2016-17

Abstract

Scientific applications are generating an ever-increasing volume of multi-dimensional data that are largely processed inside distributed array databases and frameworks. Similarity join is a fundamental operation across scientific workloads that requires complex processing over an unbounded number of pairs of multi-dimensional points. In this talk, we introduce a novel distributed similarity join operator for multi-dimensional arrays. Unlike immediate extensions to array join and relational similarity join, the proposed operator minimizes the overall data transfer and network congestion while providing load-balancing, without completely repartitioning and replicating the input arrays. We define formally array similarity join and present the design, optimization strategies, and evaluation of the first array similarity join operator. Meanwhile, if the data are rapidly updated, the join result can be considered as a view, which is defined by the similarity join. We model the process as incremental view maintenance with batch updates and give a three-stage heuristic that finds effective update plans. Moreover, the heuristic repartitions the array and the view continuously based on a window of past updates as a side-effect of view maintenance. We design an analytical cost model for integrating materialized array views in queries. A thorough experimental evaluation confirms that the proposed techniques are able to incrementally maintain a real astronomical data product in a production pipeline.

Biography

Weijie Zhao is a PhD student in the EECS graduate group at UC Merced, working with Prof. Florin Rusu. He received his BS from East China Normal University in Shanghai. His research interests include databases and scientific data management. Weijie is an avid computer programming contestant.

Abstract

In the first part of this talk, we present a sematic-aware image editing algorithm for automatic sky replacement. The key idea of our algorithm is to utilize visual semantics to guide the entire process including sky segmentation, search and replacement. First, we train a deep convolutional neural network for semantic scene parsing, which is used as visual prior to segment sky regions in a coarse-to-fine manner. Second, in order to find proper skies for replacement, we propose a data-driven scheme based on semantic layout of the input image. Finally, to re-compose the stylized sky with the original foreground naturally, an appearance transfer method is developed to match statistics locally and semantically. We show that the proposed algorithm can automatically generate a set of visually pleasing and realistic results. In the second part, a work on learning image filters for low-level vision is presented (e.g., edge-preserving filtering and denoising), in which a unified hybrid neural network is proposed. The network contains several spatially variant recurrent neural networks (RNN) as equivalents of a group of distinct recursive filters for each pixel, and a deep convolutional neural network (CNN) that learns the weights of RNNs. The proposed model does not need a large number of convolutional channels nor big kernels to learn features for low-level vision filters. Experimental results show that many low-level vision tasks can be effectively learned and carried out in real-time by the proposed algorithm.

Biography

Yi-Hsuan Tsai (https://sites.google.com/site/yihsuantsai/) received the B.S. in Electronics Engineering from National Chiao-Tung University, Hsinchu, Taiwan and the M.S. in Electrical Engineering and Computer Science from University of Michigan, Ann Arbor. He is currently working toward the PhD advised by Prof. Ming-Hsuan Yang at UC Merced and is the recipient of the Graduate Dean's Dissertation Fellowship in 2016. He was invited to attend the doctoral consortium at the IEEE Conference on Computer Vision and Pattern Recognition in 2016. His research interests include computer vision, computational photography and machine learning with the focus on visual object recognition and image editing. He also did research internships at Qualcomm Research, Max Planck Institute and Adobe Research.
Sifei Liu (http://www.sifeiliu.net/) is a Ph.D candidate in Electrical Engineering and Computer Science with Prof. Ming-Hsuan Yang. Her research interests are in computer vision and machine learning. She completed her M.C.S. at University of Science and Technology of China (USTC) under Stan.Z Li and Bin Li, and received the B.S. in control science and technology from North China Electric Power University. She received the Baidu fellowship in 2013. In 2013 and 2014, she was a intern at Baidu Deep Learning Institute. In addition, she was a visiting student at the Chinese University of Hong Kong in 2015. She was invited to attend the doctoral consortium at the IEEE Conference on Computer Vision and Pattern Recognition in 2016. Her research interests include computer vision, machine learning and computational photography.

