimise the influence of subject�s prior knowledge, most paradigms involve complex, semantically neutral and arbitrary stimulus domains.

Artificial grammar learning

Artificial grammar learning was the first of such paradigms to investigate subconscious learning (e.g. Reber 1967, 1993, Dienes, Broadbent and Berry, 1991), whereby a recurring grammatical construct is embedded in a number of study strings. Subjects are typically asked to memorise a series of such letter strings generated by a finite-state grammar, followed by a surprise test phase where they are asked to recognise previous strings from new strings. Whilst in fact the test strings are all new, some conform to the grammatical rule, and the robust finding is an above-chance persuasion to these strings, despite verbal ignorance of any grammatical rules. It has also been demonstrated that knowledge acquired about the structure of an artificial grammar using one character set could be transferred to a different character set conforming to the same underlying grammatical structure, providing further non-instance based support for the semantic nature of subconscious learning (e.g. Reber, 1967, Mathews et al. 1989).

The controversy of such studies, surrounding suggestion of subconscious learning based on the lack of verbal reasoning behind decisions has led to three main interpretations: participants have either implicitly learnt knowledge of the grammatical rules; based decisions on similarity between test and study strings; or the tests of verbal knowledge used are insensitive. Shanks and St. John (1994) summarise the main methodological flaws inherent in such studies with two main invalidities: a disparity between the information the subject is using to perform the task, and the information the researcher believes them to be using; and a lack of sensitivity of the verbal knowledge tests to actual awareness, such as low confidence or fragmented knowledge. Johnstone and Shanks (2001) also maintain that subconscious learning is actually based on fragmented episodic knowledge of exemplars, and that there is no evidence to suggest that memorisation leads to passive abstraction of rules or encoding of whole training exemplars.

It is common within artificial grammar studies that participant introspection reveals details of classification strategies irrelevant to the actual grammar, despite significant results. Dienes, Broadbent and Berry (1991) argue against the analysis of introspections such as those reported in the study of Reber and Lewis (1977), claiming that the reported strategies although not conforming exactly to the artificial grammar, may still account for the participant�s classification. The difference between the rules consciously noticed by the participant and the actual grammar are likely to account for the lack of a ceiling effect, in that only a proportion of the test stimuli conformed to the participant�s inferred rules (see also Perruchet and Pacteau, 1990). As evidence to support this notion, Dulany, Carlson and Dewey (1984) attempted to increase sensitivity to awareness by asking subjects to underline the test string of their forced-choice decision, with letters that justify their decision of the string being grammatical. It was found that the highlighted fragments of the strings corresponded to classification, suggesting that subjects were indeed conscious of their knowledge. It could be argued that artificial grammar learning therefore compromises validity through its sheer complexity and lack of control over interference from confounding factors.

Invariance learning

Invariance learning (McGeorge and Burton, 1990) provides an ideal substitute to simplify some problematic complexities found within other paradigms. This is achieved through a simple and quick methodology, a well-defined stimulus set, reliable results and a hidden rule in the form of a recurring digit providing easily verbalisable knowledge.

In McGeorge and Burton�s original study (1990), participants were exposed to a series of 4-digit study strings, through a distraction task involving simple arithmetic of the digits. Each of the strings contained the invariant digit 3, although this was not drawn to the attention of the participants. A second unexpected stage of the experiment subsequently presented participants with ten pairs of new strings under the false pretence of a recognition memory test, suggesting that one of the pair had been previously seen in the study phase. Whilst in fact neither of the pair had been seen before, one of the pair did conform to the hidden rule of the invariant digit 3 (the positive string), whilst the other did not (the negative string). The robust and replicated finding was that participants selected significantly more positive than negative strings at test, even though they expressed considerable surprise when the hidden rule was subsequently explained.

Subsequent experiments of the study investigated the level of supposed implicit memory through manipulating the encoding of study strings through visual processing, and conversion of test strings to words (e.g. �3245� at study and �seven two one three� at test). The findings were robust, maintaining significantly higher selection of positive strings chosen across different encoding and recognition modalities, similar to the transference found in artificial grammar learning. McGeorge and Burton (1990) concluded that subject�s performance at test was not explicit and should be attributed to implicitly acquired knowledge held at a semantic, rather than perceptual, level. Although it is easy to imagine that subjects may learn rules that are correlated with the complex set of rules underlying an artificial grammar (e.g. Perruchet and Pacteau, 1990), it is less straightforward to explain the invariant-learning results in terms of partial knowledge of a single rule. This is not to suggest however that invariance learning conclusively supports subconscious learning, as the paradigm has since gathered a number of alternative explanations.

Invariant selection based on explicit strategies

Cock, Berry and Gaffan (1994) found evidence to refute McGeorge and Burton�s (1990) interpretation of implicit rule-abstraction, arguing an increased similarity between study strings and positive test strings. According to this instance-based hypothesis, subjects incidentally encode instances from the learning set, to be subsequently episodically compared with each of the two test instances. Positive strings are then mostly selected due to their increased similarity with study strings, already guaranteed to share at least one digit with each. This simple post hoc cognitive function matching the test to study strings is even more probable in light of the fact that the participants were led to believe it was a memory test. Through analysis of similarity within McGeorge and Burton�s exemplars, based on a relatively simple matching index, Cock et al. found that responses were biased towards more similar test strings. However, the hypothesis was unable to classify all subject responses, such as dissimilar negative string selection, and failed to distinguish between selected and rejected positive strings.

Cock et al. (1994) also directly investigated the hypothesis through manipulation of the test strings for similarity to study strings, based on test string variables of positive or negative, and similar or dissimilar. When similarity was controlled for, selection of the positive strings was at chance level, however when both strings were positive or both negative, similar strings were chosen significantly more often. Cock et al. conclude that response of subjects in both studies was more consistent with similarity than the hidden rule.

Through manipulating the nature of the strings, participants may have been forced to adopt specific explicit strategies based on similarity, as strings are either similar or dissimilar. Responses of subjects in Cock et al�s study reported response strategies based on familiarity, checking the sums of pairs, looking at the sequence of digits, and simply selecting larger numbers. Cock et al�s claims are therefore merely speculative, and hold little support as an explanation for the original findings of McGeorge and Burton (1990), although the results do suggest that both studies may have inhibited any implicit knowledge effects through the incidental provision of explicit strategies. Whilst Cock et al. remain neutral as to whether similarity matching is an implicit or explicit process, others (Perruchet, 1994, Shanks & St. John, 1994) argue that such instance-based learning is episodic and therefore explicit. This problem is compounded by the fact that the nature of the learning under more standard conditions remains difficult to interpret.

