Date/Time : 17 Jun 2021 10:00
Venue : via Zoom
Speaker : Rui Wang, University of California at San Diego

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd Modelling complex physical dynamics is a fundamental task in science and engineering. Traditional physics-based models are interpretable but usually require significant computational resources and expertise. While deep learning (DL) provides alternatives for efficiently emulating nonlinear dynamics, it does not necessarily obey the physics laws or generalize well across different systems. Thus, there is a growing need for incorporating prior physics knowledge into DL models so that we can take the best of both types of approaches. We present three physics-guided DL approaches for spatiotemporal forecasting. We first present TF-Net, which marries the numerical techniques with custom-designed neural networks, that can predict accurate physical fields that obey desirable physical characteristics. We also provide two approaches to improve generalization across different systems. One is to design deep equivariant models that respect given symmetries of physical dynamics, and the other one is to infer shared time-invariant characteristics by leveraging meta-learning techniques.

Add to calendar :

Date/Time : 03 Jun 2021 16:00
Venue : via Zoom
Speaker : Liu Yong, Institute of High Performance Computing

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd Transformer has been developed as one of most impactful deep neural network architecture for Natural Language Processing (NLP) tasks such as machine translation, question & answer, and reading comprehension. Its effectiveness has been demonstrated by the impressive performance of large language model such as GPT-3 to produce convincing stream of human-like text. This talk will cover the mechanism of transformer and how it has been used for computer vision tasks such as image classification, followed by our recent work on transformer-based medical image segmentation for multiple diseases. Compared with representative existing methods, the proposed method consistently achieved the highest segmentation accuracy, and exhibited good cross-domain generalization capabilities.

Add to calendar :

Date/Time : 27 May 2021 16:00
Venue : via Zoom
Speaker : Niangjun Chen, Singapore University of Technology and Design

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd Many problems in data science and cloud computing can be cast as online optimization problems. In this talk, we study general online optimization problems with uncertainties, where the payoff or cost information is only known to the decision-maker up to some bounded uncertainty set. We develop an algorithmic framework, CRT, that can be applied to both online cost minimization problems with revenue constraints and online revenue maximization problems with inventory constraints. In both scenarios, we design online algorithms based on the CRT framework and prove their optimality in terms of competitive ratio. Furthermore, while prior works on online nonlinear optimization with inventory constraint require the revenue functions to be concave, we show that CRT-based algorithms have optimal competitive ratios in general online optimization settings with non-convex cost functions or non-concave revenue functions. We demonstrate our proposed algorithms’ practical performance via trace-driven simulations in a cloud computing application against prominent existing algorithms.

Add to calendar :

Date/Time : 20 May 2021 16:00
Venue : via Zoom
Speaker : Yuan FANG, Singapore Management University

JJoin Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd Heterogeneous Information Networks (HIN) often model complex graph-structured data with multi-typed relationships between different types of objects. The heterogeneity gives rise to the rich semantics on HINs, such as different user interactions on social networks, and different functional associations between proteins on a biological network. Metagraphs, which model recurring subgraph structures on HINs, emerge as a powerful tool to capture the rich semantics. Compared to their simpler variant metapaths, metagraphs are higher-order structures with stronger expressiveness and can be regarded as nonlinear functions of metapaths. Not surprisingly, metagraphs can be used to derive effective graph representations on HINs, to enable various downstream tasks such as node classification and link prediction. The efficacy of metagraph-based representations can be demonstrated in various domains including social, cyberphysical and biological networks.

Add to calendar :

Date/Time : 22 Apr 2021 16:00
Venue : via Zoom
Speaker : Anqi Qiu, National University of Singapore

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd The Laplace-Beltrami operator is a generalization of the Euclidean representation of the Laplace operator to an arbitrary Riemannian manifold. It is a self-adjoint operator and its eigenfunctions form a complete set of real-valued orthonormal basis functions. In this talk, I will introduce spectral Laplace-Beltrami wavelets and its computational algorithm. I will then demonstrate its use for smoothing and classification of the data defined on smooth surfaces embedded in the 3-D Euclidean space. Furthermore, I will discuss that the spectral Laplace-Beltrami Wavelets can be used for the construction of geometric convolutional neural network (CNN) and then introduce a vertex-based geometric CNN algorithm for regular surfaces in which translation and downsampling on surfaces can be the same as those in the regular grid. I will show the use of this method for the prediction of Alzheimer’s Disease.