Abstract

This talk is about cache, and more specifically, solid-state storage based cache for large-scale computing systems such as cloud computing and big data systems. With the increasing workload data intensity and increasing level of consolidation in such systems, storage is becoming a serous bottleneck. Emerging solid-state storage devices such as flash memory and 3D Xpoint have the potential to address this scalability issue by providing a new caching layer between main memory and hard drives in the storage hierarchy. However, solid-state storage has limited capacity and endurance, and needs to be managed carefully when used for caching. This talk will present several recent works done by the ASU VISA Research Lab for addressing these limitations and making effective use of solid-state caching.
First, the talk will introduce CloudCache, an on-demand cache allocation solution for understanding the cache demands of workloads and allocating the shared cache capacity efficiently. It is able to reduce a workload’s cache usage by 78% and the amount of writes sent to cache device by 40%, compared to traditional working-set-based approach. Second, the talk will present CacheDedup, an in-line cache deduplication solution that integrates caching and deduplication with duplication-aware cache replacement to improve the performance and endurance of solid-state caches. It can reduce a workload’s I/O latency by 51% and the amount of writes sent to cache device by 89%, compared to traditional cache management approaches. Finally, the talk will be concluded with a brief overview of the systems research at the VISA lab.

Biography

Ming Zhao is an associate professor of the Arizona State University (ASU) School of Computing, Informatics, and Decision Systems Engineering (CIDSE), where he directs the research laboratory for Virtualized Infrastructures, Systems, and Applications (VISA, http://visa.lab.asu.edu). His research is in the areas of experimental computer systems, including distributed/cloud, big-data, and high-performance systems as well as operating systems and storage in general. He is also interested in the interdisciplinary studies that bridge computer systems research with other domains. His work has been funded by the National Science Foundation (NSF), Department of Homeland Security, Department of Defense, Department of Energy, and industry companies, and his research outcomes have been adopted by several production systems in industry. Dr. Zhao has received the NSF Faculty Early Career Development (CAREER) award, the Air Force Summer Faculty Fellowship, the VMware Faculty Award, and the Best Paper Award of the IEEE International Conference on Autonomic Computing. He received his bachelor’s and master’s degrees from Tsinghua University, and his PhD from University of Florida.

Abstract

In the first part of this talk, we present a work for face parsing via conditional random field with unary and pairwise classifiers. We develop a novel multi-objective learning method that optimizes a single unified deep convolutional network with two distinct non-structured loss functions: one encoding the unary label likelihoods and the other encoding the pairwise label dependencies. Moreover, we regularize the network by using a nonparametric prior as new input channels in addition to the RGB image, and show that significant performance improvements can be achieved with a much smaller network size. Experiments show state-of-the-art and accurate labeling results on challenging images for real-world applications. In the second part, a work on video object segmentation is presented. It is a challenging problem due to fast moving objects, deformed shapes, and cluttered backgrounds. To obtain accurate segmentation across time, we propose an efficient algorithm that considers video segmentation and optical flow estimation simultaneously. For video segmentation, we formulate a principled, multi-scale, spatio-temporal objective function that uses optical flow to propagate information between frames. For optical flow estimation, particularly at object boundaries, we compute the flow independently in the segmented regions and recompose the results. We call the process "object flow" and demonstrate the effectiveness of jointly optimizing optical flow and video segmentation using an iterative scheme.

Biography

Sifei Liu (http://www.sifeiliu.net/) is a Ph.D candidate in Electrical Engineering and Computer Science with Prof. Ming-Hsuan Yang. Her research interests are in computer vision and machine learning. She completed her M.C.S. at University of Science and Technology of China (USTC) under Stan.Z Li and Bin Li, and received the B.S. in control science and technology from North China Electric Power University. She received the Baidu fellowship in 2013. In 2013 and 2014, she was a intern at Baidu Deep Learning Institute. In addition, she was a visiting student at the Chinese University of Hong Kong in 2015. She was invited to attend the doctoral consortium at the IEEE Conference on Computer Vision and Pattern Recognition in 2016. Her research interests include computer vision, machine learning and computational photography.
Yi-Hsuan Tsai (https://sites.google.com/site/yihsuantsai/) received the B.S. in Electronics Engineering from National Chiao-Tung University, Hsinchu, Taiwan and the M.S. in Electrical Engineering and Computer Science from University of Michigan, Ann Arbor. He is currently working toward the PhD advised by Prof. Ming-Hsuan Yang at UC Merced and is the recipient of the Graduate Dean's Dissertation Fellowship in 2016. He was invited to attend the doctoral consortium at the IEEE Conference on Computer Vision and Pattern Recognition in 2016. His research interests include computer vision, computational photography and machine learning with the focus on visual object recognition and image editing. He also did research internships at Qualcomm Research, Max Planck Institute and Adobe Research.

