Introduction
In 2020, we will try to use fine personal data and data science on a global scale to help improve the robustness and adaptability of the social system, so as to adapt to the world that coexists with the new crown epidemic. Today, when the process of digitization and intelligence is accelerating again, the Jizhi Club and the Zhiyuan Community have invited Zhou Xiaohua, director of the Department of Biostatistics, School of Public Health, Peking University, Cui Peng, associate professor of the Department of Computer Science and Technology, Tsinghua University, and Zhang Jiang, professor of the School of System Science, Beijing Normal University, Pittsburgh University assistant professor Wu Lingfei and many other scholars shared, as well as sharing on different topics such as co-developing causal recommendation system, graph network and time series data network, understanding complexity, high school student special session, and sharing the latest exploration of interdisciplinary with everyone. And the latest insights into complex systems. At the annual Jizhi Academic Conference, let us join the causal revolution in the field of data science and artificial intelligence and explore the complex society of entangled cause and effect!
At the same time, we are calling for angel investment sponsorship and cooperation, and we look forward to building "a research institute without walls" with us!
Scan the QR code of the poster, and you can sign up for a fee.
Five reports sharing:
1. Cui Peng| Stable learning: Discover the common foundation of causal reasoning and machine learning
2. Zhou Xiaohua| The application of causal inference in medical treatment z2. The future: how to evaluate and accelerate innovation in the intelligent era
4. Li Ruiqi | The emergence of scaling laws in shared bicycle systems
5. Li Zhouyuan | Cross-scale research on biodiversity and stability
Three sub-forums:
1.Causal recommendation system
2. Network and spatio-temporal prediction
3. Understanding complexity
Home sharing
Stable learning: Discovering the common foundation of causal reasoning and machine learning
In recent years, with the application of artificial intelligence technology, the application of artificial intelligence technology has entered the "deep water zone", exposing existing machine learning. There are many risks and shortcomings in learning methods. How to develop a learning model that is stable and robust to data changes is very important for academic research and practical applications. Causal inference is a powerful statistical modeling tool for interpretation and stable learning. In this sharing, Mr. Cui Peng focuses on causal reasoning and stable learning, aiming to explore causal knowledge from observation data to improve the interpretability and stability of machine learning algorithms.
Personal profile: Cui Peng, Associate Professor of Tsinghua University, his research interests include causal reasoning and stable prediction under big data environment, network representation learning, and its application in financial technology, smart medical and social network scenarios.
The application of causal inference in medical treatment
Since Obama launched the precision medicine plan, the term precision medicine has been in the air and has attracted much attention. Precision medicine is an emerging method of disease prevention and treatment that takes into account differences in personal genes, environment and living habits. The two main causal issues involved are heterogeneous treatment effect estimation and individualized treatment plan selection. In this lecture, Mr. Zhou Xiaohua will introduce the potential outcome model and structural equation model framework, related mathematical concepts and tools. It also introduces how to solve some problems of causal reasoning in biomedicine, such as non-compliance, death truncation, confusion control, causal intermediary analysis and missing data.
Personal profile: Zhou Xiaohua , Director of the Department of Biostatistics, School of Public Health, Peking University. His research direction is missing data, causal inference analysis, big data analysis, semi-parametric models, medical testing, health economics, and the development of new statistical methods in the field of health services.
The future of innovation: How to evaluate and accelerate innovation in the intelligent era
The process of human society from simple to complex is accompanied by a series of innovations. Today, innovation is regarded as an important driving force for social development and even the maintenance of social operations. While we continue to lament the "involution", what is behind it may be that the current innovation speed is not enough to maintain the socialRapid development of this dilemma. We cannot help asking whether the development of science, technological progress, and economic prosperity are limited by human innovation capabilities? If the answer is yes, then what limits human creativity? Is it possible for us to release it by designing collective intelligence and machine intelligence? In order to obtain information that can be used to guide future innovative research, policies and business strategies Insights, we reviewed the latest research progress in innovation, and focused on the research on innovation bottlenecks caused by knowledge fusion.