Artificial learning grammars have also shown a similar mixed influence of implicit and explicit processes. Vokey and Brooks (1992) found that the Reber studies (1967) also confounded the influence of similarity when measuring for salience of item knowledge in the learning of artificial grammars, although when controlling for the effects of similarity and artificial grammars, both were found to be major influences of performance (see also Brooks and Vokey 1991). McAndrews and Moscovitch (1985) found that whilst some subject�s classifications were based solely on one exemplar�s similarity to another, other subjects were able to implicitly distinguish between grammatical conformities within a set of exemplars sharing equal similarity to those seen in a previous study phase. It seems from this finding that similarity and hidden rules compete for dominance in influencing performance, and that either or both may contribute to the findings of McGeorge and Burton (1990).

Acknowledging the weak effect of Cock et al�s (1994) instance-based analysis of McGeorge and Burton�s (1990) study, Wright and Burton (1995) suggested that any alternative explanation of invariant learning should logically classify participant�s responses to a more accurate level than the presence of the invariant digit within the test strings. Successfully fulfilling this criteria, they then proposed a new alternative strategy in which test strings are chosen not through selection, but rejection of the alternative based on within-string repetitions of digits (e.g. 3446). Participants may perform invariant recognition indirectly through rejection of the negative strings, based on the presence of repetitions as distinctively novel strings. This is again due to the problematic construction of the strings in the McGeorge and Burton study. Variant digits were randomly chosen from the pool of 8 available digits (excluding 3 and 0), three for the positive strings (excluding the invariant) and four for the negatives. Based on the extra digit chosen for the negatives, the probability is far higher that such strings will contain a repetition. This difference is likely to reflect the difference between positive and negative selection.

Analysis of McGeorge and Burton�s exemplar sets for repetition revealed a classification system based on such distinctive strings. However, like Cock et al�s similarity based analysis, when repetition was controlled for there was still a residual effect of the invariant digit. Despite Wright and Burton�s contention that performance is more influenced by the presence of repetitions than invariant digits, they could not account for the whole effect using this measure alone.

Stadler, Warren and Lesch (2000) interestingly found that whilst distinctive repetition could not account for the residual effect of invariance learning, it did account for the cross-modal transference. When the strategy was prevented through an absence of repetitions, cross-modal transference ceased to occur, only to reappear when the strategy was again introduced. That transference is a consequence of idiosyncratic based rejection suggests that the typically supported �deep structure� of subconscious learning may in fact be an illusory consequence of an explicit confound.

In an attempt to prevent problematic effects of digit-based exemplars, Bright and Burton (1994) used clock faces to investigate subconscious learning of a hidden rule, using times between 6 and 12 o�clock as the invariant within study phase stimuli. When faced with two new clock faces in a subsequent supposed recognition task, subjects recognised novel clock faces conforming to the rule more often than not. More importantly, there was no preference for previously seen times over times that were novel but conformed to the hidden rule, implying that similarity does not dominate over the hidden rule as an influence of performance. In a continuation of their instance-based argument, Cock et al (1994) suggested that recognised novel clocks may still be based on similarity, due to the ambiguous nature of the stimuli. As they lacked hour numbers on the faces and the clock hands were ambiguously placed, the faces may have been encoded approximately by the subjects, and when coupled with an unreliable memory, a distorted preference for exact matches of previously seen exemplars may have occurred. However, some support for the effects of implicit transference was found in that invariant effects were still apparent when the test exemplars were presented in digital rather than analogue. This finding is also robust against the effects of repetition based rejection, although it may be argued that an analogous form of distinction was apparent such as on-the-hour times.

Real world support for subconscious learning based on exposure

As indicated, research into subconscious learning has relied almost exclusively on laboratory experiments, despite those focussing on incidental learning of real world regularities typically failing to show learning (e.g. Morton, 1967, Jones, 1990). This contradiction is surprising given the typically reliable learning of abstract artificial grammars (e.g. Reber, 1967, 1993) and rule abstraction (e.g. McGeorge and Burton, 1990, Bright and Burton, 1994) through exposure. Kelly, Burton, Kato and Akamatsu (2001) argue that the disparity may reflect the variable sensitivity used to measure learning following exposure. Most real world experiments have used free recall whereas laboratory methods typically use forced-choice measures of recognition. Kelly et al. examined incidental and subconscious learning of real world regularities within British and Japanese culture, using a two alternative forced-choice method typical of hidden-rule exposure experiments. Participants were found to demonstrate significantly higher correct �estimates� for their own cultural ambiguities such as the direction of the Queen�s head on a coin or stamp, than their trans-national counterparts, despite reported low confidence ratings. These results suggest that exposure to a particular stimuli allows for incidental learning, being unintentional, and implicit in that the effect is inhibited with the explicit knowledge interference of free recall measurement.

Subconscious learning in the absence of explicit strategy

It is apparent from the mentioned studies that a main problem in interpreting results of subconscious learning paradigms is the interference of explicit strategies. A majority of rule abstraction literature fails to control for the effects of instance-based learning, which may manifest either as fragmented episodic knowledge (e.g. Johnstone and Shanks, 2001) or similarity matching (e.g. Cock et al., 1994). Alternatively, variable levels of distinction found within exemplar studies provide a possible explicit strategy based on idiosyncratic rejection (e.g. Wright and Burton, 1995, Stadler, Warren and Lesch, 2000). Measures of implicitly acquired knowledge are then largely invalidated through the use of explicit strategies. Despite these methodological flaws, subconscious learning has not been reliably refuted on account of such inconclusive evidence, on the contrary its presence is suggested through evidence of residual implicit effects when the confounding variables are controlled for (e.g. Cock et al., 1994).

Meta-analyses of the literature have also concluded empirical support for the existence of subconscious learning as an entirely separate mechanism from those controlling conscious information (Lewicki, Hill and Czyzewska, 1992), although the interaction between the mechanisms is unknown (Cleeremans et al., 1998). Using the analogy of implicit and explicitly acquired knowledge to procedural and declaritive knowledge, subconscious learning has also been reliably demonstrated as resistant to psychological disorder and aging. With participants aged 50-64, Kotchoubey, Haisst, Daum, Schugens & Birbaumer, (2000) demonstrated substantial performance deficits in verbal and visual explicit learning tasks, whilst implicit perceptual learning and skill acquisition tasks were not impaired with age. This suggests that only explicit, but not subconscious learning processes are compromised in older subjects, a finding which has also been applied to young subjects (Czyzewska, Hill, & Lewicki, 1991, Vinter and Perruchet, 2000), providing compelling evidence to suggest that subconscious learning is not age-related.

