用婴儿家庭真实录音训练的大语言模型复现语言习得过程

Home language: Large language models trained on real-life speech recorded in babies’ home environments support the theory that such input is sufficient to engender their language acquisition, according to a new study. Investigators recorded 12 hours of sound per day in the homes of eight different families for the first 1,000 days of the children’s lives. Artificial-intelligence models built with these datasets learned to distinguish speech sounds and built vocabularies of thousands of words in ways that mirrored each child’s development. This approach “offers a framework for translating theories of how children learn into computational models that can be tested directly against children’s real-world development,” the authors write.

Autism research spotted this week:

• “Spatial single-cell mapping of transcriptional differences across genetic backgrounds in mouse brains” Nature Methods
• “Trends in autism spectrum disorder cumulative incidence and incidence density: A systematic review” Biological Psychiatry
• “DDX3X deficiency in mice impairs corticofugal motor circuits and causes motor deficits that are attenuated by early-life exercise” Proceedings of the National Academy of Sciences
  See also: “DDX3X loss affects male, female mouse neurons differently”
• “SARS-CoV-2 mRNA vaccination in pregnancy and child neurodevelopment at 18 to 30 months” JAMA Pediatrics
• “Sex and autism share cell type-specific effects on cortical gene expression” medRxiv
• “Optically-pumped magnetometer-MEG (OPM-MEG) can detect and map epileptiform activity in diagnostic and presurgical paediatric cohorts” Clinical Neurophysiology

注意转载授权 · 阅读原文