Reinforced statistical learning of auditory categories: A preliminary report of cognitive, cortical and computational mechanisms

摘要

Learning to categorize speech sounds is crucial for language acquisition and a struggle for second language learners. Most research has examined supervised learning in which learners receive feedback. This offers a useful model for second language acquisition. However, work in first language acquisition suggests people may acquire categories by tracking the statistics of the input without feedback. Few studies examine how these forms of learning work together and little cognitive neuroscience examines unsupervised learning. To study this interaction, we developed the reinforced statistical learning paradigm in which learners are first exposed to statistically structured stimuli in an unsupervised paradigm and then receive supervised training. We present preliminary data from two studies using this paradigm. It suggests that the unsupervised phase benefits later supervised learning, but only in limited conditions that may vary according to whether items are blocked or interleaved. We examined the cortical consequence of learning with electro-corticographic recordings from the surface of the brains of two epilepsy patients. A machine learning analysis suggests some surprising ways in which auditory cortical representations change with learning. This suggests the need for theories that explain the varied effects of unsupervised learning on auditory processing.