Knowlton, Ramus and Squire (1992) have similarly found amnesic patients able to learn classification tasks as well as intact patients in a variety of typical subconscious learning paradigms. As strong support towards semantic rule-abstraction and against the argument of instance-based learning, these positive findings were found despite amnesics� severely impaired declaritive knowledge for the exemplars used to teach the category (see also Graf, Mandler & Haden, 1982, Graf, Squire and Mandler, 1984, Knowlton and Squire, 1993, Knowlton, Squire and Gluck, 1994).

Performance based on subconscious learning through inhibition of explicit processes

Despite strong evidence to suggest the existence of subconscious learning, there has been little attempt to resist the occurrence of explicit processes interfering with possible implicit knowledge within rule-abstraction literature. Masters (1992) demonstrated with implicit and explicit motor-skill acquisition that circumstances of pressure (induced through evaluation apprehension and financial stakes) reduced performance of explicit knowledge, whereas implicitly acquired knowledge remained unaffected. Although concluding that implicit knowledge is robust in the face of pressure, the implicit knowledge inducing distraction task was stopped during the test period, which may have relieved pressure to a neutral state. When controlling for the presence of the learning phase distraction task, Bright and Freedman (1998) found no impact of stress intervention, although the induced stress was considerably lower. Hay and Jacoby (1996) similarly find support for an increased response time affecting explicit recollection, but not habitual activities.

Rathus, Reber, Manza and Kushner (1994) also found that high levels of self-reported anxiety were associated with individuals taking longer to perform the explicit task of memorizing artificial grammar strings, however once learned, the implicit task of classifying new strings was not affected. They concluded that performance of implicit processes, but not explicit processes, are resistant to the effects of pressure. It is speculated that the residual effects of implicitly acquired knowledge demonstrated in rule-abstraction literature indicate inhibition of implicit processes by the conscious use of explicit strategies.

The present study aims to investigate subconscious learning using the invariant digit paradigm of McGeorge and Burton (1990), but modified in an attempt to resist the interference of explicit knowledge. In order to prevent the use of instance based learning, Cock et al�s (1994) condition of positive dissimilar against negative similar will be implemented. Using 30 exemplars in the study phase, Cock et al found this condition to show the most significant preference for similarity (mean 4.06 positive strings out of a maximum of ten recognised, significantly below chance), therefore any selection of the positive strings cannot be accounted for by similarity. The effect of distinction based rejection of negative strings will be controlled for by carefully designing the study and test strings not only so that negative strings are in fact more similar to study strings, but that also no repetitions are present in any of the strings (this was also controlled for in the similarity manipulation condition of Cock et al�s study). Following research to suggest the inhibition of explicit processes through pressure, a new condition will be administered to the paradigm, in an effort to achieve this. Whilst previous studies have allowed relaxed response to the test exemplars, it is possible that continued effort and conscious explicit processes were used to influence performance, such as episodic fragments of knowledge from the study phase. A two-second deadline will therefore be used during the test phase in which the participants must respond. They will be encouraged to respond as quickly as possible, but viewing both of the test exemplars to ensure a choice is not made by viewing only one. As Lewicki et al. (1992) argue that implicit information processes are significantly faster than their conscious counterparts, this condition should provide an optimum environment for implicit knowledge to be expressed without opportunity for the slower processing of explicit knowledge to interfere.

A variable of exposure to study strings will also be manipulated to investigate the effects of exposure and extended practise of the study exemplars on subsequent performance. Three experimental conditions will expose participants to 10, 30 and 50 study exemplars, before measuring performance of implicit rule-abstraction with the usual 10 test pairs. A majority of invariance research (e.g. McGeorge and Burton, 1990, Cock et al., 1995, Wright and Burton, 1994) has used 30 study exemplars, despite evidence to suggest that subconscious learning is dependent on and correlated with increased exposure to the incidental learning environment (e.g. Berry and Broadbent, 1984, Lewicki, Hoffman and Czyzewska, 1987, Sanderson, 1989, Willingham, Nissen and Bullemer, 1989, Mathews, 1990). Interestingly, in one of the only invariance studies to investigate the effects of continued exposure to study stimuli, Ward and Churchill (1998) found an extended training set to not improve implicit acquisition of rule-abstraction, although there was no knowledge of how much exposure was needed until the positive strings were selected at above-chance levels. It is hoped that manipulation of the study string exposure will clarify this. An arithmetic distractor task similar to that used in a majority of invariance literature will be used, in which participants calculate the first and last two digits of the array. This should not affect the measurement of implicit acquisition as meta-reviews indicate implicit knowledge to be resistant to distraction (e.g. Lewicki et al., 1992, Cleeremans et al., 1998). It has also been found that active input during the study phase is far more beneficial to implicit knowledge acquisition, than simple observation (Kelly and Burton, 2001).

It is predicted that the two-second deadline will yield a higher level of selected positive exemplars than the similar condition in Cock et al�s (1994) study, in which similarity dominated with a relaxed response deadline. It is also predicted that the 30 study exemplar condition will provide a significantly higher number of selected positive strings, and restore the original predicted effect of subconscious learning in McGeorge and Burton�s (1990) study, in which lack of control over confounding variables led to only a residual effect. Finally, in accordance with literature to suggest the beneficial effects of practise and exposure on implicit knowledge acquisition, it is predicted that the effect of implicit rule-abstraction will increase with the number of study strings.

Method

Participants

Ninety undergraduate students at Keele University were tested individually. They were acquired through opportunity sampling and were not paid for their participation. Participants were not familiar with the nature of the study. Participants were aged between 18 and 26 (mean=20.2, SD=0.9).

Stimuli

The stimuli were divided into two distinct sections, the study phase in which participants are exposed to the invariant exemplars, and the test phase in which subconscious learning of the presence of an invariant digit is tested.

For the study phase, lists of four-digit numbers (exemplars) were created for each of the three conditions, comprising groups of 10, 30 and 50 numbers. The exemplars were created in accordance with a strict methodological protocol. Each of the numbers contained the invariant digit, in this study the invariant was chosen as 5 in order for counterbalancing of frequency to occur across the arithmetic distraction task. As in McGeorge and Burton�s (1990) study, participants were required to summate the first and last two digits of the four-digit exemplar, and indicate which was highest. For example, in the exemplar 3456, 5+6 is higher than 3+4, and so the right side would be indicated. Although McGeorge and Burton used the invariant digit 3, it was deemed that 5 would be more appropriate as it could evenly fall into the smallest or largest pairing (whereas for example 1 would most often fall into the smaller category). Aside from the invariant, the remaining three digits could not contain a repeat (including the digit 5) or 0 (zero) so as to prevent distinctive exemplars.

