Despite a large number of studies which have dissociated task performance from verbalisation of knowledge, the nature of learning in invariant learning paradigms remains contentious. Of particular concern is whether the learned knowledge is explicit or implicit, and whether the learned knowledge is instance-based or based on abstract rules.

This article focuses on the latter of these two concerns. Specifically, we present two experimental manipulations which we argue can distinguish between instance-based and abstract rule-based learning in invariant learning tasks. We believe this to be an interesting and productive proposal since there are currently few, if any, techniques for testing the contributions of instance-based knowledge or abstract rule-based knowledge to performance on implicit learning tasks.

The invariant learning task we used to examine our two experimental manipulations was the McGeorge and Burton (1990) invariant digit task. The task consists of two stages: a learning stage and a test stage. In the learning stage, subjects are required to perform a mental arithmetic orienting task on thirty strings of 4-digit numbers. Unbeknownst to the subjects, the learning set stimuli all conform to a hidden rule: all the stimuli contain an invariant feature, the digit “3”. After learning, subjects perform an unexpected “recognition” test, in which subjects make a series of ten forced-choice decisions between pairs of new strings which they had not encountered. They are led to believe that one item of each pair was an item from the learning set. In each test pair, one test item conforms to the hidden rule (a positive instance) and one test item does not (a negative instance). McGeorge and Burton (1990) found that a mean of 7.0 positive instances (max=10.0) were selected in the recognition test, a value which was significantly above chance (chance=5.0). In addition, no subject reported the common feature of the learning exemplars, despite the fact that, in principle, the regularity was very easy to report.

The invariant digit task has the great advantages for our purposes that abstract, rule-based and instance-based accounts of invariant learning on the task have been proposed, and it has proven difficult to distinguish satisfactorily between these accounts. In addition, the methodology is simple and quick, the stimulus set is well defined, the experimental results are reliable, and the hidden invariant is easily verbalisable. Three accounts of the selection bias (one abstract rule-based account and two instance-based accounts) are presented below.

According to the abstract rule learning hypothesis, subjects abstract the hidden consistency from the learning set through exposure to successive instances. The abstract rule links stimuli which contain a 3 with inclusion in the original learning set. At test, this rule is applied to the two test instances. Application of the rule will result in more positives being selected than negatives, because the positive instances always contains a 3, and the negative instances always do not.

Evidence in favour of abstract rule-learning has been obtained by Bright and Burton (1994), using a similar rationale and methodology to that used by McGeorge and Burton (1990). Subjects were presented with thirty pictures of analogue clocks which varied in design, but all shared one hidden, invariant property: upon each clock in the learning set was printed a time between 6 o’clock and 12 o’clock. The subjects were requested to note down the time on each clock.  After all the stimuli had been presented, Bright and Burton (1994) asked the subjects to perform an unexpected recognition test. The test consisted of ten pairs of clocks, none of which had been presented previously. One of the alternatives of each pair conformed to the hidden consistency (a positive instance), the other alternative did not (a negative instance). Subjects were requested to select the clock face that they thought they had previously seen, guessing if they were unsure. Immediately following the recognition test, the subjects were asked a set of graded questions to discover whether the subjects could report any common property of the clocks in the stimulus set. Bright and Burton (1994) found that significantly more positive instances than would be expected by chance were selected in the recognition test, despite the fact that no subject was able to verbally recall the invariant feature. This bias persisted using a different orienting task, and when the perceptual features of the clocks at test varied from the learning stage. Interestingly, no significant differences were found between the selection of positive instances which had not been in the learning set and original learning set items. Bright and Burton (1994) concluded that the “nature of the processing in this task seems to be best described as implicit, abstractive, and influenced by the type of task performed during learning”.