People's Government of Zhejiang Pronvice
Zhejiang University
Alibaba Group
After summarizing the relevant aspects of a relatively new paradigm of computing, namely Intelligent Computing (IC), as described by some of the leading IC researchers, I will discuss many recent developments in the data management area that have relevance to IC. As someone who has been, and who continues to be, active as an individual contributor in the database (DB) area for over four and a half decades, in this talk, I will give a broad overview of the evolution of the data landscape. I will discuss not only research topics but also the trends in the commercial and standards arenas. Some of the topics that will be covered are cloud DBMSs, hardware-software co-design, multi modal integrated DBMSs, deeper integration of AI and DBs (including Large Language Models), vector indexes, and edge analytics.
We describe some recent advances in the development of general-purpose numerical algorithms for linear, semidefinite programming (LP/SDP), and AI Training optimization, including fast pre-solving and diagonal-preconditioning, LP cross-over, ADMM, First-order Hype-Gradients, GPU Implementations. Most of these techniques have been implemented in an emerging optimization solver COPT, and they increased the average solution speed by over 3x in the past three years on a set of benchmark problems. For certain problem types, the speedup is more than 100x, and problems that have taken days to solve or never been solved before are now solved in seconds or minutes to high accuracy.
In this talk, I will share our experience and lessons learned during the development of Tongyi Qianwen, also known as Qwen, a state-of-the-art large language model at Alibaba Cloud. I will first outline the key steps taken to construct a high-performing model with the ability to generate creative text, comprehend intricate instructions, tackle mathematical problems, and more. Subsequently, I will describe a variety of systems challenges in building large language models and present our innovative design in the areas of distributed storage, high performance networking, resource scheduling, and execution frameworks. Such techniques significantly enhance the efficiency of handling complex AI workloads in a distributed environment.
Today’s autonomous driving system are facing severe challenges in highly dynamic, random and dense traffic scenarios. Existing hierarchical design method, for example, that with rule-based decision and linear motion controller, is lack of sufficient adaptability and flexibility. As a biologically inspired artificial intelligence, reinforcement learning (RL) is promising to provide the self-evolving ability for automated cars, which has the potential to generalize to unknown driving scenarios. This talk will discuss recent advances in learning-based autonomous driving systems for high-level self-driving intelligence. An interpretable and computationally efficient framework, called integrated decision and control (IDC), is proposed to fulfill more flexible functionality, in which the standard actor-critic architecture can be subtly utilized to train its decision and control neural networks. Some technical breakthroughs in safe reinforcement learning and high-fidelity simulator are also discussed for the purpose of training more accurate neural network controllers.
In this talk, we introduce a biologically-inspired hierarchical learning framework as a potential solution to the simulation-to-reality gap problems in humanoid robot control. The proposed three-layer hierarchical framework enables the generation of multi-contact, dynamic behaviors even with low-frequency policy updates of whole-body model-based optimization. The upper layer is responsible for learning an accurate dynamics model with the objective of reducing the discrepancy between the analytical model and the real system. This enables the computation of effective learning of whole-body humanoid control policies. We present the evaluation results of the proposed method on a a real robot system.
Astronomy is currently in a golden age. In the 21st century, the Nobel Prize has been awarded 7 times to a total of 10 achievements and 19 individuals in astronomy. The latest major frontier in astronomy is the study of extreme explosive phenomena in the dynamic universe, represented by Fast Radio Bursts (FRBs). In 2023, the Shaw Prize in Astronomy was awarded for the discovery of FRBs. In 2024, the Marcel Grossmann Award was given for in-depth research on FRBs. Within a thousandth of a second, FRBs can release the equivalent of the total radiation energy of the Sun in a decade. The yet unknown origin of FRBs manifest that Humans are beginning to confront the cosmic "time frontier": extreme energy released in extremely short timescales. Currently, there are two main radio telescopes at the forefront of observing FRBs: the CHIME array in Canada and the Chinese FAST (Five-hundred-meter Aperture Spherical radio Telescope). The former leads in the discovery of new FRBs, while the latter leads in in-depth observations of FRBs. Since 2020, these two facilities have produced most of the Nature/Science papers in this field. In order to detect the energy characteristics within the shortest observable timescale and preserve the cosmic EM field information as completely as possible, we have proposed the novel concept of the "Cosmic Antennae (CA)" phase-field telescope, which uses unprecedented large-scale high-throughput astronomical signal computation and artificial intelligence to restore the dynamic universe field. With the support of the Zhejiang Laboratory, the first generation CA will achieve an unprecedented 10,000 square degrees FoV with high-cadence monitoring. CA expects to achieve an FRB discovery efficiency 30 times better than that of FAST. The goal is to lead this new frontier by sensing the dynamic universe through computation.
