Wednesday, April 19, 2017

What’s statistical about learning? Insights from modelling statistical learning as a set of memory processes

Statistical learning is a term used to describe the discovery of patterns within incoming information, and has been widely applied to language learning. Language is composed of many different, and learnable, patterns. For example, knowing that the two syllables “ba” and “nan” will be followed by “a”, to make the word “banana”. Or, learning the pattern “is verb-ing”, such as “is running”, “is hopping” or “is skipping”. Infants are sensitive to these patterns when learning language, and considerable research over the last two decades has highlighted statistical learning as one of the primary ways in which humans acquire language.

This paper distinguished between two types of patterns, or statistics, learned via statistical learning. The first are conditional statistics, which refer to the likelihood of two or more elements co-occurring, such as syllables or words. Conditional statistics are useful because they go beyond the simple frequency of co-occurrence. For example, given the phrase “the dog”, you know they are two distinct words because of the low conditional relationship between them; Although “the dog” is a frequent phrase, the conditional relationship between “the” and “dog” is low because the word “the” occurs in combination with many other words. The other type of statistic is known as a distributional statistic, which refers to how the variability of items stored in memory shapes what is subsequently stored within memory. For instance, if a newly encountered item matches something that is already stored in memory, it won’t be stored as a unique item. However, if this item is distinct from what is already stored in memory, it will be stored as a unique item. Learning the distributional regularities within a language plays an important role in learning the phonemes within your native language, and other patterns across language.

The statistical learning of conditional and distributional statistics are natural extensions of processes we know to exist in memory more generally. These memory processes include recognizing familiar items already stored in memory, gradually forgetting items stored in memory if they are not re-encountered, and interference between similar items stored in memory. Looking at statistical learning from the perspective of memory has a couple of important implications. First, it connects statistical learning, and thereby language learning, to other types of learning. Second, it gives and explanation as to why the ability to learn a language tends to decline with age. If statistical learning is linked to memory processes, and memory changes with age, changes in memory over the lifespan may help explain why language learning tends to decline with age. Exploring the connection between statistical learning and memory is important for theories of language acquisition because it connects language acquisition with what we know about human cognition.  

Blogger: Nicolette is a Psychology Ph.D. student, supervised by Dr.s Lisa Archibald and Marc Joanisse

1 comment: