Upcoming Lectures

图片

Upcoming Lectures

当前位置: Home >> Upcoming Lectures >> 正文

Lecture 9 On Presuppositions of Machine Learning: A Best-fitting Theory

如何突破机器学习的先验假设?

日期: 2022-10-24 点击:

Speaker Bio


Zongben Xu was born in 1955. He received his Ph.D. degrees in mathematics from Xi’an Jiaotong University, China, in 1987. His current research interests include applied mathematics and mathematical methods of big data and artificial intelligence. He established the L(1/2) regularization theory for sparse information processing. He also found and verified Xu-Roach Theorem in machine learning, and established the visual cognition based data modelling principle, which have been widely applied in scientific and engineering fields. he initiated several mathematical theories, including the non-logarithmic transform based CT model, and ultrafast MRI imaging, which provide principles and technologies for the development of a new generation of intelligent medical imaging equipment. He is owner of the Tan Kan Kee Science Award in Science Technology in 2018, the National Natural Science Award of China in 2007and winner of CSIAM Su Buchin Applied Mathematics Prize in 2008. He delivered a 45-minute talk on the International Congress of Mathematicians 2010. He was elected as member of Chinese Academy of Science in 2011.

Zongben Xu was the vice-president of Xi’an Jiaotong University. He currently makes several important services for government and professional societies, including the director for Pazhou Lab (Guangzhou), director for the National Engineering Laboratory for Big Data Analytics, a member of National Big Data Expert Advisory Committee and the Strategic Advisory Committee member of National Open Innovation Platform for New Generation of Artificial Intelligence.

徐宗本,中国科学院院士,数学家、信号与信息处理专家、西安交通大学教授。主要从事智能信息处理、机器学习、数据建模基础理论研究。曾提出稀疏信息处理的L(1/2)正则化理论,为稀疏微波成像提供了重要基础;发现并证明机器学习的“徐-罗奇”定理, 解决了神经网络与模拟演化计算中的一些困难问题,为非欧氏框架下机器学习与非线性分析提供了普遍的数量推演准则; 提出基于视觉认知的数据建模新原理与新方法,形成了聚类分析、判别分析、隐变量分析等系列数据挖掘核心算法, 并广泛应用于科学与工程领域。曾获国家自然科学二等奖、国家科技进步二等奖、陕西省最高科技奖; 国际IAITQM 理查德.普莱斯(Richard Price)数据科学奖; 中国陈嘉庚信息技术科学奖、中国CSIAM苏步青应用数学奖;曾在2010年世界数学家大会上作45分钟特邀报告。

曾任西安交通大学副校长,现任人工智能与数字經济广东省实验室(琶洲实验室)主任、西安数学与数学技术研究院院长、陕西国家应用数学中心主任、大数据算法与分析技术国家工程实验室主任,是国家大数据专家咨询委员会委员、国家新一代人工智能战略咨询委员会委员。


Abstract

Machine learning have been applied with a set of prerequisite or hypotheses, the optimal setting of which is a `the chicken or the egg’ problem. Those hypotheses include in particular (i) the Large Capacity Hypothesis on hypothetical space, (ii) the Independence Hypothesis on loss function, (iii) the Completeness Hypothesis on training data, (iv) the Prior-Determine-Regularizer Hypothesis on regularization terms, and (v) the Euclidean Hypothesis on analysis framework. We analyze the role, effect and limitations of those hypotheses in this talk, and propose a systematic way, could named as a best-fitting theory, to break through each of the hypotheses.

More specifically, we propose the model driven deep learning approach to burst the Large Capacity Hypothesis, develop a noise modeling principle to breach the Independence Hypothesis, suggest the axiomatic curriculum/self-paced learning approach for the Completeness Hypothesis, the implicit regularization method for the Prior-Determine-Regularizer Hypothesis, and Banach space geometry for the Euclidean Hypothesis. In each case, we show the best-fitting strategy, substantiate the value and outcome of the breaking though. We show also that the continuing effort for bursting the hypotheses of ML is needed, which is then opening new hot directions of ML research.  

机器学习是人工智能的最基础、最核心技术(算法),但机器学习的执行通常都是以一组基本的先验假设为前提的,这些基本假设包括: 假设空间的大容量假设、训练数据的完备性假设、损失度量的数据独立假设、正则项的先验决定假设、分析框架的欧氏空间假设等。本报告分析这些假设的作用、局限及其影响,提出突破这些基本假设的可能途径与方法。特别,我们提出突破假设空间大容量假设的模型驱动深度学习方法、突破训练数据完备性假设的课程-自步学习方法、突破损失度量数据独立假设的误差建模原理、突破正则项先验决定假设的隐正则化方法、突破分析框架欧氏空间假设的Banach空间几何方法。每一情况下,我们举例说明新突破带来新价值。所有这些尝试构成机器学习的适配性理论,是当下机器学习研究的一个新方向。