Moderator: Michael G. Somekh, Zhejiang Lab, Fellow of the Royal Academy of Engineering, UK
Panelist: C. Mohan, Hong Kong Baptist University & Tsinghua University, Member of NAE, USA
Qiuming Cheng, China University of Geosciences (Beijing), Academician of CAS
Stefanie Walch-Gassner, University of Cologne, President of the German Astronomical Society
Can Wang, Zhejiang University, China
Graph is a ubiquitous non-Euclidean structure, describing the intercorrelated objects in the complex system, ranging from social networks, transportation systems, financial transaction networks to biochemical and molecule structures. Nowadays, graph neural networks are becoming the de facto solution for learning on graphs, generating node or graph embeddings in representation space, such as the traditional Euclidean space. However, a natural and fundamental question that has been rarely explored is: which representation space is more suitable for complex graphs? In fact, the "flat" Euclidean space is suitable for grid structures but is not geometrically aligned with generic graphs with complex structures. Thus, it is interesting to explore deep graph learning in different geometric spaces. This talk will delve into the fascinating facts and properties of various geometric spaces (e.g., hyperbolic space and hyperspherical space), and discuss some preliminary works on tasks such as classification, clustering, contrastive learning, graph structure learning, and continual graph learning. These endeavors pave the way for the next generation of deep graph learning.
In this talk, we will try to clarify different levels and mechanisms of intelligence from historical, scientific, and computational perspective. From the evolution of intelligence from phylogenetic, to ontogenetic, to communal, and to artificial intelligence, we will try to shed light on how to understand precisely what the seemingly dramatic advancement in machine intelligence in the past decade has truly accomplished. This includes to provide a principled mathematical explanation to the practice of deep learning from the perspective of compressive data encoding and decoding. This reveals limitations of the current practice and suggests natural ways to develop more correct and complete learning systems. Eventually, we will clarify the difference and relationship between knowledge and intelligence, which may guide us to pursue the goal of developing truly autonomous intelligent systems.
Human intelligence constantly inspires the development of AGI. While human intelligence manifests in our complex interaction with the world, we propose that the evaluation of AGI should be rooted in dynamic embodied physical and social interactions (DEPSI). Specifically, we propose the importance of conducting evaluations targeting a cognitive infrastructure rooted in the theory of mind that enables communicative learning, a U space describing the physical laws and social norms that agents master in reality, and a V space describing the value system. Furthermore, we propose three critical characteristics to be considered as AGI benchmarks and to be evaluated in TongTest: performing infinite tasks, self-driven task generation, and value alignment. The Tong test describes a value- and ability-oriented testing system that delineates five levels of AGI milestones through a virtual environment with DEPSI, allowing for infinite task generation. We contrast TongTest with classical AI testing systems in various aspects and propose a systematic evaluation system to promote standardized, quantitative, and objective benchmarks and evaluation of AGI. In addition, we build a prototype AGI agent called “Little Girl Tong-tong” to participate in TongTest, and the results show that TongTest can benchmark AGI agents from various dimensions.
Molecular clouds (MCs) condense out of the warm interstellar medium (ISM) on scales of several 100 pc and host filamentary substructures on sub-pc scales. They consist of molecular hydrogen (H2), which can only be traced indirectly in observations, and are subject to supersonic turbulence. In the SILCC project, we investigate how a multi-phase interstellar medium (ISM) is shaped and stirred by feedback from massive stars. We carry out 3D magneto-hydrodynamic simulations with the FLASH code, for which we developed new physical models for gravity, radiation transport, and chemistry. In SILCC, we compare the impact of the explosion sites of supernovae and find that a significant fraction needs to explode in low-density environments to develop an ISM that is in agreement with observations of the Milky Way. Furthermore, we include star cluster formation using sink particles, as well as the massive star feedback from these clusters in the form of ionizing radiation and stellar winds. With these high-performance computer simulations, we can show that the early feedback by massive stars limits the accretion of fresh gas, i.e. the growth of the star-forming MCs, and thus regulates the overall star formation efficiency. In novel zoom-in simulations based on adaptive mesh refinement, we can now investigate the process of molecular cloud formation in high resolution.
