Title: | Linear Algebra Computation in Data Science |
Speaker: | Prof. Zhaojun BAI, Department of Computer Science, University of California, USA |
Time/Place: | 11:30 - 12:30 FSC1217, Fong Shu Chuen Library, HSH Campus, Hong Kong Baptist University |
Abstract: | Linear algebra computation is one of the seven computational giants for massive data analysis, along with basic statistics, generalized N-body problems, graph-theoretical computations, optimization, integration and alginment problems [National Research Council, Frontiers in Massive Data Analysis, The National Academies Press, 2013]. In this talk, we will discuss a number of linear algebra problems arising from data analysis techniques. In particular, we will discuss Rayleigh-Quotient (RQ) and RQ-type optimizations in robust classification to handle uncertainty and in constrained partitioning to incorporate a prior information. We will show that many of these RQ and RQ-type optimizations can be reformulated as linear or nonlinear eigenvalue problems. The use of modern solvers for these eigenvalue problems will also be discussed. |
Title: | On an inverse problem in photolithography |
Speaker: | Prof. Luca RONDI, Department of Mathematics and Geosciences, University of Trieste, Italy |
Time/Place: | 16:30 - 17:30 SCT909, Cha Chi-ming Science Tower, HSH Campus, Hong Kong Baptist University |
Abstract: | The inverse photolithography problem is a key step in the production of integrated circuits. In this talk I present a regularization and computation strategy for this inverse problem. The method is based on a variational approach whose key feature is a regularization procedure for a suitable thresholding operation. The validity of the method is shown by a convergence analysis and by numerical experiments. This is a joint work with Fadil Santosa (University of Minnesota) and Zhu Wang (University of South Carolina). |
Title: | Learning methods for safely using unlabelled data |
Speaker: | Dr. Yufeng LI, National Key Laboratory for Novel Software Technology, Nanjing University, China |
Time/Place: | 11:00 - 12:00 FSC1217, Fong Shu Chuen Library, HSH Campus, Hong Kong Baptist University |
Abstract: | When the amount of labelled data is limited, it is usually expected that learning methods exploiting additional unlabelled data will help improve learning performance. In many situations, however, it is reported that learning methods using unlabelled data may even decrease the learning performance. This phenomenon affects the deployment of unlabelled data learning methods in real-world situations. It is thus desirable to develop “safe” unlabelled data learning methods that often improve performance, while in the worst cases do not decrease the learning performance. In this talk, I introduce our recent progresses in this direction. By considering the performance degeneration may be caused by not-good-enough solution in large-scale optimization, we present a scalable method which is able to optimize a tight convex upper bound of the objective in an efficient manner. By considering that the performance degeneration may be caused by the uncertainty of model selection, we present a learning method which optimizes the worst-case accuracy improvement and hence avoids the harm of uncertain model selection. Furthermore, by considering that real-world applications require variants of performance measures and the performance degeneration may be caused by the difficulty in optimizing complicate performance measures, we present to optimize the worst-case performance gain under complicate performance measures and show that when the performance measure is AUC, F1 and Top-k precision, the minimax convex relaxation of the objective could be solved globally and efficiently. Experiments validate the effectiveness of the proposal method. |
Title: | Numerical methods for uncertainty quantification for multiphase flow in porous media |
Speaker: | Prof. Rolf JELTSCH, ETH, Zurich, Switzerland |
Time/Place: | 11:30 - 12:30 FSC1217, Fong Shu Chuen Library, HSH Campus, Hong Kong Baptist University |
Abstract: | Multiphase flow models are used to understand the extraction of oil or gas in porous media or store Carbon Dioxide in empty reservoirs. Mathematically the models consist of a system of nonlinear partial differential equations. The phase pressure is modeled using an elliptic equation while the phase saturation is a hyperbolic system if capillary pressure is neglected. Many physical input variables, e.g. rock permeability, relative permeability are determined by measurement processes and hence are prone to uncertainty. Hence one needs to model uncertainty to be able to compute the effects. We report on a recently started research project in Brazil and review the current state of the art. |
Title: | SDNA: Stochastic Dual Newton Ascent for Empirical Risk Minimization |
Speaker: | Dr. Zheng QU, Department of Mathematics, The University of Hong Kong, Hong Kong |
Time/Place: | 11:00 - 12:00 FSC1217, Fong Shu Chuen Library, HSH Campus, Hong Kong Baptist University |
Abstract: | We propose and analyze a randomized block coordinate descent methods with overlapping blocks. In each iteration our method picks a random subset of the coordinates, following an arbitrary probability law, and maximizes, exactly, the objective restricted to the random subspace spanned by the coordinates. Equivalently, this can be seen as the solution of a proximal subproblem involving a random principal submatrix of the Hessian of the quadratic function. Hence, our method is capable of utilizing all curvature information available in the random subspace in which it operates, which leads to striking striking improvements in both theory and practice, sometimes by orders of magnitude. Note that this is very different from the update strategy of parallel/minibatch coordinate descent methods. Indeed, while these methods also update a random subset of variables in each iteration, they instead only utilize curvature information present in the diagonal of the Hessian. We then instantiate this method to derive a new algorithm-Stochastic Dual Newton Ascent (SDNA)-for solving regularized empirical risk minimization problem with smooth loss functions and strongly convex regularizer. |
Title: | 香港浸會大學數學系四十五週年校友講座 - 數學教師成【長】之路 |
Speaker: | 黃廣威校友 黃志揚校友, Hong Kong |
Time/Place: | 14:30 - 16:30 RRS905, Sir Run Run Shaw Building, HSH Campus, Hong Kong Baptist University |
Title: | HKBU MATH 45th Anniversary Distinguished Lecture - Dirichlet Forms and Applications |
Speaker: | Prof. Zhi-Ming Ma, Institute of Applied Mathematics, Chinese Academy of Sciences, China |
Time/Place: | 16:30 - 17:30 (Preceded by Reception at 4:00pm) WLB103, The Wing Lung Bank Building for Business Studies, Shaw Campus, Hong Kong Baptist University |
Abstract: | The theory of Dirichlet forms is a powerful mathematical framework which connects deterministic potential theory and stochastic analysis. In this talk I shall briefly introduce the notion of Dirichlet forms and review some of its applications. I shall also present some recent development and results in this research direction. |
We organize conferences and workshops every year. Hope we can see you in future.
Learn MoreProf. M. Cheng, Dr. Y. S. Hon, Dr. K. F. Lam, Prof. L. Ling, Dr. T. Tong and Prof. L. Zhu have been awarded research grants by Hong Kong Research Grant Council (RGC) — congratulations!
Learn MoreFollow HKBU Math