Personal profile: Wu Lingfei , assistant professor at the University of Pittsburgh, is interested in how to measure innovation in science and technology, discover effective team mechanisms to accelerate innovation, and design AI tools to achieve automated innovation.
The emergence of the scaling law in the shared bicycle system
The basic law of human movement is a very important research issue in urban research, but there are few studies on the pattern of movement between vehicles and passengers. Such a problem is in the era of sharing economy. Very critical. Through the analysis of the detailed information of each trip recorded by the dockless bicycle sharing system, it reveals the behavior patterns hidden behind the trip. There are universal scaled behaviors, which has positive guiding significance for green travel under the sharing economy.
Personal profile: Li Ruiqi , associate professor of the School of Information Science and Technology, Beijing University of Chemical Technology, and director of the Urban Network Laboratory. His main research direction is urban big data analysis and modeling, complex network research and epidemic transmission dynamics.
Cross-scale research on biodiversity and stability
"Biodiversity and stability across scales: from species to community" (Biodiversity and stability across scales: from species to community), this report will take the diversity from ecology Starting from the classic speculation about the relationship with stability, based on the empirical analysis of species observation data, discuss the driving factors of community stability, show the stabilization law of populations with spatial scale changes, and share a new understanding of the relationship between diversity and ecosystem functions.
Personal profile: Li Zhouyuan , Lecturer at Beijing Forestry University, research direction is remote sensing and geographic information application, land use and climate change, biodiversity and ecosystem function.
sub-forum 1:
causal recommendation system topic
1. Content background introduction
In the early days of the intelligent era, we were exposed to various recommendation systems. If the goal of the Internet is to connect everything, then the role of the recommendation system is to establish more efficient connections. In essence, recommendation is a special form of information extraction, which uses past behavior and user similarity to generate a series of information tailored to the end user's preferences. This allows the recommendation system to connect users with content and services more efficiently, thereby saving a lot of time and cost. The task of the
recommendation system is often to answer a causal question. For example, if we "force" users to watch a movie, what will the score be? The main tool for answering this question in the industry is A/B testing. The high price and the limited nature of answering questions, the industry has a strong demand for using observational data to answer causal questions, and the use of causal reasoning Debias has become a very promising direction. This sub-forum focuses on Causal debias. We will start from the problem, explore the design of appropriate statistical indicators and their estimation methods, evaluate the Bias in the recommendation system, and use causal information to optimize the recall and ranking algorithms in the recommendation system.
2. Sharing process and introduction
Dong Zhenhua: Application and challenges of recommendation system model
Sharing topic introduction: Recommendation system is one of the most important technologies in the commercial application of artificial intelligence and big data. It is widely used in social media, online shopping platforms, In the information flow and advertising system of short video platforms, mobile browsers, etc. This sharing will introduce the main technical models used in the current recommendation system, as well as the paranoid challenges and interpretability problems that may exist in this data-oriented model-oriented technical model. Finally, we will also discuss the above-mentioned problems under the inspiration of causality. s solution.
Personal profile: Dong Zhenhua , the current technical expert of Huawei Noah’s Ark Laboratory, is responsible for the research and implementation of the recommendation system and machine learning cutting-edge technology, and helps Huawei build a recommendation system for multiple products, including: Huawei browser, Huawei application market, Advertising system, one-screen information flow and direct service, financial product recommendation system, etc. Research directions: recommendation system, information retrieval, causal reasoning, counterfactual learning.