The astronomy research is undergoing a profound transformation as more and more big data is generated by large survey telescopes. The influence of data-driven methods on scientific discovery has penetrated into all fields of astronomy research, and the new idea of AI for Science represented by artificial intelligence technology has brought new opportunities for scientific research. This presentation will introduce the requirements of astronomical big data processing against the background of data generated by large survey telescopes around the world. The presentation will also show some exploration in combination of the multiband survey data with the multi-modal AI, which will greatly improve data mining capabilities and change research paradigms in astronomical data mining.
Earth system science stands at the forefront of 21st-century earth science, similar to the revolutionary breakthroughs introduced by plate tectonic theory in the last century. However, it has yet to establish a unified theoretical model or define its core scientific challenges. Unlike the movement paradigms established by key discoveries like continental drift, mid-ocean ridge rifting, and mantle convection in plate tectonics, earth system science grapples with greater complexity, involving interactions among multiple spheres: the atmosphere, hydrosphere, lithosphere, and biosphere. The primary challenges within this intricate system include elucidating the coupling mechanisms between these spheres, understanding the causes and precise predictions of extreme events, and exploring the long-term evolution of the earth system. Although modern science boasts vast observational data from fields such as remote sensing, meteorology, geology, and deep-sea exploration, critical areas like the deep Earth, deep-time and polar regions remain underrepresented. The complexity and multidimensionality of existing data further complicate processing and interpretation, with traditional analytical methods and models showing significant limitations in addressing nonlinear feedback relationships among these spheres. The integration of big data, large models, and high computing power offers a promising avenue for theoretical breakthroughs in earth system science. By combining and analyzing data from diverse sources and scales, and leveraging advanced technologies like machine learning and high-performance computing, scientists can construct more sophisticated integrated models. These models enhance understanding and predictive capabilities regarding multi-sphere interactions, capturing dynamic changes and revealing nonlinear coupling effects.
Density-functional theory (DFT) is a cornerstone of modern computational chemistry. Specifically, Kohn-Sham DFT greatly reduces computational costs by focusing on the real-space electron density rather than the many-body wavefunction in a higher-dimensional space. However, the catch is that the exchange-correlation (XC) functional in Kohn-Sham equation is unknown and needs to be approximated. This leads to the errors that are too large to be predictive. We proposed and developed two approaches to address this problem, i.e., Δ-learning technique to calibrate the results of conventional DFT, and machine learning (ML)-based XC functional. As approximated XC functional leads to systematic error, a simple ML model with a little extra information will be capable for calibrating the less precise results to better counterparts. Δ-learning is thus designed to learn this error with a small amount of data. Our pioneering work back in 2003 proposed a framework, proving that as simple as a one-hidden-layer neural network, together with several molecular descriptors, is enough for calibrating DFT-level results to experimental level. Later in 2022, we replaced the time-consuming DFT calculation with a graph neural network, enabling the prediction of experimental-level heat of formation in nearly no time. Δ-learning method has also been applied to calibrate photophysical properties, open circuit voltages of lithium-ion batteries, etc. Another solution to improve the accuracy of DFT is to find a better XC functional. In 2004, we refined the three hybrid parameters of B3LYP XC functional to make them depend on molecular descriptors included the number of electrons, dipole moment, quadrupole moment, kinetic energy, and spin multiplicity. The functional exhibits a remarkable alignment with experimental data. Thanks to the holographic electron density theory, in 2019, we further developed a deep-learning-based algorithm that maps quasi-local electron densities to the exact XC potential.
In the last few years, we have seen a tremendous amount of scientific progress made as a result of the AI revolution, both in our much expanded ability to make use of the fundamental principles of nature, and our much expanded ability to make use of experimental data and the literature. In this talk, I will start with the origin of the AI for Science revolution, review some of the major progresses made so far, and discuss how it will impact the way we do research. I will also discuss some of the ongoing projects that we are working on, with the objective of constructing a new set of infrastructure for scientific research.
This talk will provide a brief history of Science as a journal and will also give an overview of the Science family of journals, including Science Robotics. The talk will also delve into what journals like Science Robotics looks for in an article. It will also highlight the peer review process at the journal, as well as some tips to consider when preparing your work for submission to the journal.