Speaker: Meng Lu （卢萌, Peking University）
Time: 10:00 - 12:00, 2024-11-16
Venue: A6-1
ZOOM: 637 734 0280
PW: BIMSA
Organizer: Rongling Wu
Abstract
Intelligence emerges from the intricate structural and dynamic organization of the neuronal system. From a bottom-up perspective, this system is hierarchically constructed, progressing from single molecules to cells, and ultimately to neuronal networks. Understanding the structural dynamics at each level, from a combined perspective of biology, physics, and mathematics, is crucial for uncovering the physical and functional basis of intelligence.
At the molecular level, super-resolution microscopy, combined with coupled diffusion-advection equations and fractal geometry, revealed a phase-dependent mechanism of aggregate formation: active transport transitions to diffusion as aggregates grow, providing new insights into neurodegenerative disease markers. At the cellular level, using state-of-the-art deep learning models and graph theory, we uncovered a causal relationship between the structure of the endoplasmic reticulum (ER) and lysosomal motion. Our findings demonstrate that lysosomes actively regulate ER reshaping, a process critical for both metabolic adaptation and neuronal development. At the network level, we developed a multi-scale recording system based on an integrated optical-electrode device, enabling simultaneous high-resolution spatial and temporal measurements of neuronal activity and structure. This system captured the structural dynamics of neurons, spanning from large-scale network organization involving thousands of neurons to detailed calcium activity at individual synapses.
Beyond the physical organization of intelligence, we propose a theoretical framework for modeling its abstract representation, organization, and evolution. This framework conceptualizes tokens as fundamental units embedded in manifold in a high-dimensional intelligence space. The varying curvatures in manifolds reflect the intrinsic structure of intelligence. Thought flow一the sequential activation of tokens一follows geodesic trajectories on these manifolds. Consciousness integrates external stimuli with the ongoing thought flow, dynamically adjusting token embeddings, manifold curvature, and geodesic paths.
In this talk, I will present our multidisciplinary approach to studying intelligence, spanning from single-molecule dynamics to large-scale neuronal networks, and from the physical organization of neurons to a mathematical framework based on Riemannian geometry, providing a holistic view of the structural dynamics underlying intelligence.
Speaker Intro
本人于2024年1月在北京大学成立人工智能与动态结构实验室，致力于通过理论与实验相结合的方法探索智能的本质。实验上，我们开发先进的工具，获取高质量的数据并基于此构建理论，通过理论对实验进行指导从而获得更好的数据，推动理论的不断完善。本人希望通过跨学科的协同研究，促进人工智能与神经科学领域的深度融合与发展，逐步深入对智能本质的探索。近五年发表相关论文20篇，其中在Nature Methods、Science Advances等期刊以独立一作发表论文8篇，受邀于瑞士联邦理工（EPFL）、国际理论物理研究中心（ICTP）等知名机构和国际学术会议做报告，任英国阿尔兹海默研究基金会（Alzheimers Research UK）基金评委，并于2023年受剑桥大学出版社之邀成为生物成像和分析领域的国际期刊Biological Imaging 的副编辑（Associate Editor）。
Science for AI: 探索构建基于黎曼几何的智能与意识理论框架，研究智能的几何化本质，核心思想可总结为：the geometry of intelligence guides how the consciousness navigate, the consciousness dictates how the geometry of intelligence evolves.
AI for Science: 开发先进的人工智能工具，并将其与其他技术结合，研究神经元网络的结构动力学。主要成果包括：1）研发了基于石墨烯微电极阵列的多模态神经元结构与活动检测系统，实现了光学成像与电生理记录的同步，首次实现了神经元动作电位与突触级亚结构形变的同时检测；2）结合超分辨成像技术开发了人工智能视频图像分析工具ERnet，成功解决了内质网结构的精确识别与定量分析难题，精准量化了其在不同疾病模型中的形态变化。
