뇌인지과학과에서는 7월 셋째 주 목요일에 세계적인 석학을 초빙하여 BCS 10-10 Focus Lecture Series 행사를 개최합니다.

Blake Richards 교수님은 McGill 대학교 교수이자, 딥러닝의 창시자 중 한 명인 Yoshua Bengio가 설립하고 deep leanring 연구를 선도하고 있는 Mila (Quebec Artificial Intelligence Institute) 소속으로, neuroscience기전 기반 AI 알고리즘을 개발하며, universal principles of learning and memory 분야 연구를 활발히 수행하고 계십니다.

학부 시절에는 AI를 전공하고, 박사 및 박사후 과정에서는 wet-lab 기반의 시각신경과학(visual neuroscience) 및 hippocampal engram 연구(Josselyn/Frankland Lab)를 수행하신 이력이 있어, 뇌과학과 인공지능에 관심 있는 학생들에게도 큰 영감을 줄 수 있으리라 기대됩니다.

관심있는 분들의 많은 참여를 부탁드립니다.

ㅣ일시 : 7월 17일(목) 오후 4시
ㅣ장소 : 501동 목암홀(1층)
ㅣ연사 : Blake Richards, Ph.D (Associate Professor, McGill University & Core Faculty Member, Mila)
ㅣ주제 : "Representational geometry in self-supervised learning"
ㅣ초록 : The geometry of representations in a neural network can significantly impact downstream generalization. Data from animals suggests that brains employ representations with geometric properties that are ideal for downstream generalization. We show that deep neural networks develop similar representational geometry during self-supervised pre-training. First, in vision models, we demonstrate that the best performing models are those with brain-like geometries. Next, we examine the geometry of representations in large-language models (LLMs) during pretraining. We find that LLM representations also converge towards more brain-like geometry, but in a non-monotonic manner. In fact, LLMs exhibit distinct phases of geometry during pretraining, and these phases predict how well the LLM will perform following fine-tuning. Altogether, our work provides a set of brain-inspired evaluations for assessing the quality and dynamics of self-supervised learning in deep neural networks.

※ 문의 : 뇌인지과학과 (02.880.8011 / hjinkim1@snu.ac.kr)