From Sound Primitives Learning to Structure Extraction: Brain-inspired Model of Infant Early Vocal Learning
Résumé
The underlying mechanism of the biological neural machinery of the human brain receives
an upsurge of attention from machine learning scientists, as developmental learning gives
insight into open-ended unsupervised learning. In particular, language acquisition remains an
exploitable field in not only machine learning but psychology and biology as well.
Psychologically speaking, infants’ perception of speech sounds in early infancy contributes to
later language acquisition, making it one of the main paths of exploring how infants’ brains
manage to process early information. It is noticeable that adults’ perception of speech sounds
becomes constrained compared to that of newborns, as influenced by the phonetic patterns of
their native language, which appears by the end of the first year after birth. Correspondingly,
in machine learning’s perspective, it is doubtful if unsupervised learning in adulthood might
happen without supervised learning. Thus developmental learning until infants’ first
anniversary is believed to be the most similar mechanism to unsupervised machine learning.
As a result, our research focuses on modeling computationally the learning mechanism of speech
sound in early infancy, more precisely before 12 months of age, and we propose a bio-inspired
early-infant-based neural network for the learning of speech sounds, which lays a foundation
for the later grammatical structure extraction neural network for language acquisition.
Origine | Fichiers produits par l'(les) auteur(s) |
---|