Zou Hao: Using confusion variable balance to achieve stable learning
Sharing topic introduction: Nowadays, machine learning technology has achieved excellent results in many fields. But interpretability and stability are still problems and dilemmas that have not been fully resolved by machine learning. These two problems exist because most existing machine learning models are based on unexplainable and unstable association statistics. Therefore, we try to combine causal science in machine learning to achieve the goal of stable learning. This sharing will introduce the causal effect estimation based on the balance of confusion variables, and how to introduce this idea into machine learning to achieve stable learning.
question outline:
stable learning which applications and integration points in the industry such as recommendation systems. What are the more in-depth analysis and interpretations of
stable learning from the perspective of causal analysis.
Personal profile: Zou Hao , a third-year doctoral student in the Department of Computer Science and Technology, Tsinghua University, and his supervisor Cui Peng. The research direction is causal reasoning and counterfactual learning.
Wu Peng: Application of non-parametric/semi-parametric statistics in recommender systems
Sharing topic introduction: If we want to conduct causal inference in observational research, the usual approach is to use the reciprocal of the propensity score as a weight to eliminate confounding. But when the propensity score is close to 0 or 1, the final estimate will be very unstable. This sharing will introduce how to use the least model assumptions (semiparametric or nonparametric) to estimate the causal parameters of interest when the value of the propensity score is allowed to be close to 0 or 1, and discuss how to apply this method to recommendation System and machine learning.
question outline:
recommendation system evaluation criteria.
recommends the formal definition of various biases in the system. The causal framework in the
recommendation system. How to deal with the problem of excessive variance caused by the propensity score value close to 0 or 1 in the
recommendation system.
recommends how to combine a small amount of randomized test data (uniform data) and a large amount of observational data in the system to eliminate bias. The link between
Rerandomization and stable learning.
Personal profile: Wu Peng , post-doctorate at Peking University. The research direction is causal inference, non-parametric/semi-parametric statistics and 10+ published papers.
3. Topic discussion
1. What we expect from the recommender system
2. How to implement a recommender system with causal knowledge and reasoning ability
Sub-forum 2:
graph network and time series data network topic
1. Content background introduction
zz graph neural network has become a deep neural network One of the hottest directions in the learning field, it has been widely used in natural language processing/computer vision/recommendation systems and other fields. However, the current research on graph neural networks mainly focuses on homogeneous graphs composed of the same type of nodes and edges. However, graph data in real life is often a heterogeneous graph composed of multiple types of nodes and edges, so it is more in line with the needs of actual industrial applications. How to design a heterogeneous graph neural network architecture and apply it in actual industrial scenarios is an urgent problem to be solved. In this sub-forum, we will use spatiotemporal data mining, heterogeneous graph neural network model construction and practical applications to understand the structure and relationships in society.
2. Sharing process and introduction
Wang Huandong: Spatio-temporal data mining and user privacy protection
Sharing theme introduction: With the implementation of the national urbanization strategy in recent years, the rapid development of large cities and urban agglomerations has brought transportation, environment, etc. A series of urban problems. The root cause is our inadequate understanding of the nature of urban operations and insufficient control over its development laws. Urban big data provides us with valuable opportunities to study its laws, butWhen we want to use a data-driven approach to solve urban problems, we are faced with problems such as uneven data quality and natural separation of multi-source data. In response to the above problems, we have developed a series of intelligent spatio-temporal data enhancement technologies, which can recover global data from local data, estimate high-resolution data from low-resolution data, and form cross-domain fusion data based on multi-domain separated data. The leap from small data to urban big data. Based on these data, we have realized the modeling of the user's temporal and spatial movement patterns from the individual level to the group level, and studied the privacy protection issues.
Personal profile: Wang Huandong is a postdoctoral researcher in the Department of Electronic Engineering of Tsinghua University. He obtained a bachelor's degree in information and communication engineering and a second bachelor's degree in mathematics and applied mathematics from Tsinghua University in 2014 and 2015, respectively, and in 2019 Obtained a Ph.D. in Information and Communication Engineering from Tsinghua University. Research interests include urban computing, mobile big data mining, and reinforcement learning.
