如何发现 AI+Science 中的下一个 AlphaFold 和 ChatGPT?
导语
从微观到宏观,跨越广阔的空间和时间尺度,AI + Science 在发现基本粒子、量子计算、蛋白质模拟、材料设计、可控核聚变、气象预测、碳捕捉等政策设计、探索浩瀚宇宙等各个方面,都发挥着重要作用。一方面,各个科学领域中的重大问题为 AI 研究带来全新的挑战和机会;另一方面,最新的 AI 技术为解决科学领域的问题提供了强大的工具。
在集智俱乐部 AI + Science 读书会第一期,斯坦福大学计算机科学系博士后研究员吴泰霖从 AI for Science 和 Science for AI 两个方向,探讨为何要将 AI 与 Science 结合,以及 AI + Science 下一步关注的重要问题和未来面对的挑战。今天的文章整理自此次分享。
吴泰霖 | 讲者
陶如意 | 整理
梁金 | 编辑
本文由第一期 AI + Science 读书会第一期总结而成。在此次分享中,讲者吴泰霖主要就以下几个方面展开论述:
我们为什么要讨论 AI + Science?
AI for Science 关注的核心问题,以及前沿进展有哪些;
Science for AI 领域的核心问题,以及前沿进展有哪些;
AI + Science 下一步关注什么问题?
1. 我们为什么要讨论 AI + Science ?
1. 我们为什么要讨论 AI + Science ?
[1]Lam, Remi, et al. "GraphCast: Learning skillful medium-range global weather forecasting." arXiv preprint arXiv:2212.12794 (2022).[2]Lumper, John, et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596.7873 (2021): 583-589[3]Degrave, Jonas, et al. "Magnetic control of tokamak plasmas through deep reinforcement learning." Nature 602.7897 (2022): 414-419[4]AI Feynman 2.0 [1]: rediscover top-100 physics equations in Feynman lectures
[5] Victor Garcia Satorras, Emiel Hoogeboom, Max Welling,E(n) Equivariant Graph Neural Networks [2022] [6] Zhang L, Han J, Wang H, et al. Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics. Physical review letters, 2018, 120(14): 143001.[7] Sohl-Dickstein J, Weiss E, Maheswaranathan N, et al. Deep unsupervised learning using nonequilibrium thermodynamics.[8] Greydanus, Samuel, Misko Dzamba, and Jason Yosinski. "Hamiltonian neural networks." Advances in neural information processing systems 32 (2019)
2. AI for Science
2. AI for Science
图2. 动力系统仿真示意图。μt是系统的初始状态,可以是一个连续函数,或者一个图;f*是演化动力学,可以是偏微分方程的演化,或者是真实世界的演化;a 是系统不随时间变化的静态参数;∂X是系统的边界条件。
图3. 使用图神经网络模拟复杂系统。| 图片来源: Sanchez-Gonzalez et al. Learning to Simulate Complex Physics with Graph Network. ICML 2020. http://proceedings.mlr.press/v119/sanchez-gonzalez20a/sanchez-gonzalez20a.pdf
[8]Sanchez-Gonzalez et al. Learning to Simulate Complex Physics with Graph Network. ICML 2020.[9] Wu, Tailin, et al. "Learning large-scale subsurface simulations with a hybrid graph network simulator." SIGKDD 2022[10]Lam C Y, Lu J R, Udalski A, et al. An Isolated Mass-gap Black Hole or Neutron Star Detected with Astrometric Microlensing[J]. The Astrophysical Journal Letters, 2022, 933(1): L23.11]Pfaff, Tobias, et al. "Learning mesh-based simulation with graph networks." ICLR 2021[11]Wu T, Wang Q, Zhang Y, et al. Learning large-scale subsurface simulations with a hybrid graph network simulator. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2022: 4184-4194.[12]Prantl L, Ummenhofer B, Koltun V, et al. Guaranteed Conservation of Momentum for Learning Particle-based Fluid Dynamics. NeurIPS 2022[13]Li, Zongyi, et al. "Fourier neural operator for parametric partial differential equations." ICLR2021