Add to calendar :

Date/Time : 08 Apr 2021 16:00
Venue : via Zoom
Speaker : Yinping YANG, Institute of High Performance Computing, A*Star, Singapore

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd What are social technologies? Why they are useful in understanding people in the digital era? How do these technologies help to understand how the population respond to the pandemic throughout its development? This joint presentation will introduce A*STAR’s affective and social research and advanced emotion analysis technologies, and discuss how they are developed and put in good use in a range of practical industry applications, including the context of COVID-19. We will first review conceptual concepts surrounding human emotions, machine learning, natural language processing, and system architecture that lay the foundation of our social technologies. We will then share a few key learning points from our COVID-19 social media analytics studies in collaboration with communication scientists and public health practitioners. Among our first published findings at JMIR Public Health and Surveillance, we discovered that at a global scale, public emotions shifted strongly from fear to anger over the course of the pandemic, while sadness and joy also surfaced. Our presentation will be concluded with a few key takeaways and discussion for future research.

Add to calendar :

Date/Time : 01 Apr 2021 16:00
Venue : via Zoom
Speaker : Bo AN, Nanyang Technological University

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd In January 2017 CMU’s Libratus system beat a team of four top-10 headsup no-limit specialist professionals, which was the first time an AI had beaten top human players in this game. Libratus’s success is purely based on algorithms for solving large scale games and has nothing to do with deep learning! Over the last few years, algorithms for solving large scale games have also been applied to many domains such as security, sustainability, ad-word auction, and e-commerce. For some complex domains with strategic interaction, reinforcement learning is also used to learn an efficient policy. This talk will discuss key techniques behind these success and their applications in domains including games, security, e-commerce, and urban planning.

Add to calendar :

Date/Time : 25 Mar 2021 16:00
Venue : via Zoom
Speaker : Joe Kileel, University of Texas at Austin

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd Tensors are higher-order matrices, and decomposing tensors can reveal structure in datasets. In recent years, tensor decomposition has found applications in statistics, computational imaging, signal processing, and quantum chemistry. In this talk, we will present a new numerical method for low-rank symmetric tensor decomposition, building on the usual power method and ideas from classical algebraic geometry. The approach achieves a speed-up over the state-of-the-art by roughly one order of magnitude. We will also describe an “implicit” variant of the algorithm for the case of moment tensors which avoids the explicit formation of higher-order moments, analogously to matrix-free techniques in linear algebra. Time permitting, we will include a quantitative analysis of the non-convex optimization landscape associated with our algorithm. This is based on joint works with Joao Pereira, Tammy Kolda and Timo Klock.

Add to calendar :

Date/Time : 18 Mar 2021 16:00
Venue : via Zoom
Speaker : Yifei Wang, Stanford University

Join Zoom Meeting https://nus-sg.zoom.us/j/88213418159?pwd=K2lPSXEvemZGSXhiVGYxTVUzQ0d3QT09 Meeting ID: 882 1341 8159 Password: pcasvd Abstract We introduce a framework for Newton’s flows in probability space with information metrics, named information Newton’s flows. Extending the relationship between overdamped Langevin dynamics and Wasserstein gradient flows of Kullback-Leibler (KL) divergence, we derive Newton’s Langevin dynamics from Wasserstein Newton’s flows. We design sampling efficient variational methods in affine models and reproducing kernel Hilbert space (RKHS) to approximate Wasserstein Newton’s directions. Convergence results of the proposed information Newton’s method with approximated directions are established. Several numerical examples from Bayesian inference problems are shown to demonstrate the effectiveness of the proposed method.

Add to calendar :