Ji Houye: Heterogeneous graph neural network and its application in Alibaba's recommendation business
Sharing topic introduction: Graph neural network has become one of the most popular directions in the field of deep learning, in natural language processing/computer vision/recommendation systems and other fields Has been widely used. However, the current research on graph neural networks mainly focuses on homogeneous graphs composed of the same type of nodes and edges. However, graph data in real life is often a heterogeneous graph composed of multiple types of nodes and edges, so it is more in line with the needs of actual industrial applications. How to design a heterogeneous graph neural network architecture and apply it in actual industrial scenarios is an urgent problem to be solved. This report will introduce the heterogeneous graph neural network and its application in Alibaba's sharing recommendation business.
Wen Haomin: Intelligent practice of urban and end-of-line distribution networks-taking the order of package prediction as an example
Sharing theme introduction: The rapid development of the e-commerce industry in recent years has placed higher requirements on the logistics industry. We will introduce the intelligent scenes of the city and the end-point distribution network, and briefly describe our related work on the prediction of the order of collection. Prediction of parcel path is essential for understanding the delivery behavior of couriers and optimizing the dispatch system. This problem is essentially a sorting problem with strict time and space constraints, and different couriers have different decision-making preferences. In response to the above problems, we proposed a deep neural network model to learn the decision-making experience and preferences of couriers by modeling the historical behavior of couriers.
Personal introduction: Wen Haomin , PhD student (Bo 2) at the Institute of Network Science and Intelligent Systems (INSIS), Beijing Jiaotong University. Main research direction: Spatio-temporal data mining. Currently, he has published a paper on ICDE, the top conference.
Lin Yan: Context-sensitive location representation learning
Sharing topic introduction: Representing location learning on spatiotemporal trajectories is a very important task, which can help multiple trajectory data mining tasks to achieve better performance. In the real world, a place often carries multiple functions. In different contexts, users visit a place for different purposes. However, the existing location representation learning methods usually learn a static embedding vector for each location, without considering the dynamic versatility of the location. In response to the above problems, we propose a context-sensitive location representation learning method, which dynamically calculates its representation according to the specific context of the target location. In this way, the multi-functional features of the location can be more accurately integrated into its embedding vector.
Personal introduction: Lin Yan, a first-year PhD student of INSIS, Beijing Jiaotong University. His research interests are spatiotemporal data mining and representation learning.
3. Topic discussion
1. Problems and bottlenecks of graph neural networks.
2. The next step of graph neural network development.
sub-forum 3: Understanding complexity
thinking and measurement of universe, life, consciousness
1. Content background introduction
People live forever, see the world, see all beings, and see themselves. As we continue to think, see, and testify, we will ask the same series of questions: Why is the world complicated? Can this complexity be recognized? Can people be known by themselves?
, walking on different roads, entered the field of complex system research in order to answer these common questions. We follow the pioneers of system research and change the perspective fromDifferent research hotspots have shifted, from chaos, to dissipation, to control, to cluster emergence, to power law, to network, to deep learning, to cause and effect, to graph network...Our research objects are different, from quantum , To the weather, to the economy, to the society, to the city, to the intelligent machine, to the game, to the life, to the brain, to the consciousness, to the universe...We are amazed at the spontaneous, self-organized, self-adapted system, and confused by the complexity The directionality and autonomy of sexual development, we hope to understand, measure, simulate and even reconstruct this magical system. Can the complexity of
be recognized? Complexity is being recognized, cracked, compromised, squeezed... This process is still going on, we are witnesses, or participants. We can rethink and examine, the complexity of what we keep talking about, or is it not "the synthesis emerging in science"? How does the brain recognize and learn complexity, and how does it recognize itself? Can we reproduce this process through understanding? Do the evolution of the physical world and the life world share the same set of rules of complexity and level formation, and can they be measured in the same way? What is the mechanism of self-replication? Is the subject the starting point of life?