[14]Li, Zongyi, et al. "Fourier neural operator for parametric partial differential equations." ICLR2021[15] Li, Zongyi, et al. "Neural operator: Graph kernel network for partial differential equations." arXivpreprint arXiv:2003.03485 (2020).[16]Raissi et al., Journal of Computational physics 378 (2019): 686-707
[17]Xu, Minkai, et al. "Geodiff: A geometric diffusion model for molecular conformation[18]Guo, Ruchi, Shuhao Cao, and Long Chen. "Transformer meets boundary value inverse problems." NeurIPS 2022[19]You, Jiaxuan, et al. "Graph convolutional policy network for goal-directed molecular graph generation." Advances in neural information processing systems 31 (2018).[20]Degrave, Jonas, et al. "Magnetic control of tokamak plasmas through deep reinforcement learning." Nature 602.7897
图12. 使用符号回归发现物理定律。| 来源:Udrescu, Silviu-Marian, et al. "AI Feynman 2.0: Pareto-optimal symbolic regression exploiting graph modularity." NeurIPS 2020 https://arxiv.org/abs/2006.10782
[21]Liu, Ziming, and Max Tegmark. "Machine learning conservation laws from trajectories." Physical Review Letters 126.18 (2021): 180604.[22]Udrescu, Silviu-Marian, et al. "AI Feynman 2.0: Pareto-optimal symbolic regression exploiting graph modularity." NeurIPS 2020[23] Wu, Tailin, and Max Tegmark. "Toward an artificial intelligence physicist for unsupervised learning." Physical Review E 100.3 (2019): 033311.[24] Mundhenk, Terrell, et al. "Symbolic regression via deep reinforcement learning enhanced genetic programming seeding." NeurIPS 2021[25] Du, Yilun, Shuang Li, and Igor Mordatch. "Compositional visual generation with energy based models." NeurIPS 2020[26] Wu, Tailin, et al. "Zeroc: A neuro-symbolic model for zero-shot concept recognition and acquisition at inference time." NeurIPS 2022[27] Cao, Kaidi, Maria Brbic, and Jure Leskovec. "Concept learners for few-shot learning." ICLR 2021[28] Brown, Tom, et al. "Language models are few-shot learners." NeurIPS 2020
3. Science for AI
3. Science for AI
[29] Xu, Yilun, et al. "Poisson flow generative models." NeurIPS 2022[30] Liu, Ziming, et al. "Towards understanding grokking: An effective theory of representation learning." Advances in Neural Information Processing Systems 35 (2022): 34651-34663.
4. AI + Science 下一步关注什么问题?
4. AI + Science 下一步关注什么问题?
AI+Science是一个逐渐兴起的研究方向。在这个领域我们接下来应该关注的是什么问题呢?讲者认为,一个足够重要且有潜力的问题应该满足以下几个条件:
1)普适且影响深远。也就是说这个问题的解决方案可以被用于解决其他很多问题。
2)这个问题本身可能目前看起来是模糊的,但在2-3年内是有希望解决的;
3)有充足的数据;
吴泰霖是斯坦福大学计算机科学系的博士后研究员,由Jure Leskovec教授指导。他从麻省理工物理学博士毕业,其毕业论文主题为AI for Physics and Physics for AI,本科毕业于北京大学。他的研究兴趣为AI+Science,包括开发机器学习方法用于大规模科学和工程模拟,开发神经符号方法用于科学发现,以及由科学问题启发的表示学习(运用图神经网络、信息理论和物理等方法)。他的工作发表在NeurIPS、ICLR、UAI等顶级机器学习会议以及物理学顶级期刊上,并被MIT Technology Review报道。他是美国国家科学院院刊(PNAS)、Nature Communications、Nature Machine Intelligence、Science Advances等综合期刊的审稿人。
个人主页:https://tailin.org/
AI+Science 读书会启动
“后ChatGPT”读书会启动
推荐阅读