Interpretation of KDD2016 Papers (1)

Joint compilation: Zhang Min, Gao Fei, Chen Yang Yingjie

Introduction: KDD2016 is the premier interdisciplinary conference gathering researchers and practitioners in data science, data mining, knowledge discovery, large-scale data analysis and big data.

Paper 1 title: Stable fluid approximation convolutional neural network

Summary

For aerodynamically relevant design, analysis, and optimization problems, the flow field is simulated using a computational fluid dynamics (CFD) solver. However, CFD simulation is usually an iterative process that is computationally expensive, memory demanding, and time consuming. These shortcomings of CFD have limited the opportunities for design space exploration, and at the same time they have broken the idea of ​​interactive design. We propose a general and flexible approximation model for real-time prediction of non-uniform steady-state layer fluids in two- or three-dimensional domains based on convolutional neural networks (CNNs). We explored alternatives for geometric representation and network architecture for CNNs. We show that the convolutional neural network estimates the speed of the velocity field, which can be two orders of magnitude faster than a GPU-accelerated CFD solver, and four orders of magnitude faster than a CPU-based CFD solver (at the expense of very low errors). This approach provides immediate feedback on real-time interaction design during the early stages of design. CNNs ensure an effective assessment of the entire velocity field compared to existing aerodynamic models. Further, designers and engineers can directly apply the CNN approximation model to their design space exploration algorithms without the need to train additional, lower-dimensional alternative models.

Keywords : convolutional neural network; surrogate model; computational fluid dynamics; machine learning

Author introduction

(1) Xiaoxiao Guo

School: University of Michigan - University of Michigan

Research area: Machine Learning Artificial intelligence

More related field papers:

· Deep learning for real-time Atari game play using offline Monte-Carlo tree search planning(2014 )

· Action-conditional video prediction using deep networks in atari games(2015)

· Deep Learning for Reward Design to Improve Monte Carlo Tree Search in ATARI Games(2016)

(2) Wei Li

Organization: Autodesk Research

Research direction: Human-computer interaction mapping, artificial machine intelligence, cross-science

More related papers:

· Automated Extraction of System Structure Knowledge from Text (IDETC / CIE 2016 Conference)

· Four-Bar Linkage Synthesis Using Non-Convex Optimization (CP 2016 Conference)

(3) Francesco Iorio

Organization: Autodesk Research

Research directions: Software scalability, hybrid computing, accelerator systems on various high-performance platforms, general-purpose design, and parallel algorithm development on a variety of core and multi-core systems.

More related papers:

· Four-Bar Linkage Synthesis Using Non-Convex Optimization (CP 2016 Conference)

· Automated Extraction of System Structure Knowledge from Text (IDETC / CIE 2016 Conference)

Paper Two: Structured Double Random Matrices Based on Image Clustering

Summary

As one of the most important topics for machine learning research, clustering algorithms have been widely used in many different fields. At present, the widespread application of clustering algorithms in scientific research and industrial practice has attracted a great deal of attention. Researchers have proposed a large number of clustering algorithms, of which the most important one is the use of similarity matrix based image clustering methods. Recent research has used dual random matrices to normalize input similarity matrices and to improve image-based clustering models. Although this dual random matrix can improve the performance of the clustering model, its clustering structure is not as clear as expected. Therefore, post-processing steps are needed to extract the final clustering results, and the results may not be ideal. In order to solve this problem, we propose a new convex model, a low-rank constraint Laplacian matrix, and use this new model to learn structured double random matrices. Our new structured double random matrix can directly display the clustering structure and encode the probability of paired data points to be linked together, thereby improving the clustering results. We have obtained an efficient optimization algorithm to achieve new research goals. In addition, when the input information is different, we can use the method to get in touch with the K-means and the spectrum cutting model respectively. In the article, we will also discuss the topic at the theoretical level. We conducted experimental studies based on synthetic and benchmark data sets to validate the performance of our proposed method. The empirical research results prove that our model provides methods and approaches to better solve the K-means clustering problem. The clustering indication provided by our model is used as the initial value, and the K-means convergence is equal to a smaller objective function value for better clustering performance. In addition, the performance of the clustering performance of our model is compared with the performance of the general clustering method and the associated dual-random model. In all data sets, the performance of our model can be comparable to other related models, and even better than other models.

Keywords: double random matrix; image Laplacian matrix learning, K-means clustering algorithm, spectral clustering algorithm

Author introduction

(1) Xiaoqian Wang

School: Doctor of Computer Science and Engineering, University of Texas at Arlington, USA

Main research areas: Machine Learning and Computer Biology

Related academic achievements:

· Clustering and projected clustering with adapted neighbors. (Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 2014. P977-986.)

· Higher-order dispersion compensation to enable a 3.6 μs wavelength-maintaining delay of a 100 Gb/s DQPSK signal (Optics letters. 2010. p. 2985-2987.)

