Scaling Law-based Functional Clustering: A Statistical Model for Compressing Large Multimodal Models

Time:10:30 - 12:30, 2024-06-08

Venue:ZOOM: 637 734 0280 PW: BIMSA

Speaker:Rongling Wu (邬荣领 BIMSA, YMSC)


Models across different application domains are exploding in size. This disproportionate increase in model sizes will quickly obstruct the sustainability of deep learning. From my personal view, functional clustering, a model derived to classify a population of variables into distinct modules, can serve as an alternative tool to compress and optimize large multimodal models in any size. I will be presenting this argument, assuming that the audiences have blank background.

Speaker Intro

Rongling Wu, received a Ph.D. in Quantitative Genetics from the University of Washington (Seattle) in 1995. He was a Distinguished Professor of Statistics and Public Health Sciences at Pennsylvania State University, and Director of the Center for Statistical Genetics. He is currently the Zeng Siming Chair Professor of Yau Mathematical Sciences Center, Tsinghua University. He is also a researcher at Yanqi Lake Beijing Institute of Mathematical Sciences and Applications, and also serves as editor-in-chief, associate editor, special editor and editorial board member of several journals in the fields of genetics, bioinformatics and computational biology. He was selected as a fellow of the American Association for the Advancement of Science and the American Statistical Association, and won the Distinguished Researcher Award of the American Institute of Applied Mathematics and Statistics (SAMSI), the University of Florida Research Fund Professor Award, the Pennsylvania State University Distinguished University Professor Award, and the Floyd Science Innovation Award. Research interests include: developing interdisciplinary statistical methods to reveal the genetic control mechanisms of complex traits and human complex diseases. The proposed functional mapping method can effectively discover the genetic rules of trait development and describe the key patterns of gene effects changing over time and space. Combining functional mapping with evolutionary game theory, scale theory, and prey-predator theory, a series of computational methods have been developed to construct multi-level, multi-space, and multi-scale genotype-phenotype relationships from molecules to phenotypes The three-dimensional network provides analysis tools for systems biology, systems medicine, and systems pharmacology research. Published more than 400 SCI papers in important international journals such as Nature Reviews Genetics, Nature Communications, PNAS, Journal of the American Statistical Association, Annals of Applied Statistics, Physics of Life Reviews, Physics Reports, Briefings in Bioinformatics, Cell Reports, Evolution, etc. The research results have been cited or highlighted by important journals such as Science and Cell.

DATEJune 7, 2024
Related News
    • 0

      Model Selection for Optimal Regression Learning

      In statistical learning, various mathematical optimalities are used to characterize performances of different learning methods. They include minimax optimality from a worst-case standpoint and asymptotic efficiency from a rosy view that the regression function to be learned sits there to be discovered. When multiple models, e.g., trees, neural networks and support vector machines, are considere...

    • 1

      Large cardinals and small sets: The AD+ Duality Program

      Speaker: Hugh WoodinW. Hugh Woodin is a Distinguished Professor Emeritus of Mathematics at the University of California, Berkeley and a Professor of Philosophy and of Mathematics at Harvard University. Woodin is a leading set theorist. He has made many notable contributions to the theory of inner models and determinacy, to functional analysis, and recursion theory. A type of large cardinal, the...