Abstract

Imagine computer displays become a space to augment human thinking. Essential human activities such as seeing, gesturing, and exploring can couple with powerful computational solutions using natural interfaces and accurate visualizations. In this talk, I will present research effort to quantify visualization techniques of all kinds. Our ongoing work includes research in: (1) perceptually accurate visualization – constructing a visualization language to study how to depict spatially complex fields in quantum-physics simulations and brain-imaging datasets; (2) using space to compensate for limited human memory – developing new computing and interactive capabilities for bat-flight motion analysis in a new metaphorical interface; and (3) extending exploratory metaphors to biological pathways to make possible integrated analysis of multifaceted datasets. During the talk, I will point to a number of other projects being carried out by my team. I will close with some thoughts on automating the evaluation of visualizations and venture that a science of visualization and metaphors now has the potential to be developed in full, and that its success will be crucial in understanding data-to-knowledge techniques in traditional desktop and immersive settings.

Biography

Jian Chen is an Assistant Professor in the Department of Computer Science and Electrical Engineering at the University of Maryland, Baltimore County (UMBC), where she leads the Interactive Visual Computing Lab (http:// ivcl.umbc.edu) and UMBC’s Immersive Hybrid Reality Lab (http://tinyurl.com/ ztnvdmf). She maintains general research interests in the design and evaluation of visualizations (encoding of spatially complex brain imaging, integrating spatial and non-spatial data, perceptually accurate visualization, and event analysis) and interaction (exploring large biological pathways, immersive modeling, embodiment, and gesture input). She has garnered best-paper awards at international conferences, and her work is funded by NSF, NIST, and DoD. She is also an UMBC innovation fellow and a co-chair of the first international workshop on the emerging field of Immersive Analytics. Chen did her post-doctoral research at Brown University jointly with the Departments of Computer Science (with Dr. David H. Laidlaw) and Ecology and Evolutionary Biology. She received her Ph.D. in Computer Science from Virginia Tech with Dr. Doug A. Bowman. To learn about Jian Chen and her work, please visit http:// www.csee.umbc.edu/~jichen.

Abstract

Communication is an intricate dance, an ensemble of coordinated individual actions. Imagine a future where machines interact with us like humans, waking us up in the morning, navigating us to work, or discussing our daily schedules in a coordinated and natural manner. Current interactive systems being developed by Apple, Google, Microsoft, and Amazon attempt to reach this goal by combining a large set of single-task systems. But products like Siri, Google Now, Cortana and Echo still follow pre-specified agendas that cannot transition between tasks smoothly and track and adapt to different users naturally. My research draws on recent developments in speech and natural language processing, human-computer interaction, and machine learning to work towards the goal of developing situated intelligent interactive systems. These systems can coordinate with users to achieve effective and natural interactions. I have successfully applied the proposed concepts to various tasks, such as social conversation, job interview training and movie promotion. My team's proposal on engaging social conversation systems was selected to receive $100,000 from Amazon Inc. to compete in the Amazon Alexa Prize Challenge.

Biography

Zhou Yu is a graduating PhD student at the Language Technology Institute under School of Computer Science, Carnegie Mellon University, working with Prof. Alan W Black and Prof. Alexander I. Rudnicky. She interned with Prof. David Suendermann-Oeft in ETS San Francisco Office on cloud based mulitmodal dialog systems in 2015 summer and 2016 summer. She also interned with Dan Bohus and Eric Horvitz in Microsoft Research on human-robot interaction in 2014 Fall. Prior to CMU, she received a B.S. in Computer Science and a B.A. in Linguistics from Zhejiang University in 2011. She worked with Prof. Xiaofei He and Prof. Deng Cai on Machine Learning and Computer Vision, and Prof. Yunhua Qu on Machine Translation.

Abstract

In this talk, I will first present the results of my first research project on autonomous driving as a PhD student in UC Merced. This work was conducted in collaboration with Berkeley Deep Drive (http://bdd.berkeley.edu/project/implicit-communication-through-motion). With progress in enabling autonomous cars to drive safely on the road, it is time to start asking how they should be driving. From the days of Alvinn to the latest autonomous driving technologies like NVidia’s Drive PX, researchers have used Learning from Demonstration to teach autonomous cars how to drive. Therefore, when it comes to customization of autonomous driving, a common answer is that the car should adopt the user’s style. In this project, we questioned this assumption and conducted user research in driving simulator to test our hypothesis. We found that users tend to prefer a significantly more defensive driving style than their own. Interestingly, they prefer the style they think is their own, even though their actual driving style tends to be more aggressive. The results show that conventional Learning from Demonstration algorithms will be inadequate for personalizing autonomous driving. In the second part of the talk I will discuss the implications of this result in greater detail and present some of the potential algorithms that we can use to augment user demonstrations.