This is a reunion of a group of "old Jizhi", each speaker has a decade of experience in the Jizhi community. We were pleasantly surprised to find that in the past ten years, everyone is still thinking and exploring the original problems. It's like every encounter during this period is the same feeling: time never seems to pass, we just turned around and had a cup of coffee, and the discussion could continue.
has gone a long way, and when we look back at the Prime Minister, we still see the hardcore teenagers.
understand the complexity, today you, come and play with us~
2. Sharing process and introduction
ten three-dimensional: a closure theory about consciousness
sharing topic introduction:
brain information processing process can be at multiple time and space scales However, only coarse-grained information on a specific scale seems to be able to be perceived and used for consciousness. The information closure theory of consciousness believes that the process of consciousness is a process of forming a non-trivial information closure (NTIC) on the environment at some coarse-grained level. This theory not only proposes a new quantitative definition of the content and level of consciousness, but also This phenomenon of consciousness has been explained and predicted, and it has naturally reconciled many existing consciousness theories.
ten three-dimensional , a data operation supervisor of an airline company, resident author of Jizhi Club, graduated from the Department of Mathematics of Capital Normal University, and has long been concerned with interdisciplinary research such as complex systems and cognitive science.
Yuan Mingli: Use the Serengeti game to explore the concept of natural numbers to form
Sharing topic introduction:
after a brief review of the history of number systems and the definition of natural numbers, proposed a scenario based on reinforcement learning, trying to reconsider natural numbers from the perspective of learning The concept is formed.
Yuan Mingli , a programmer for more than ten years, wandering between the key words of engineering, language, knowledge, and intelligence. His heart is full of curiosity and confusion about the world. Taste the story of Da Liu’s "Mountain" from a friend. With brute force, he can hollow out the thick wall and see the stars.
Wang Dong: The logic of deep learning and scientific discovery
Sharing topic introduction:
scientific discovery can be automated? Can new scientific discoveries be made by only relying on data and algorithms, or does it require the brilliance of some genius? The topic of this sharing is the logic of deep learning and scientific discovery. It is an introduction to discuss the issue of scientific discovery automation with everyone.
Wang Dong, with a background in philosophy of science, focusing on intelligence and cognitive philosophy.
Yuan Xingyuan: From network literature to world model
Sharing topic introduction:
With the emergence of the new trend of natural language processing ultra-large-scale corpus, is the neural network language model still a magical revised statistical model? From the weather forecast to the continuation of the novel, will the deep laws of the world's operation be hidden in the network connection? Can a language simulator be used as a strategy function of a reinforcement learning agent? Please enjoy the irresponsible brain hole
Yuan Xingyuan : Caiyun Technology CEO since 2020.
Yu Dejun: Based on Avida's self-replication evolution law exploration
Sharing topic introduction:
1. A brief introduction to the self-replication problem, its research process and the main problems. 2. Introduction to artificial life avida platform and self-replication work based on avdia 3. Inferences and conjectures.
Yu Dejun , a senior engineer in petroleum companies, is engaged in the informatization of natural gas and pipeline related enterprises, and is interested in the research of artificial intelligence and self-replication systems. Z2z
Wang Xiong: Thoughts on the origin of subjectivity From particles to complex systems
Sharing topic introduction:
How does a world governed by objective laws produce free will? One possible direction is to understand the foundation of the universe from particle models to complex system models. I will systematically review the concepts of particles and fields in theoretical physics, and imagine a basic theory that replaces particles and fields with systems.
Wang Xiong, Jizhi scientist. Engaged in complexity science and quantitative investment research and teaching. Since the time of university, he has been studying the two extreme scientific problems of the unity and complexity of natural laws.
Little Wooden Ball: Exploration of the Concept of Former Life Subjects and Secondary Emergences
Sharing topic introduction:
Early complexity research is based on the dynamic analysis of chaotic systems, dissipative systems and cluster systems. In recent years, it has focused on the understanding of networks and control . But for a multi-level complex system like life, does its hierarchical generation process share the same set of rules with the physical world? We will re-examine the emergence and hierarchical generation issues in the study of complexity from the discussion of the relationship between autocatalysis and pre-living subjects and the concept of secondary emergence.