Excluding the invariant, the three remaining digits within the exemplars could not be anagrammatical of each other, e.g. 5342 could not be presented in a trial with 2453. Taking these factors into account, there are only 56 possible combinations of three digit numbers (listed in Appendix C), as previously noted in the Cock et al study (1994). The number of combinations appropriate for the experimental condition were selected and placed with the invariant. Selection from the list was systematic for each condition, e.g. every fifth number for the 10 study-exemplar condition. This was to balance the variant digit frequency, e.g. if the first ten combinations were chosen they might have contained an unrepresentative frequency of the digit 1. Following selection, Any imbalance of digit frequency (except of course 5 which was equal to the number of exemplars specific to the condition) was corrected through swapping the three-digit combinations with others from the list of 56. This was important to ensure that the variant digits were seen in equal amounts, to prevent particular digits from interfering with the invariant.

Once digit frequency was relatively even, the four-digit arrays were scrambled so that 5 fell into the first, second, third and forth position of the array with as equal frequency as possible. This was to reduce suspicion of the invariant digit, e.g. if each array began with 5 the invariant rule would surely become explicit. It was also ensured that half of the arrays produced a first pair summation larger than the second, and half a second pair summation larger than the first. In no array was the summation of the first digit pair equal to the summation of the second digit pair.

This process was repeated for all three conditions, producing lists of 10, 30 and 50 four-digit arrays.

For the test phase, pairs of exemplars were created to test for implicit knowledge of the invariant digit through a forced-choice decision paradigm. For each condition, a list of 10 pairs of four-digit exemplars were designed. Within each pair, only one exemplar would contain the invariant digit.

To reduce the likelihood of similarity accounting for the seemingly implicit choice of the invariant digit as suggested by Cock et al. (1994) and Vokey and Brooks (1992), similarity of the test strings to the study strings was manipulated using the same method as that developed by Cock et al. (1994), specifically using the positive-dissimilar, negative-similar paradigm. The exemplar of each test pair not containing the invariant digit (negative) was designed to be as similar as possible to an exemplar in the previously presented study phase.

The negative exemplars were therefore constructed by simply repeating an exemplar from the study phase, but substituting the invariant digit for anything other than 5 or 0, e.g. 3587 in the study phase may be modified as the negative-similar exemplar 3487 in the test phase. The substitutes digits were also counterbalanced to maintain as even digit frequency as possible, and no repeats were made within an array. For the paired four-digit array containing an invariant digit (positive), a new combination was chosen from the list of 56, and a new list of ten exemplars were produced in a similar fashion to the first section, so that they bared minimal similarity to the associated study phase. For the condition containing 50 stimulus arrays, only 6 new three-digit combinations were available (the total limit of combinations being 56). This being the case, 4 combinations from the study phase were presented with the 6 new, but in a scrambled fashion, so as to reduce similarity. Whilst it seems perhaps appropriate to use 4 combinations presented at the start of the study phase to reduce a recency effect of similarity, this was not possible due to the random sequencing of the arrays in both sequences.

The ten pairs of arrays were presented so that half contained the invariant digit on the left array, and half on the right. Appendix A contains the specific exemplars used for each phase of each condition, whilst Appendix B contains example stimuli.

Apparatus

A computer program was used to generate random sequencing and presentation of the four-digit exemplars for the study phase of each experimental condition, either 10, 30 or 50 exemplars. Following this, the program then presented random sequencing of the 10 paired exemplars for the test phase. Although the participant was not timed for the arithmetic distraction task of the study phase, the custom software was programmed to allow only 2 seconds for a response within the forced-choice decision of the test phase.

Simple standardised instruction texts were presented prior to each study and test phase of each condition (see Appendix B). These were constructed using graphics software, presented as black text on white, Arial font size 24, with appropriate paragraphing. The exemplars in both phases for each experimental condition were similarly presented as black Arial 24 on a white background (examples also in Appendix B). All exemplars were presented centrally within the screen. The stimuli were presented on 13� monitors, at 72 dpi. Responses were made using a standard QWERTY configured keyboard, with Z appropriately on the left and M on the right used to indicate the side of highest summation for the distraction task of the study phase. The participant was comfortably seated approximately 600mm from the screen.

Design

Participants were randomly allocated to one of the three experimental conditions. Within each condition, thirty participants each performed a single trial. The independent variable was the number of study strings presented within the study phase, either 10, 30 or 50. The dependant variable was the number of positive test strings chosen, out of ten, during the test phase. There were no significant differences between age, sex, and length of time allowed to complete the test phase across conditions. The same apparatus, program and set-up were similarly used by all participants.

Procedure

Participants were tested individually and all study strings and test pairs were presented in a randomised order to each participant.

The participant was informed verbally by the experimenter as to what was required through a simple set of standardised instructions. These were then repeated through presentation on the monitor. For the study phase, this read �You are about to see 4-digit numbers. Your task is to add the 1^st two and last two digits, and indicate which is highest by pressing Z for left, and M for right. For example: 2751 � press Z (as 2+7 is higher than 5+1). 5289 � press M. There are [10/30/50] numbers, [so/but] it won�t take long. Press any key to start�� (see Appendix B for further details of the instructions). It was verified that the participant understood what was required of them before they commenced the first phase of exemplars. The exemplars were presented one at a time. Once an appropriate response was made according to the arithmetic distraction task, a 500ms interval consisting of a blank white screen was presented before the next exemplar appeared in the centre.

When the final exemplar in the study phase was responded to, the participant was thanked and then informed as to what was required for the second, surprise test phase of the trial. The accompanying instruction screen similarly read �Thank you, it�s almost over� For the final section, you will now see 10 pairs of 4-digit numbers, side by side. One will be a repeat of a number you just saw in the previous sequence, the other will be new. Please indicate which of the pair is most likely a repeat, pressing Z for left, and M for right. You have only 2 seconds to make a decision, but be sure to view each number before doing so. Press any key to start�. Again comprehension was verified before continuing. It was made explicit that responses must be rapid, that both strings should be viewed before making a decision, and to guess if unsure (as this is likely to be the case for most if not all test strings, due to the deceptive nature of the design in that all test strings were actually new). Failing to make a response within two seconds as asked by the instructions, the participant was presented with a screen flashing �You MUST make a response within 2 seconds!�. Data from any participant failing to respond within 2 seconds on more than two occasions was excluded from the study. Extra time for responding would allow for explicitly contrived decisions and would therefore not be measuring the immediate response based on a more implicitly processed knowledge. Once the 10 test strings had been completed, a final screen was presented thanking the participant, and this was again appreciated verbally. Responses from both phases were recorded by the program.

Finally, each participant�s awareness of the invariant rule was examined using a short semi-structured questionnaire, based on the following questions:

Did you notice anything unusual about the numbers presented in the first sequence?
If I was to tell you that a single recurring digit was present within all of the numbers shown in the first sequence, would you know what that digit was?

