为促进学术交流，拓宽学生专业视野，推进学科发展，医学技术研究院特邀请北京大学人工智能研究院杜凯博士来我院进行学术交流，并做专题学术讲座，欢迎学院师生参加活动并交流。

讲座时间：2023年11月21日 下午16:00-18:00

讲座地点：病理楼231室

讲座题目：Biophysically Detailed Brain Simulation: Next Frontier in Brain Science and AI

报告摘要：

The recent success of the Super AI model, "ChatGPT," is transforming our understanding of intelligence. One captivating notion is that ChatGPT's intelligence seems to automatically "emerge" from an immense and intricate artificial neural network with over 170 billion parameters. This leads us to ponder: how might intelligence emerge from our own brains? Within the human brain, there are nearly 100 billion neurons, which can be categorized into thousands of distinct types. Each type of neuron has its own unique and elegant structure, along with an abundance of ion channels. Cutting-edge experimental advancements suggest that due to their complex dendritic structures and ion channels, individual neurons possess extraordinary computational power, rivaling that of five-layer deep learning networks. Consequently, individual biological neurons should not be viewed as mere "point-models," but rather as highly complex neural networks. Hence, our brain may contain a hyper-large neural network with parameters 4-5 orders of magnitude greater than those of ChatGPT.

Biophysically detailed brain simulations stand as the sole mathematical approach capable of capturing the complexity of dendritic structures, ion channels, and synaptic interactions. However, high computational costs severely limit its application in the neuroscience and AI fields.    

In this presentation, I will commence with a historical overview of dendritic computation research, elucidating the theoretical foundations that have led to biophysically detailed brain simulations. Following this, I will explore contemporary theoretical studies on dendritic computation and its potential applications in AI, highlighting how this evolving understanding is reshaping both neuroscience and AI paradigms. Finally, I will present our latest developments in a GPU-based simulation framework that achieves up to a 1,500-fold speed up over traditional CPU-based NEURON simulator (Zhang, et.al.,Nat. Comm.,2023). This technological advance opens the door to simulating and training biologically realistic neural networks on an unprecedented scale, unlocking the potential of human-level intelligence.

嘉宾介绍：    

杜凯博士于2002年在北京航空航天大学飞行器动力工程系获得学士学位，并在瑞典卡罗琳斯卡医学院神经科学系取得博士学位，随后在该院进行了博士后研究。在2013至2016年间，他是欧盟脑计划“大脑仿真平台”瑞典团队的主要成员。2020年，加入了北京大学人工智能研究院，并参与创建了北京智源人工智能研究院的生命模拟部门。他的主要贡献包括构建了首个针对基底核脑区的精细神经元模型，并开发了一种基于GPU的高性能计算框架—DeepDendrite。该框架不仅显著提高了大脑模拟的计算速度，还成功地将树突计算原理与人工智能模型进行了紧密的整合。他的研究成果已作为第一作者或通讯作者发表在《PNAS》、《Nat. Commun.》等国际一流学术期刊。杜凯博士目前的主要研究方向是大脑的精细仿真、树突计算，以及基于大脑精细模型的新型人工智能系统和理论。