Xiaomuqiu, PhD in tumor molecular genetics, Jizhi scientist, engaged in the research and development of new biological drugs. He has organized 24 "Out of Control" book clubs in Jizhi Club, and has long been concerned about the complexity of life and the origin of life.
Fu Wocheng: Understanding the intelligibility of complex systems
Sharing topic introduction:
Einstein once said: "The most incomprehensible thing about the universe is that it is actually understandable." The same is true for complex systems. There may be a simple micro mechanism behind a complex system, or we can establish a macro phenomenological theory of the system. Why can seemingly complex systems with high degrees of freedom be reduced to lower dimensions? In the discussion, I will combine some new advances in statistical physics (spin glass), biological evolution (mutation stability), neuroscience (neural variability), and machine learning to try to answer related questions.
Fu Wocheng, PhD in Physics, Department of Physics, Nanjing University, Postdoctoral Fellow at the University of Tokyo, Jizhi Scientist, Honorary Member of "Zhihu Salt Club 2014", published the Zhihu Salt series of e-books "Writing on the Edge of Physics" and "Critical: "Intelligent Design Principles", the one-hour e-book "Conservation of Energy", and the popular science book "Where did the universe come from". The research directions include statistical physics and its applications in life sciences.
3. Topic discussion on the origin and complexity of the
order: problems-tools-applications
researchers and the Jizhi science community
special session for high school students
this year's high school student special is still planned by high school student Feng Ruiyang , and several of her are invited Friends of high school students to share.
high school students share their own small researches in the past year and their own understanding and perceptions of scientific research, and share the project for three to five minutes.
The following is a brief introduction to the theme shared by the sharers and everyone:
Lan Zhengyang , a Beijing high school student who is ashamed of the structural injustice of the society; the sharing theme brought by
is "in a destined unfair What should students learn before the new society?".
Li Yunyun , a psychology lover who dreams of Debug the World, science and technology are her language to communicate with the world;
Guo Yuze , hope to combine science and technology with psychological knowledge, and communicate with the world between reason and sensibility;
The shared theme brought is "The influence of body language on human emotions."
Wang Yuheng , 11 years of Canadian high schoolI am reading, like hands-on, love science; what
shared is "Explore wireless power transmission technology-Wireless Power Transmission".
Zhang Fengyuan , an 11th grade high school student in Canada; the sharing theme brought by
is "software reverse engineering".
Wu Yujun , currently studying in the third AP of ICC High School Attached to the National People's University; the topic of
this time is his award-winning research in the IMMC International High School Student Mathematical Modeling Competition "Electronic product retail stores respond to snap-up decoration and layout adjustment".
Liu Yuanxin , a high school student who has the same passion for physics and history; the theme shared by
is "Exploring the relationship between glass and colored glaze".
Introduction to previous annual meetings
Jizhi Club 2019 Academic Annual Meeting
Angel Investment Sponsorship and Cooperation
Our annual meeting is still based on directional invitations, and the number of participants is expected to be around 200; the theme of the event is "causal entanglement", and at the same time It also covers other technical topics, such as statistical physics, deep learning, causal science, graph networks, artificial intelligence, etc.; our participants are mainly research scholars, technical engineers/products/operations, masters and doctors who explore in scientific research, It is composed of active volunteer representatives, company founders or investors in the community.
In order to successfully host this annual meeting, we sincerely invite companies to actively sponsor. We will try our best to meet your business needs in recruitment, brand promotion, technology discussion, and industry-university-research docking. At the same time, we promise that Jizhi Club will be a non-profit Organization, all the funds you sponsor will be used for the operation of this annual meeting and the Jizhi Club!
If you are willing to cooperate, please contact Wang Ting, the person in charge of this event, to discuss detailed cooperation matters.