The participant�s data was excluded on the following counts:

If they answered yes to question 1, and demonstrated explicit knowledge of the invariant rule.
If they correctly answered 5 as the recurring digit, whether guessed or otherwise. (The last two situations led to exclusion despite Cock et al�s claim that post hoc reporting of invariant digit awareness did not significantly affect performance. Whilst participants aware of the rule may only see it as a peculiar commonality, such knowledge might still inhibit the same knowledge acquired implicitly, and so confound the validity of the trial as a measurement of unconsciously guided decision making. Also, it is acknowledged that the second situation logically excludes approximately one in nine participants based on the likelihood of simply guessing the digit 5).

If they gave an explicit rule on which they were operating during the test phase, whether based on the invariant digit or not. For example, reporting a selection based on probabilistic sequencing (e.g. left, right, left etc.), or choosing exemplars based on larger numbers.
If during the test they performed in a manner which was not measuring implicit awareness, e.g. closing eyes during test phase, operating an obvious left right left right response to the test phase, not responding within 2 seconds on more than two occasions.

All participants were then debriefed, and made aware of the invariant rule and true nature of the experiment. This was appropriate due to the deceptive nature of the experiment. If requested when prompted, participants were made aware of their �score� through knowledge of the mistakes they made during the study phase, and number of positive exemplars chosen in the test phase. Participants were kindly asked to not disclose the invariant rule to anyone in the unlikely event that a friend may subsequently participate in the experiment.

Results

Overall, 30 participant�s results were discarded, 11 for response behaviour and 19 for invariant-digit awareness. There were 8 counts of failing to respond to at least 8 out of 10 test strings, and 3 counts of inappropriate response behaviour (displaying obvious explicit strategies for random response). Nineteen participants reported the invariant rule, whether explicitly or guessing the digit when made aware of an invariant digit by the experimenter. Incidentally, the frequency of invariant digit awareness, explicit or guessed, was 3, 6 and 10 for the 10, 30 and 50 study exemplars respectively. No participants expressed awareness to the novelty of the test strings. Data collection continued until 30 participants within each condition had responded to at least 8 test pairs and failed to report explicit awareness of the invariant rule, including incorrect guessing of the invariant digit. The results are shown in Table 1, and Figure 1. Raw data are in Appendix A.

Table 1: Mean no. of positive invariant exemplars chosen (out of 10)

Condition (No. of study strings)	A (10, N=30)	B (30, N=30)	C (50, N=30)
Mean (St Dev) invariants chosen	4* (1.17)	5.95* (0.98)	5.63** (1.61)

* p<0.001.

** p<0.05.

Figure 1: Mean number positives chosen at test. The largest amount of positives were chosen after 30 study exemplars, the lowest with 10.

For the 10 study exemplars condition (A), the mean number of positive strings chosen at test was 4, out of a maximum of 10, with a standard deviation of 1.17. In the 30 study exemplars condition (B), the mean of positive strings chosen at test was 5.67, with a particularly robust standard deviation of 0.98. The 50 study exemplars condition (C) gave a mean of 5.63 positive strings chosen at test, with a standard deviation of 1.61. Figure 2 shows the frequency of chosen positive exemplars for the three conditions.

Figure 1: Frequency of positives chosen at test. It is clear that condition B (30 study exemplars) yielded the smallest variance of selection. Condition C (50 study exemplars) had the largest variance, which interestingly was split into two distinct sections.

T-test analysis found condition A to be significantly below chance level (t=-4.66, df=29, p<0.01), whilst both conditions B and C were significantly above chance level (t=5.21, df=29, p<0.01, t=2.16, df=29, p=0.039, respectively).

A one-way anova found the between groups variance to be significantly different (f=20.682, df=2, p<0.001). Tukey�s post-hoc analysis was used to test the means for individual differences between conditions. There was no significant difference between conditions B and C (p>0.05). However, both of these conditions gave a significantly higher number of chosen positive exemplars than condition A (B>A: MD=-1.93, p<0.01; C>A: MD=-1.67 p<0.01, both one-tailed).

Discussion

The present study aimed to investigate implicit acquisition of knowledge in the form of rule-abstraction, using the paradigm of invariance learning. In an attempt to control for such confounds as similarity (Cock et al., 1994) and distinctive strings (Wright and Burton, 1995), the experiment aimed to provide an optimally valid measurement of implicit knowledge. This was achieved through replicating the positive, dissimilar vs. negative, similar (PDNS) structure of Cock et al�s (1994) similarity analysis, which found a significant bias towards similarity, against the invariant rule. No repetitions were used, to control for distinctive strings. A two-second deadline was also introduced to the paradigm to prevent relaxed explicit processing of the strings, which would encourage such conscious processes based on episodic knowledge. Finally, numbers of study strings were manipulated to investigate effect of exposure on subconscious learning.

When 10 study strings were used, a mean number of 4.00 positive strings were chosen at test, out of a maximum of ten. When study strings were increased to 30, the number of positive strings selected rose to a 5.95, which was significantly above chance. Strangely, when the number of study strings was increased to 50, the number of positive exemplars fell again to 5.67, but was still above chance level.

Congruence with invariance literature

The findings were interesting in that the condition most closely matching a condition within Cock et al�s (1994) study yielded very different, essentially opposite, results. Cock et al found a mean of 4.06 positives chosen for the PDNS condition after 30 study exemplars. As this is significantly below chance, it strongly suggests a dominance of similarity over the presence of an invariant digit as an influence to recognition performance. However in the present study, 30 study exemplars led to a mean of 5.95 positive, dissimilar strings chosen at test phase, with 4.05 negative, similar strings chosen. The standard deviation was similarly robust (0.98, compared to 1.06 for Cock et al.). Interestingly, a similar finding to Cock et al. was found when only 10 exemplars were used at study, with a mean 4.00 positive strings chosen at test phase.

The 30 study exemplar condition was also similar to the Experiment 1 of Wright and Burton�s (1995) testing for the effect of absent repetitions, in that no repetitions were used in any of the strings. Again, significantly higher results were found (mean of 5.95 positives chosen compared to 5.00). Interestingly however, concordance was found between the present findings and those in Experiment 2 of Wright and Burton�s within measures manipulation of repetition. For the neutral condition where repetition was absent, 5.9 positives were chosen. These figures were unfortunately found to be at chance level however, due to the small proportion of such test strings in the within measures design. The unaccounted for positive selection in Wright and Burton�s study indicates that performance based on implicitly acquired knowledge may have been inhibited by the insensitive nature of the test phase.