Biography

Chandrayee Basu (http://chandrayee-basu.squarespace.com) is a 2nd year Ph.D. student in UC Merced advised by Prof. Mukesh Singhal, UC Merced and Prof. Anca Dragan, UC Berkeley. She is applying Human-Robot Interaction Algorithms to integrate human interaction into motion planning of autonomous cars. Chandrayee has multi-disciplinary research experience in design, applied machine learning, smart environment and human-robot interaction that she acquired as a Graduate student at UC Berkeley and Carnegie Mellon University.

Abstract

In this talk I will compare and contrast research vs startup innovation, based on my experiences at Stanford, the MIT Media Lab and Bay Area startups. I’ll discuss how the desired outcomes of each context encourages different kinds of risk and exploration, takeaways from my research experiences, and how we structure the early ideation process at Lemnos Labs where I am an Entrepreneur In Residence.

Biography

David Merrill is a technology executive and hardware startup founder with a computer science and human computer interaction background. His tactile learning system startup Sifteo - based on his Ph.D. from MIT - was acquired by drone-maker 3D Robotics in 2014 to become the kernel of a new consumer product group. At 3D Robotics he took various roles on the team that launched Solo: the Smart Drone in 2015, then he led R&D and IP. Alumnus of MIT, Stanford Computer Science and Symbolic Systems. TED speaker. Human-computer interaction expert. Drone builder. His work has been featured by the Discovery Channel, Popular Science, Wired, and the New York Times. Merrill is currently Entrepreneur in Residence at Lemnos Labs, an early-stage VC firm in San Francisco, working on the next project.

Abstract

In this talk I will present new approaches for achieving full-body motion coordination for humanoid virtual characters. I will first present a parametric data-based mobility controller with known coverage and validity characteristics, achieving flexible real-time deformations for locomotion control. I will then present a method for switching between different types of locomotion in order to navigate cluttered environments. The proposed method incorporates the locomotion capabilities of the character in the path planning stage, producing paths that address the trade-off between path length and locomotion behavior choice for handling narrow passages. In the last part, I will introduce a new approach for the coordination of locomotion with manipulation. The approach is based on a coordination model built from motion capture data in order to represent proximity relationships between the action and the environment. The result is a real-time controller that can autonomously produce environment-dependent full-body motion strategies. The obtained coordination model is successfully applied on top of generic walking controllers, achieving controllable characters which are able to perform complete full-body interactions with the environment.

Short Bio

Alain Juarez-Perez is a Ph.D. candidate in the Electrical Engineering and Computer Science graduate group of the University of California, Merced. His work is being developed at the Computer Graphics Lab under the advice of Prof. Marcelo Kallmann and has been supported by a UC-Mexus Doctoral Fellowship. He received his B.S. in Computer Science in 2012 from the University of Guanajuato, and in 2014 he was a visiting research assistant at the USC Institute for Creative Technologies. His research interests include Computer Animation, Data-Driven Algorithms, Motion Capture, Computational Geometry, Machine Learning, Computer Graphics and Motion Planning.

Abstract

In this talk, I will introduce my recent research projects on Internet of Things (IoT) security. First, I will introduce a physical layer authentication method for RFID tags. Second, I will talk about a fast and reliable protocol for authentication and key agreement among multiple IoT devices based on wireless signal information. Third, I will introduce an IoT data communication framework that guarantees data authenticity and integrity.

Biography

Chen Qian is an Assistant Professor in Department of Computer Engineering at University of California Santa Cruz. He was on the faculty of Computer Science at University of Kentucky 2013-2016. He got his PhD degree from The University of Texas at Austin, where he worked with Simon Lam. His research interests include computer networking and distributed systems, Internet of Things, network security, and cloud computing. He has published more than 60 papers, most which appeared in top journals and conferences including ToN, TPDS, ICNP, INFOCOM, ICDCS, SIGMETRICS, CoNEXT, CCS, NDSS, Ubicomp, etc.

