Human sequence learning under incidental and intentional conditions.

Abstract

This research explored the role that dissociable associative learning and hypothesis-testing processes may play in human sequence learning. Two two-choice SRT tasks were conducted, one under incidental conditions and the other under intentional conditions. In both cases an experimental group was trained on four sub-sequences (i.e. XXX, XYY, YYX and YXY). To control for sequential-effects, sequence learning was assayed by comparing their performance to a control group that had been trained on a pseudo-random ordering, during a test-phase in which both groups experienced effectively the same trial-order. Under incidental conditions participants demonstrated learning of the sub-sequences that ended in an alternation, but not of those that ended in a repetition. In contrast, under intentional conditions XXX showed the greatest evidence of learning. This dissociation is explained using a two-process model of learning, with an associative process (the Augmented SRN) capturing the incidental pattern, and a rule-based process explaining the advantage for XXX under intentional conditions. 3 ACKNOWLEDGEMENTS

Modeling human sequence learning under incidental conditions.

“…Augmented SRN to produce our pattern of results in this experiment and still generate the 284 pattern of results found by Jones and McLaren (2009). A typical set of simulation results for 285 the current experiment is shown in the bottom panels of Figure 3.…”

“…This is because the sub-sequences XXX and YYY can be expected 49 to be very difficult to learn based on these earlier findings, and both these sub-sequences fall 50 in Group Same. Intriguingly, when we ran the Augmented SRN on this new experiment with 51 the same parameters as those used to model Jones and McLaren (2009), the pattern we 52 obtained was actually the reverse, with Group Same sub-sequences learnt better than Group 53 Different sub-sequences. Thus evidence-based intuition and the model seem to be in conflict, 54…”

“…This is the variant of the two-choice serial 14 reaction time (SRT) task recently developed by Jones and McLaren (2009). In this task 15 participants observe two circle outlines on a screen and are given two response keys, one for 16 each circle.…”

“…These 290 parameters modulate the relative weightings of the contributions to learning from the 291 different inputs and the context units, and were deliberately chosen to allow us to simulate 292 our data. The learning rate parameters for the fast and slow weights were set to 0.5 and 0.2 293 respectively, and other parameters were the same as those used in Jones and McLaren (2009). 294 different conditions / groups in their appropriate relative positions.…”

“…Is its success 316 simply a consequence of greater flexibility in fitting the data contingent on this increase in 317 free parameters? We believe this is not the right explanation, because when we simulated the 318 Jones and McLaren (2009) data with the new model we did not vary the parameters at all, 319 unlikely to represent "overfitting" of the data. Instead, we believe that the inclusion of the 321 units corresponding to the circle stimulus on the current trial is the critical feature making the 322 difference, and if we take this out of the simulation leaving everything else unchanged then 323 the model reverts to predicting that Group Same should perform better on test than Group 324 Different.…”