The difference between results within subconscious learning experiments points to the highly sensitive nature of the methodology used. Kelly et al (2001) demonstrated the importance of measurement sensitivity, with recall and recognition based methods respectively influencing performance based on explicit and implicit knowledge. It has been demonstrated that even type font has affected apparent implicit acquisition of knowledge (Jacoby & Hayman, 1987). Shanks and St. John (1994) acknowledge the inherent invalidity of subconscious learning paradigms at measuring the same knowledge as that which is actually used in the performance of such tasks. As participants failed to guess the invariant digit when made aware of its presence, the present study conforms to the definition of subconscious learning as proposed by Berry and Dienes (1993).

The effect of a response deadline

The success of the pressure to respond within 2 seconds was obviously of major importance in this experiment, and may account for the disparity between the present study and both Cock et al�s and Wright and Burton�s. This is more apparent considering the significant bias towards the invariant rule at response, despite a failure to report any strategies or the presence of the invariant rule, and the negative strings bearing more similarity than the positives to the study strings.

Cock et al reported that participants had reported a reliance on �memory� or �familiarity� (Cock et al 1994, p1026) when asked of any strategies used during the test phase. This supports the claim of the present study to have prevented the dominant influence of explicit rules. Participants did not report any strategies of decision making within this short interval, and explained choices purely on random estimation. The results show however that the results were not random, being significantly different from chance when 10, 30 or 50 exemplars were used.

If explicit processes were inhibited as hoped for by the rapid response, it is argued that subconscious learning of the hidden rule was able to influence the two-alternative forced choice response, over any potential explicit processes such as similarity matching or distinctive rejection. Hay and Jacoby (1996) similarly found that habitual processes based on implicitly acquired knowledge were unaffected by a response deadline in probability matching. Masters (1992) supports the notion with the performance of an implicitly learned skill seemingly unaffected by additional pressures. Both of these studies however reported a decreased if not absent level of explicit strategy in the face of pressure. Meta analyses of literature have concluded empirical support for the existence of separate mechanisms for implicit and explicit information (Lewicki, Hill and Czyzewska, 1992). Amnesia studies suggest that these mechanisms are also physiologically differentiated (Ramus and Squire, 1992), and the interaction between the two is currently unknown (Cleeremans et al., 1998). Lewicki et al. (1992) argue that nonconscious information-acquisition processes are incomparably faster and structurally more sophisticated than their conscious counterparts. This may account for the efficient performance of implicitly acquired knowledge when circumstances inhibit the slower, less-sophisticated consciously processed knowledge. Schmidt (1982) succinctly captures this idea with the notion of a pianist, whom when asked to concentrate on describing the actions of the fingers, suffers a decrease in performance. Applied to the previous studies of invariant learning, it could be that the nature of the test phase inhibited any implicit processes through effortful explicit thought, which may then utilise overriding strategies such as similarity matching or string rejection. This was apparent in the present study only in the lack of sufficient exposure to the invariant rule. When only 10 study exemplars were used, similarity was still the main influence of performance. This suggests that similarity may still occur as an implicit based process, in the absence of verbal awareness, despite popular belief otherwise (e.g. Perruchet, 1994, Shanks & St. John, 1994). The implications also place a dominance of explicit processes on performance in the absence of implicit knowledge. The interaction between implicit and explicit knowledge based performance therefore seems to depend on a complex matrix of variables, which undoubtedly include exposure to learning material and pressure to perform.

The applications for efficient performance under pressure when based in implicitly acquired knowledge are vast, and should be further explored across various paradigms including real world settings.

The effects of exposure

The effects of exposure were also significantly apparent. The results of Cock et al�s study using 30 exemplars were only matched in the present study in the condition using 10 study exemplars. It is possible that the small exposure to the hidden rule in this condition was insufficient to allow implicit rule-abstraction, and so similarity matching prevailed in the absence of any implicit knowledge. However when 30 and 50 study exemplars were used, significantly more responses were made towards the positive exemplars, suggesting implicit acquisition of the invariant rule. This is consistent with claims that richer and more abstract knowledge is acquired when a longer study phase is used (Mathews, 1990). The effects of increased exposure also account for the implicit acquisition of cultural regularities in a real world setting (Kelly et al., 2001).

It is interesting that performance was not improved when the study exemplars were increased to 50, as might be predicted following the significant improvement between 10 and 30 study strings. The finding supports a previous study by Ward and Churchill (1998) in which the training set was also extended in an invariant task. Using both the presence of an invariant digit 3, and the absence of a 3 as two distinct invariances, Ward and Churchill found that subconscious learning of both conditions increased, despite rule-abstraction hypotheses predicting better performance for the presences of a 3 than the absence. When the training set was extended, it was found that performance in the No 3�s condition also decreased. The authors were unable to explain this finding, although it seems to be relatively robust in light of the present findings. However, as there was no significant difference between performance after 30 or 50 study strings, the decrease must be accounted for by chance.

There are a number of possible reasons as to why neither implicit rule abstraction or test string similarity significantly improved performance above that found for 30 study strings. It is likely that the sheer size of the study phase exemplar list inhibited performance based on implicit knowledge. With so many exemplars, a great level of metacognitive episodic interference is likely to have occurred at memory encoding, making a recognition task seem totally beyond the conscious ability of the participant. Despite the presence of an invariant digit within all of the study exemplars, subconscious learning is likely to have been overridden by a conscious lack of confidence and feeling of despair in the participant, increased by the pressure of a two-second deadline. Responses may have then been made with a consciously random sequence, with little allowance for the role of unconscious processes to guide decision making. This is supported by exclusion of three participants for inappropriate response behaviour within the 50 study exemplar condition, the only three such exclusions in the entire experiment, as well as many reports of despair during the test phases of included participants. Two participants made immediate decisions without first viewing both exemplars, following an obvious left-right-left counterbalanced sequence (incidentally, the chances of the program randomly generating a test phase in such a sequence is 1 in 500). The third closed their eyes throughout the test phase as a method of random response, which is clearly an invalid measurement of memory-based decision making.

These suggestions are in accordance with the findings of Masters (1992) and Bright and Freedman (1998), who found that implicit skill acquisition (but not performance) was inhibited in situations of pressure, by conscious effort and explicit processing. Despite the short deadline it is also likely that the influence of similarity was inhibited on account of the study exemplars. As the negative test strings are designed to share similarity with only 1 out of the 50 study exemplars, it is unlikely that this information was extracted at recognition during the two-second deadlines of the test phase. Therefore, it is speculated that implicitly acquired knowledge of the invariant rule was inhibited by the pressure of the two-second deadline, demonstrated by the reported lack of confidence of the participants in their ability to perform the required task.