Abstract

This seminar presents a brief and selected survey on the advances on stochastic distribution control, where the purpose of the controller design is to control the shape of output probability density functions (pdf) of non-Gaussian and general stochastic systems. This research was motivated through the requirement of distribution shape control of a number of practical systems. In recent years much research has been performed internationally and journal special issues and invited session at major conferences have been seen since 2001. It is expected that this seminar will provide with some up-to-date information on this new area.

Biography

Dr. Hong Wang joined PNNL in February 2016 as a Laboratory Fellow. He is based in the Controls team within the Electricity Infrastructure and Buildings Division of the Energy and Environment Directorate. Prior to joining PNNL, he was a full (chair) professor in process control at the University of Manchester in the U.K. Dr. Wang 's research interests are in advanced modelling and control for complex industrial processes, and fault diagnosis and tolerant control. He originated the research on stochastic distribution control where the main purpose of control input design is to make the shape of the output probability density functions to follow a targeted function. This area alone has found a wide spectrum of potential applications in modelling, data-mining, signal processing, optimization and distributed control systems design. Dr. Wang is the lead author of five books and has published over 300 papers in international journals and conferences. He is a member of three International Federation of Automatic Control Technical Committees and associate editor for IEEE Transactions on Control Systems Technology, IEEE Transactions on Automation Science and Engineering, and seven other international journals. He has been the associate editor of IEEE Transactions on Automatic Control, and has served as IPC member and conference chairman for many international conferences. Dr. Wang has also received several best paper awards from International Conferences including the best paper awards at Int. Conf. Control 2006, the Jasbar Memorial Prize for his outstanding contribution in the Science and Technology Development for paper industries in 2006, best theory paper award at World Congress on Intelligent Control and Automation in 2014 and one of the five finalists for the best application paper prize at 2014 IFAC World Congress. Dr. Wang holds a Ph.D. degree from Huazhong University of Science and Technology (HUST) in P R China.

Abstract

Robot grasp planning has been extensively studied in the last decades and it often consists of two different stages, i.e., where to grasp an object and how to measure the quality of a tentative grasp. Because these two processes are computationally demanding, form closure grasps are more widely used in practice instead of force closure grasps, even though the latter is in many cases preferable. In this talk, we introduce our framework to improve grasp quality evaluation. We accelerate the computation of the grasp wrench space, used to measure the grasp quality, by exploiting some geometric insights in the computation of convex hull. In particular, we identify a cutoff sequence to terminate the convex hull calculation with guaranteed convergence to the quality measure. Furthermore, we study how noise at each joint of the manipulator affects grasp quality. Different arm configurations will generate different noise distributions at the end-effector which have a huge impact in the robustness of grasping. In the last part of the talk, I will introduce a grasp planner taking into account the local geometry of the object to be grasped. In particular, for concave objects we exploit the fact that grasping at the concave region can make the grasp more robust. These insights are studied in theory and validated on an experimental platform.

Biography

Shuo Liu received his B.Sc. degree in computer engineer from University of Minnesota in 2012 and his B.Sc. degree in mathematics from Beijing Jiaotong University in 2012. In his Junior year of college, he participated in 2011 Robocup and won the champion title in the Middle Size League. He also participated in IROS 2016 Grasping and Manipulation Challenge and won 2nd place in automation track. Since August 2012 he is pursuing a Ph.D. degree in electrical engineering and computer science at the University of California, Merced, working with Dr.Stefano Carpin. His interests include manipulation, grasping, and computational geometry.

Abstract

Flying robots, such as multicopters, are increasingly becoming part of our everyday lives, with current and future applications including personal transportation, delivery services, entertainment, and aerial sensing. These systems are expected to be safe and to have a high degree of autonomy. This talk will discuss the dynamics and control of multicopters, including some research results on trajectory generation for multicopters and fail-safe algorithms. Finally, we will present the application of a failsafe algorithm to a fleet of drones performing as part of a live theatre performance on New York's Broadway.

Biography

Mark W. Mueller joined the mechanical engineering department at UC Berkeley in September 2016. He completed his PhD studies, advised by Prof. Raffaello D'Andrea, at the Institute for Dynamic Systems and Control at the ETH Zurich at the end of 2015. He received a bachelors degree from the University of Pretoria, and a masters from the ETH Zurich in 2011, both in Mechanical Engineering. http://www.me.berkeley.edu/people/faculty/mark-mueller