Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Language Acquisition in Brains and Algorithms: towards a systematic tracking of the evolution of language representations using stereoelectroencephalography recordings in children and deep learning.
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster A100 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Also presenting in Lightning Talks A, Tuesday, October 24, 10:00 - 10:15 am CEST, Auditorium
Linnea Evanson1,2, Christine Bulteau2, Mathilde Chippaux2, Georg Dorfmüler2, Sarah Ferrand-Sorbets2, Emmanuel Raffo2, Sarah Rosenberg2, Pierre Bourdillon2,4,5, Jean-Remi King1,3,5; 1Laboratoire des systèmes perceptifs, École Normale Supérieure, PSL University, 2Hospital Foundation Adolphe De Rothschild, Paris, 3Meta AI, 4Paris-Cité University, 5Equal Contribution
The human brain, unlike other species, has evolved a unique ability to communicate through language by combining a limited set of known elements (words) into a representation that is both novel and meaningful. However, how humans learn to perform this task, through computation algorithms and structural organisation in the brain, remains largely unknown. In this project, we investigate how the representations of language in the human brain change as children mature, using classic models of linguistic features, and artificial neural network models. To this end, we have collected stereoelectroencephalography (sEEG) signal in 30 children, between 3 and 19 years old, while they listen to an audiobook, Le Petit Prince by Antoine de Saint-Exupéry. This forms the first dataset of its kind on naturalistic linguistic stimuli. We present a decoding analysis on phoneme and lexical level linguistic features, as well as a comparison with the artificial neural network Wav2Vec2.0. We present three main findings from decoding the classical linguistic phenomena. The first finding is that phoneme features such as voicing, as well as lexical features such as word frequency can be decoded in an autistic child of three years old, replicating what is seen in neurotypical adults. Secondly, we show that there is a decrease in reaction time with age, for lexical features, but also an increase in total processing time. This indicates that older children process lexical information more quickly but also maintain that information in the brain for longer, allowing it to be integrated with other information for higher level abstraction. Thirdly, we show the sequence of processing that occurs topographically in the brain. In young children, response times are similar across brain areas indicating that it is only lower level information that is represented even in the frontal cortex. However, in older children there is a delay that increases with distance from the primary auditory cortex indicating that there is integration of multiple levels of information, and more abstract features are represented in the highest brain areas. We also present a comparison with Wav2Vec2.0, where we show that early layers in the model are good predictors of the primary auditory cortex, and deeper layers are able to explain activations in higher brain areas but only in the older children. This research is significant as it provides a better understanding of how language is represented in the brains of children of different ages. While most features can be decoded in even the youngest children, their spatial organisation differs from mature brains, and the presence of delays, and lack of delays across brain areas indicates a hierarchy of processing algorithms that are not yet in place in the youngest children, and must be learned during development.
Topic Areas: Language Development/Acquisition,