Limitations of the invariance paradigm

Additional findings included that out of the two most popularly chosen positive exemplars for each of the three conditions, 5 out of 6 were the lowest number of the pair for the particular string (see Appendix A). For example, the positive exemplars were chosen in the test strings 3856 7642 and 9873 2579, which were also the lowest numbers when seen as a whole. Again, this is interestingly the reverse of that found in Cock et al�s study, where participant�s reports of test phase response strategy were varied, but included a simple unexplained preference for larger numbers. It is speculated that a preference for low numbers at test phase may develop through simple classical conditioning. As the participant is likely to find sums involving low numbers easier to solve, they may develop a preference for such numbers during the study phase, providing a perfectly logical test phase strategy when under pressure.

This is consistent with previously outlined explicit strategies available to participants in such studies (e.g. McGeorge and Burton, 1990) and suggests that the two-second deadline may not be successful at preventing explicit processing per se, but only the more influential processes. It is of course also possible that the effect was due to chance, and this is further likely following that the effect was largely inconsistent below the two most popular. It is still worth considering however in further research that the effect is counterbalanced across the test string variables, or the range of the pairs first digits minimised, e.g. 3546 4273. The popular exemplar characteristics may be further studied using the raw data in Appendix A.

Two further potential methodological complications were found within the invariance learning paradigm, both shared with most other similar experiments. Firstly, the nature of the verbal tests used to measure explicit knowledge of the hidden rule may invalidate the measurement of performance as sensitive to implicitly acquired knowledge of the invariant rule. Following debriefing that an invariant digit was present within all study strings, participants are typically asked to guess which digit it was (e.g. Cock et al., 1994, Wright and Burton, 1995). This may be seen as a forced choice measure not entirely dissimilar from that of the test phase, in which participants are asked to provide knowledge to a best-fit, based on exposure to a series of study strings conforming to a hidden rule. Wright and Burton even presented the participants with a list of digits, asking them to circle the guessed (at least explicitly) invariant, excluding results of those who guessed correctly. Is it not possible that implicit knowledge leading to invariant digit selection at test would not also manifest in the forced choice test of explicit awareness? Indeed, there is very little separating the tests of implicit and explicit awareness. However, the surprising lack of correct answers to this debriefing question despite significant implicit performance at test only goes to suggest a further distinction between motor and articulatory processes, the former being more related to implicitly acquired knowledge. This is opposed to the common belief that articulatory and task-related processes have access to a shared database (e.g. Hayes and Broadbent, 1988).

Secondly, it may be argued that the effect of an ever present invariant digit is weakened during the test phase. As participants view both alternatives, they inevitably view increasing numbers of exemplars not containing the invariant digit (the negative exemplars. For example, consider the condition with only 10 study strings. By the tenth test exemplar, participants overall will have seen 19 positive and 9 negative strings. Therefore, only 32% of exemplars contained the invariant digit. With so few study strings, this will invariably dilute the effect of invariance learning. In fact, even after the first test string, the invariant digit as a �rule� is abolished, and therefore all subsequent test strings are measuring a different level of acquired knowledge. It is proposed in light of this inherent paradox that further studies measure performance using only a single test string, so that the invariant digit as a rule is preserved. Of course, vast numbers of participants will be necessary, making the procedure all the more elusive.

Some further minor methodological limitations of the study may include the construction of test exemplars. The study exemplars were constructed separately, each using a systematic selection of variant digits from the list of 56 possibilities. As the test exemplars were then constructed in accordance with the study exemplars, the test exemplars were, like the study exemplars, the same within conditions but different between conditions. This is unlikely to affect the results however due to the equal digit frequencies and random presentation of strings. Care was also taken to prevent any distinctive exemplars, not only in the form of repeated digits, but also anything starting in 19 which as a year may have semantic value.

The similarity manipulation, which was controlled using a simple method not dissimilar to that developed by Cock et al. (1994), is not empirically tested and may hold certain invalidities with regards to preventing post-hoc similarity matching, also acknowledged by Cock et al (1994). Similarity matching may be based on far more complicated processes possibly including shapes within certain groups of digits, relationships between adjacent digits based on range, memory of odd and even numbers and other more abstract features. However as the same method was used in both studies, it is controlled for across the different experiments. The disparity between results can therefore be attributed to other factors.

Conclusions

The effects of increased exposure and a response deadline were shown to increase the level of performance based on implicitly acquired knowledge of an invariant rule. It is argued that increased exposure strengthened the implicit rule-abstraction, whilst the response deadline inhibited the confounding influence of explicit strategy common within invariance research.

The effect of the response deadline as a possible method of inhibiting explicit knowledge is a powerful technique in the search for support of subconscious learning, and should be further explored, possibly involving transference to alternate modalities or paradigms. It is known that subconscious learning effects are vulnerable to many slight superficial variances, which may also account for some of the disparity between similar studies.

As subconscious learning has been shown to be robust in the face of aging and amnesia of declarative memory (e.g. Kotchoubey et al., 2000, Graf, Mandler & Haden, 1982, Graf, Squire and Mandler, 1984, Knowlton and Squire, 1993, Knowlton, Squire and Gluck, 1994), the results hold great applicable value in suggesting a potential method of learning procedural information which may of benefit to the elderly or amnesic. It is for these reasons that further replication and study in this area is encouraged to further understand the complex variables which affect subconscious learning, and to what extent.

References

Berry, D.C., & Broadbent, D.E. (1984). On the relationship between task performance

and associated verbal knowledge. The Quarterly Journal of Experimental

Psychology, 36 (A), 209-231.

Berry, D.C. & Dienes, Z. (1993). Subconscious learning: Theoretical and empirical issues.

Hove: Lawrence Erlbaum.

Bright, J.E.H. & Burton, A.M. (1994). Past midnight: semantic processing in an

subconscious learning task. Quarterly Journal of Experimental Psychology, 47 (A), 71-90.

Bright, J.E.H., & Freedman, O. (1998). Differences between implicit and explicit

acquisition of a complex motor skill under pressure: An examination of some

evidence. British Journal of Psychology, 89, 249-263.

Brooks, L.R., & Vokey, J.R. (1991). Abstract analogies and abstracted grammars:

comments on Reber (1989) and Mathews et al. (1989). Journal of

Experimental Psychology: General, 120, 316-323.

Cleeremans, A., Destrebecqz, A., & Boyer, M. (1998). Subconscious learning: news from

the front. Trends in Cognitive Sciences, 2 (10), 406-416.

Cock, J.J., Berry, D.C. & Gaffan, E.A. (1994). New strings for old: the role of

similarity processing in an incidental learning tasks. The Quarterly Journal of

Experimental Psychology, 47A (4), 1015-1034.

Dienes, Z., Broadbent, D., & Berry, D. (1991). Implicit and explicit knowledge bases

in artificial grammar learning. Journal of Experimental Psychology: Learning,

Memory and Cognition, 17 (5), 875-887.

