Reconsidering the Automaticity of Visual Statistical Learning

Abstract

Statistical learning refers to the process of extracting regularities from the world without feedback. What are the necessary conditions for statistical learning to arise? It has been argued that visual statistical learning (VSL) is “automatic”, such that subjects will passively and even unconsciously extract statistical regularities from streams of visual input as long as they attend to the stimuli. In contrast, our data indicate that simply attending to stimuli is not, on its own, sufficient for learning. In Experiments 1 & 2, we provided incidental exposure to regularities in a stream of images and observed little to zero VSL across a range of conditions. In Experiment 3, we found that explicitly instructing participants to seek regularities dramatically improved their performance on direct measures of learning, but not on an indirect response time measure. Finally, in Experiments 4 & 5, we demonstrated that a methodological confound in prior work using the indirect response time measure could account for some previous evidence of automatic and implicit VSL.Overall, we found very little evidence of learning using direct measures of VSL, and no evidence of learning using an indirect response time measure. Participants who recognized visual sequence regularities in a forced-choice task could also often recreate the sequences when explicitly probed, indicating their knowledge was not entirely implicit. We suggest that some form of active engagement with stimuli may be needed to extract sequential regularities, and that VSL does not occur automatically.

What canonical online and offline measures of statistical learning can and cannot tell us

“…This correlation reached significance in only one of the three methods we used to evaluate the relationship, despite the fact that we followed two methods from earlier papers that identified a significant relationship. [21], [26] The weak and variable correlation we observed, in combination with previous findings reporting variable to null results [11], [18], [20], [23], [25], calls into question whether these two measures of SL are in fact evaluating learning of the same or different features of the input.…”

“…Himberger and colleagues recently argued that the graded RT effect observed in numerous SL studies is an artifact unrelated to the regularities that experimenters expect participants to learn, but rather a consequence of general RT facilitation, combined with a design that confounds position in the triplet with position in the stream. [25] We demonstrate empirically that this confound is not present in our data. RTs in both experiments did not trend towards linearly faster responses over the course of a trial/stream; RTs in Experiment 1 hovered around the mean, while RTs in both conditions in Experiment 2 increased for a majority of the trial.…”

“…Lastly, we wanted to address a possible confound that could have helped generate the RT effect. As argued by Himberger et al [25], the widely observed graded RT effect could be driven by a trivial, overall speeding-up of RTs combined with the fact that targets appear later in the stream as well as later in the pseudoword or triplet. To ensure that this confound is not present in our data, we ran a linear model with triplet position and stream position (1-216) within each trial as predictors.…”

“…Meanwhile, a larger number of studies report no correlation between these measures. [11], [18], [20], [23], [25] The weak to nonexistent correlation between these SL measures has been largely discussed on a theoretical level. [18], [22] (But see [20] for an empirical treatment of the subject.)…”

“…[13], [24] Specifically, Himberger et al have demonstrated that RT facilitation could be observed independently of any regularity-learning if the task design confounds the position of the target within a triplet sequence with its position in the test stream. [25] The authors argued that RTs to targets in the first position of a triplet may be slower than those to targets in later positions due to trivial facilitation, as participants became more familiar with the task.…”