Search Abstracts | Symposia | Slide Sessions | Poster Sessions
The ‘what’, ‘how’, and ‘where’ of statistical learning: A comprehensive synthesis and new neurocognitive theoretical framework
Poster D20 in Poster Session D, Saturday, September 13, 5:00 - 6:30 pm, Field House
Christopher Conway1, Karolina Janacsek2,3, Joshua Buffington4, Michael Ullman4; 1Grinnell College, 2University of Greenwich, 3Eotvos Lorand University, 4Georgetown University
During recent decades it has become clear that the ability to learn environmental regularities—known as statistical learning—is crucial for human language learning, as well as cognition more generally. Considerable advances have been made in this area of research, which is growing at a near-exponential rate. However, as pointed out in recent influential critiques, the field appears to be at a crossroads. In particular, there remain fundamental questions about just ‘what’ statistical learning is, ‘how’ it occurs, and ‘where’ in the brain it is implemented. Here we address these questions by adopting a framework inspired by Marr’s (1982) classic levels of analysis, which broadly correspond to the questions of ‘what’ (the computational level), ‘how’ (the algorithmic level), and ‘where’ (the implementational—i.e., neurobiological—level). At the ‘what’ level, we suggest that statistical learning may be most usefully defined prototypically, as an umbrella term that encompasses a cluster of learning situations that tend to have certain characteristics, namely the learning of structured patterns, encountered over multiple exposures, under incidental conditions. We focus on three tasks that often meet these criteria: embedded pattern (e.g., word segmentation), artificial grammar, and serial reaction time tasks. At the ‘how’ level, we focus on two widely-studied learning algorithms: prediction-based learning and chunking. At the ‘where’ level, we focus on two well-studied learning and memory structures—the basal ganglia and medial temporal lobe—as well as neocortex. We comprehensively review and synthesize the cognitive and neural literature on statistical learning and then map the different levels to each other. We propose and show that statistical learning is supported by “many-to-many” mappings. Specifically, we suggest statistical learning can be carried out by different cognitive algorithms (both prediction-based learning and chunking) and different neural structures (in particular the basal ganglia and medial temporal lobe, but also certain neocortical regions), and moreover each structure can subserve more than one algorithm and task. Importantly, the relative reliance of statistical learning on particular algorithms and structures may be a function of a wide range of task, item, context, subject, and other factors. We suggest that this “many-to-many” viewpoint is a more advantageous way to think about statistical learning compared to perspectives that emphasize one algorithm or neural substrate over another. Finally, we discuss future directions. We emphasize that the cognitive algorithms and neural structures implicated in statistical learning are quite well-understood based on independent research from the broader study of the neurocognition of learning and memory. This independent knowledge generates a wide range of specific and often novel predictions that might not be made in the more circumscribed study of statistical learning alone. In sum, our framework has the potential to substantially advance the field of statistical learning by offering a clear account of the conditions under which statistical learning occurs and the cognitive and neural substrates that support it.
Topic Areas: Language Development/Acquisition,