Evidence for dissociable learning processes from the SRT task

Abstract

Current evidence suggests that it may be fruitful to model learning using dissociable rule-based and associative processes. Such an account would predict that people should be able to learn the same sequences qualitatively differently under hypothesis-testing and incidental conditions. The aim of this research was to test this prediction in a two choice, serial reaction time (SRT) task setting. All the SRT experiments conformed to the following general design. An Experimental Group was trained on sequences constructed by concatenating the subsequences XXX, XYY, YYX and YXY, in a random order (where X and Y were the two trial types in the two-choice task). In addition, a Control Group was trained on a pseudo-random ordering for the same number of trials. Both groups were subsequently tested on a pseudo-random ordering, to ensure that any group differences could be attributed to sequence learning, as opposed to 'sequential effects'. The extent to which the Experimental Group's responding on the third trials differed from Control was used as an index of sequence learning. When the SRT task was carried out in incidental conditions and with an RSI of 500ms, the Experimental Group showed evidence of having learnt those subsequences that ended in an alternation (XXY and YXY) but not those that ended in a repetition (XXX and XYY). In contrast, under hypothesis-testing conditions subjects performed relatively better on those subsequences that began in a repetition, probably because they were expressing knowledge related to longer strings of Xs than Ys. Furthermore, when the contingencies were made less noisy by marking out the triplets in a second hypothesis-testing condition, the pattern of subsequence learning again qualitatively differed from that in the incidental condition, with subjects now learning XXX earliest. These dissociations are consistent with the operation of separable learning processes; with the hypothesis-testing patterns being attributable to a rule-based process that finds runs of repetitions especially salient, and the incidental pattern being explicable in associative terms, by the APECS SRN (but not by other contemporary associative models). According to this account, sequential effects and sequence learning had largely additive effects on the Experimental Groups' performance. This assumption was tested by replicating two of the previous conditions at a new RSI of 50ms, and so with a changed pattern of sequential effects; the idea being that an unaltered pattern of sequence learning would suggest additivity. However, the results were ambiguous; though they were not inconsistent with the above account. In conclusion, in the absence of an adequate single-process alternative, the experiments reported in this thesis support the rule-based/associative distinction.