Modeling a reaction time variant of the Perruchet effect in humans

This paper presents a reaction time (RT) experiment that follows on from the work of Perruchet, Cleeremans, and Destrebeceqz (2006), investigating the extent to which reaction times (RTs) are governed by the conscious expectancy of a particular response. In this experiment, participants were presented with a single stimulus (which we will call the conditioned stimulus; CS) followed by one of two outcomes (which we will call unconditioned stimuli; USs); to which participants had to make an appropriate instrumental response. On every trial we recorded the time taken to make this response and participants were asked to rate their expectancy that one of the USs (US1) was going to occur. We found that the expectancy rating for US1 correlated negatively with RT on US1 trials. Over successive runs of reinforcement, when participants rated US1 as less likely to occur they were slower to respond to US1 (lower ratings, higher RTs). When we calculated the expectancy for US2 as the complement of that for US1 expectancy, expectancy of US2 correlated positively with RTs. Thus, across runs of reinforcement, participants responded more quickly to US2 when considering US2 less likely (low rating, low RT). We argue that the requirement to make a conscious expectancy rating results in participants attending more to US1 occurrences than those of US2. This results in a qualitatively different relationship between conscious expectancy and automatic responses that cannot be reconciled by a single processing system account. A dual processing system explanation of learning is proposed to explain these results. In support of this position, we successfully modeled our US2 RT data using a modified version of the Augmented simple recurrent network (Yeates, Jones, Wills, McLaren, & McLaren, 2013).