· The Contrained Laplacian Rank Algorithm for Graph-Based Clustering (Thirtieth AAAI Conference on Artificial Intelligence.)

(2) Feiping Nie, Nie Feiping,

School: Professor and PhD, Department of Computer Science and Engineering, University of Texas at Arlington, USA

Research direction: Research and development of machine learning and related application fields (pattern recognition, data mining, computer vision, image processing, information retrieval, etc.).

Related academic achievements:

· New Primal SVM Solver with Linear Computational Cost for Big Data Classifications (The 31st International Conference on Machine Learning (ICML), 2014.)

· Optimal Mean Robust Principal Component Analysis (The 31st International Conference on Machine Learning (ICML), 2014.)

· Robust Distance Metric Learning via Simultaneous L1-Norm Minimization and Maximization (The 31st International Conference on Machine Learning (ICML), 2014.)

(3)Heng Huang

School: Professor of Computer Science and Engineering, University of Texas at Arlington, Adjunct Professor of Clinical Sciences, University of Texas Southwestern Medical Center, Major Research Interests: Big Data Mining, Machine Learning, Data Science, Bioinformatics, Systems Biology, Neuroinformatics, brain science, computer vision, medical image analysis, computational sustainability, natural language processing, etc.

Related academic achievements:

· Structured Doubly Stochastic Matrix for Graph Based Clustering (22nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2016), Research Track, accepted to appear.)

· Robust and Effective Metric Learning Using Capped Trace Norm (22nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2016), Research Track, accepted to appear.)

· New Multi-Task Learning Model to Predict Alzheimer's Disease Cognitive Assessment (18th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2016), accepted to appear.)

Paper 3 Topic: Structured Deep Web Embedding

Summary

Network embedding is an important method to learn the low-dimensional representation of the vertices of the network, aiming to capture and preserve the network structure. Almost all existing network embedding methods use shallow models. However, due to the complexity of the underlying network structure, the shallow model cannot capture a highly nonlinear network structure, resulting in a sub-optimal network representation. Therefore, how to find a method that can effectively capture a highly nonlinear network structure while preserving global and local structures is an open but important issue. In order to solve this problem, this paper proposes a structured deep network embedding method, namely, SDNEL. More specifically, we first propose a semi-supervised deep model that has a multilayer nonlinear function so that it can capture highly nonlinear network structures. Then, we propose to use the first-order approximation and the second-order approximation to preserve the network structure. The second-order approximation is used by unsupervised components to capture the global network structure. The first-order approximation is used as supervision information in the supervising component and is used to preserve the local network structure. By co-optimizing in a semi-supervised deep model, our method can preserve both local and global network structures and is robust to sparse networks. Based on experience, we conducted experiments in five real-world networks, including a language network, a citation network, and three social networks. The results show that compared to the benchmark, our method can obviously reconstruct the original network better, and there are continuous gains in three applications, namely multi-label classification, link prediction and visualization.

Keywords: network embedding, deep learning, network analysis

The author of the paper introduces:

(1) Daixin Wang (王岱鑫)

School: Department of Computer, Tsinghua University

(2) Peng Cui ( Cui Peng)

School: Associate Professor, Department of Computer Science, Tsinghua University

Research direction: machine learning, data mining, social media, etc.

(3) Wenwu Zhu (Zhu Wenwu)

School: Tsinghua University Computer Department National "Thousand Person Plan" Specially-appointed professor, doctoral tutor, deputy director of the Department, the national special expert, the national 973 project chief scientist. Previously, he was a Principal Investigator at Microsoft Research Asia, a Chief Scientist at Intel China Research Institute, and a researcher at Bell Labs in the United States. IEEE Fellow, SPIE Fellow, ACM Distinguished Scientist.

Research direction: Three-dimensional space big data computing, social multimedia computing, multimedia cloud computing, future multimedia communications and networking.

Paper Download:

Paper One

Paper 2

Paper 3

Via:KDD2016 accepted-papers





VCM Parts

VCM (Voice Coil Motor), the Voice Coil Motor in electronics, is a kind of Motor. Because the principle and loudspeaker similar, so called voice coil motor, with high frequency, high precision characteristics.

Its main principle is in a permanent magnetic field, by changing the motor coil DC current size, to control the stretching position of the spring, so as to drive up and down movement.

VCM is widely used in mobile phone cameras to realize the function of autofocus. Through VCM, the position of the lens can be adjusted to present clear images.

VCM is include of Shield Case, Frame, F.Spacer, F.Spring, Yoke, Magnet, Coil, Carrier, B.Spacer, B.Spring, Base.

Components Vcm Gasket,Components Vcm Gaskets,Camera Vcm Terminals,No Notch Terminal

SHAOXING HUALI ELECTRONICS CO., LTD. , https://www.cnsxhuali.com