報告人簡介
李遠寧，上海科技大學(xué)生物醫(yī)學(xué)工程學(xué)院研究員、助理教授、博導(dǎo)、獨立課題組組長，美國卡內(nèi)基梅隆大學(xué)神經(jīng)計算與機器學(xué)習(xí)博士，加州大學(xué)舊金山分校神經(jīng)外科博士后，長期從事計算認知神經(jīng)科學(xué)研究，入選國家高層次青年人才計劃，曾獲美國國立衛(wèi)生研究院“神經(jīng)科學(xué)杰出學(xué)者獎”，2023年“腦科學(xué)與類腦智能科創(chuàng)新青年30人”，第一及通訊作者成果發(fā)表在Nature Neuroscience,Nature Communications,Science Advances,PNAS等期刊，參與科技創(chuàng)新2030 “腦科學(xué)與類腦研究”重大項目等科研項目。
內(nèi)容簡介
Understanding the shared coding of speech and language between deep neural network models and the human brain Speech and language play crucial roles in human communication, cognition, and social interactions. This talk explores the convergence between deep neural network (DNN) models and the human auditory and language processing systems. By analyzing the neural coding from the auditory nerve to the speech cortex using state-of-the-art DNN representations, we found that self-supervised learning (SSL) models, such as Wav2Vec2 for speech and GPT-2 for language, demonstrated high prediction accuracy of neural responses. Shared components between speech and language DNNs suggest that contextual and acoustic-phonetic information encoded in these models contribute to distinct spatiotemporal dynamics in brain activity. These findings reveal a significant alignment between DNN model representations and the neural processes underlying speech and language, offering novel insights into the modeling of neural coding in the auditory and language networks.
