Statistical Learning

Statistical learning is the capacity to detect and represent regularities in the environment through passive exposure, without explicit instruction or reinforcement. In the context of language acquisition, it refers specifically to the ability of infants and young children to extract structure from the auditory input — identifying word boundaries from transitional probabilities between syllables, detecting distributional patterns in syntactic frames, and using frequency information to bootstrap into higher-level grammatical knowledge.

The demonstration that infants as young as eight months can segment continuous speech into word-like units based solely on statistical cues — pioneered by Jenny Saffran and colleagues in the 1990s — fundamentally changed the empiricism-nativism debate. It showed that general learning mechanisms, operating over the statistical structure of the input, could account for at least some aspects of grammatical acquisition that the Poverty of the Stimulus argument claimed were unlearnable.

Statistical learning is not limited to language. The same mechanisms support visual pattern learning, musical structure acquisition, and social prediction. The question is not whether statistical learning exists — it does, robustly — but whether it is sufficient for the full range of grammatical knowledge, or whether it requires supplementation by innate constraints, social scaffolding, or other domain-specific mechanisms.