MPI Colloquium Okko Räsänen

Computational modeling of early language acquisition without strong linguistic priors

Abstract:

Infants learn their native language simply by interacting with their environment and being exposed to speech input. While this learning process may seem almost effortless, it is in fact an enormous challenge from an information processing perspective. How do infants, when faced with the highly complex and continuous acoustic stream without explicit instruction, learn the properties of their language? How do they acquire the phonemic system of their native language? How do they learn to segment words from running speech or associate them with their referential meanings? More broadly, what drives the language learning process, given that solutions to individual sub-problems (e.g., segmentation or categorization) do not provide immediate ecological value to the learner? What kind of prior knowledge is required to bootstrap language learning from raw acoustic input?

In this talk, I will discuss recent developments in computational modeling as a means to investigate the cognitive underpinnings of early language development. I will focus particularly on models that learn directly from raw acoustic input, potentially paired with concurrent visual input available to the learner. Rather than relying on predefined linguistic representations, these models follow the general framework of predictive processing, in which learning emerges from the minimization of prediction error within and across perceptual modalities—a principle compatible with multiple theories of language acquisition. I will demonstrate how such models can bootstrap phonemic and lexical learning from raw speech without strong linguistic priors, learning different aspects of language simultaneously with the same core mechanism. The talk will also provide a brief glimpse of how contemporary learning simulations are becoming increasingly realistic with respect to model training and evaluation methodologies.