Dulany, D.E., Carlson, R.A., & Dewey, G.I. (1984). A case of syntactical learning and

judgement: how conscious and how abstract? Journal of Experimental

Psychology: General, 113, 541-555.

Graf, P., Mandler, G., & Haden, P. E. (1982). Simulating amnesic symptoms in

normal subjects. Science, 218, 1243-1244.

Graf, P., Squire, L.R., & Mandler, G. (1984). The information that amnesic patients

do not forget. Journal of Experimental Psychology: Learning, Memory and Cognition, 10 (1), 164-178.

Hay, J.F. & Jacoby, L.L. (1996). Separating habit and recollection: memory slips,

process dissociations, and probability matching. Journal of Experimental

Psychology: Learning, Memory, and Cognition, 22 (6), 1323-1335.

Homa, D. Sterling, S., & Trepel, L. (1981). Limitations of expemplar-based

generalisation and the abstraction of categorical information. Journal of

Experimental Psychology: Human Learning and Memory, 7 (6), 418-439.

Jacoby, L. L., & Hayman, G. (1987). Specific visual transfer in word identification.

Journal of Experimental Psychology: Learning, Memory, and Cognition, 13,

456-463.

Johnstone, T., & Shanks, D.R. (2001). Abstractionist and processing accounts of

subconscious learning. Cognitive Psychology, 42 (1), 61-112.

Jones, G.V. (1990). Misremembering a familiar object: When left is not right.

Memory and Cognition, 18, 174-182.

Kelly, S.W., & Burton, A.M. (2001). Learning complex sequences: no role for

observation? Psychological Research, 65 (1), 15-23.

Kelly, S.W., Burton, A.M., Kato, T. & Akamatsu, S. (2001). Incidental learning of

real-world regularities. Psychological Science, 12 (1), 86-89.

Knowlton, B.J., Ramus, S.J., & Squire, L.R. (1992). Intact artificial grammar learning

in amnesia: dissociation of classification learning and explicit memory for

specific instances. Psychological Science, 3, 172-179.

Knowlton, B.J., & Squire, L.R. (1993). The learning of categories: parallel brain

systems for item memory and category knowledge. Science, 262, 1747-1749.

Knowlton, B.J., Squire, L.R., & Gluck, M.A. (1994). Probabilistic classification

learning in amnesia. Learning and Memory, 1, 106-120.

Kotchoubey, B., Haisst, S., Daum, I., Schugens, M., & Birbaumer, N. (1999).

Learning and self-regulation of slow cortical potentials in older adults.

Experimental Aging Research, 26 (1), 15-35.

Lewicki, P., Hoffman, H., & Czyzewska, M. (1987). Unconscious acquisition of

complex procedural knowledge. Journal of Experimental Psychology, 13 (4),

523-530.

Lewicki, P., Hill, T., & Bizot, E. (1988). Acquisition of procedural knowledge about a

pattern of stimuli that cannot be articulated. Cognitive Psychology, 20, 24-37.

Lewicki, P., Hill, T., & Czyzewska, M. (1992). Nonconscious acquisition of

information. American Psychologist, 47 (6), 796-801.

Masters, R.S.W. (1992). Knowledge, knerves and know-how: the role of explicit

versus implicit knowledge in the breakdown of a complex motor skill under

pressure. British Journal of Psychology, 83, 343-358.

Mathews, R.C. (1990). Abstractness of implicit grammar knowledge: Comments on

Perruchet and Pacteau�s amalysis of synthetic grammar learning. Journal of

Experimental Psychology: General, 119, 412-416.

Mathews, R.C., Buss, R.R., Stanley, W.B., Blanchard-Fields, F., Cho, J., & Druhan,

B. (1989). The role of implicit and explicit processes in learning from

examples: A synergistic effect. Journal of Experimental Psychology: Learning, Memory and Cognition, 15 (6), 1083-1100.

McAndrews, M. & Moscovitch, M. (1985). Rule-based and exemplar based class-

ification in artificial grammar learning. Memory and Cognition, 13, 469-475.

McGeorge, P., & Burton, A.M. (1990). Semantic processing in an incidental learning

task. Quarterly Journal of Experimental Psychology: Human Experimental

Psychology, 42, 597-609.

Morton, J. (1967). A singular lack of incidental learning. Nature, 215, 203-204.

Nisbett, R.E., & Wilson, T.D. (1977). Telling more than we can do: verbal reports on

mental processes. Psychological Review, 84, 231-259.

Perruchet, P., & Pacteau, C. (1991). Implicit acquisition of abstract knowledge about

artificial grammar: Some methodological and conceptual issues. Journal of

Experimental Psychology: General, 120 (1), 112-116.

Rathus, J.H., Reber, A.S., Manza, L., & Kushner, M. (1994). Subconscious learning and

explicit learning: Differential effects of affective states. Perceptual and Motor

Skills, 79 (1), 163-184.

Reber, A.S. (1993). Subconscious learning and tacit knowledge. Journal of Experimental

Psychology: General, 118, 219-235.

Reber, A.S., & Lewis, S. (1977). Subconscious learning: An analysis of the form and

structure of a body of tacit knowledge. Cognition, 5, 333-361.

Sanderson, P.M. (1989). Verbalizable knowledge and skilled task performance:

Association, dissociation and mental models. Journal of Experimental

Psychology: Learning, Memory and Cognition, 15 (4), 729-747.

Schmidt, R.A. (1982). Motor Control and Learning: A Behavioural Emphasis.

Illinois: Human Kinetics.

Shanks, D. R., & St John, M. F. (1994). Characteristics of dissociable human learning

systems. Behavioral and Brain Sciences, 17 (3), 367-447.

Stadler, M.A., Warren, J.L., & Lesch, S.L. (2000). Is there cross-format transfer in

implicit invariance learning? Quarterly Journal of Experimental Psychology:

Human Experimental Psychology, 53 (1), 235-245.

Vinter, A., & Perruchet, P. (2000). Subconscious learning in children is not related to age:

Evidence from drawing behaviour. Child Development, 71 (5), 1223-1240.

Vokey, J.R. & Brooks, L.R. (1992). Salience of item knowledge in learning artificial

grammars. Journal of Experimental Psychology: Learning, Memory and

Cognition, 18, 328-344.

Ward, G., & Churchill, E.F. (1998).Two tests of instance-based and abstract rule-

based accounts of invariant learning. Acta Psychologica, 99 (3), 235-253.

Willingham, D.B., Nissen, M.J., & Bullemer, P. (1989). On the development of

procedural knowledge. Journal of Experimental Psychology: Learning,

Memory, and Cognition, 15, 1047-1060.

Wright, R.L., & Burton, A.M. (1995). Subconscious learning of an invariant: Just say no.

Quarterly Journal of Experimental Psychology, 48A (3), 783-796.