社会大数据跨尺度系统学习理论与方法

Cross-scale systematic learning for social big data:theory and methods

以GPT-4为代表的AI大模型时代正加速而至,深刻改变着社会生活的方方面面.大模型巨参数深度学习是破解复杂大数据智能学习瓶颈的一种有效途径.大模型在展现出强大学习能力的同时也面临着高能耗、大算力挑战.研究表明,平均一个AI大模型训练产生的能耗约等于五辆汽车一生排放的碳总量,驱动AI大模型所需算力每3.5个月翻一番.作为一种有益的补充,内嵌规律的跨尺度系统学习是破解复杂大数据智能学习瓶颈的另一种有效途径.跨尺度系统学习已经在某些专业领域展现出了巨大的成功,如2021年诺贝尔物理学奖授予复杂物理系统跨尺度建模及其在全球气候变暖中的应用.事实上,我国科学家甚至更早开拓了复杂系统跨尺度学习研究,如北京航空航天大学暗物质大数据分析团队利用跨尺度系统学习方法实现了在PB级数据中实时学习KB级关键数据,精度达到万分之一.本文从微观尺度、介观尺度和宏观尺度上分析了跨尺度系统学习的基本原理,构建了内嵌规律跨尺度系统学习的普适方法,并以社会大数据为例开展了典型应用示范.社会大数据跨尺度系统学习应用于疫情防控、舆情分析等领域,并取得显著成效,为我国社会治理数字化、网络化、智能化发展提供了新的成功样本.

The era of AI large models,represented by GPT-4,is accelerating,and profoundly transforming various aspects of societal life.Large models with massive parameters in deep learning offer an effective approach to unraveling the bottleneck of complex big data intelligent learning.While these large models showcase powerful learning capabilities,they also face challenges of high energy consumption and computational power requirements.Research indicates that the average energy consumption produced during the training of one AI large model is roughly equivalent to the total carbon emissions from five cars throughout their lifetimes,and the computational power needed to drive AI large models doubles every 3.5 months.As a beneficial complement,law-embedded cross-scale systematic learning presents another effective approach to address the challenges of complex big data intelligent learning.Cross-scale systematic learning has demonstrated significant success in some professional domains,such as the 2021 Nobel Prize in Physics awarded for cross-scale modeling of complex physical systems and its applications in global climate change.In fact,Chinese scientists have pioneered research in cross-scale learning of complex systems,with the team analyzing dark matter big data at Beihang University utilizing cross-scale systematic learning methods to achieve real-time learning of critical data in petabyte-scale datasets,achieving precision at the level of one in ten thousand.This paper analyzes the fundamental principles of crossscale systematic learning at micro,meso,and macro scales,establishes a universal method for law-embedded cross-scale systematic learning,and conducts typical application with demonstrations using social big data.The applications of cross-scale systematic learning in areas such as epidemic prevention and control,and public opinion analysis have achieved remarkable results,providing new successful examples for the digitization,networking,and intelligence development of China’s social governance.