2016

Abstract

Increasingly, performance variability is an obstacle to understanding the throughput of large-scale supercomputers. Two runs of the same code, on the same system, may yield vastly different runtimes, depending on compiler flags, system noise, dynamic scheduling, and shared resources such as memory, filesystems and networks. Understanding an application's performance characteristics requires an increasingly large number of trial runs and measurements. Analyzing performance measurements from such runs is a data-intensive task. To address these issues, Livermore Computing is deploying Sonar, a "big data" cluster that will store and analyze performance data from LLNL’s entire HPC center. Sonar aggregates measurements from the network fabric, filesystem nodes, cluster nodes, applications. It will serve as a central data warehouse for measurements collected by tools. We will give an overview of the Sonar cluster and the tools we have integrated with it. We will also discuss some early techniques for analyzing performance data gathered from this system.

Biography

Todd is a computer scientist in the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory. His research focuses on scalable tools for measuring, analyzing, and visualizing performance data from massively parallel applications. Todd is also involved with many production projects at LLNL. He works with Livermore Computing’s Development Environment Group to build tools that allow users to deploy, run, debug, and optimize their software for machines with million-way concurrency. Todd received his Ph.D. in computer science from the University of North Carolina at Chapel Hill in 2009. His dissertation investigated parallel methods for compressing and sampling performance measurements from hundreds of thousands of concurrent processors. He received his B.A. in Computer Science and Japanese from Williams College in 2002. He has also worked as a software developer in Tokyo and held research internships at the University of Tokyo and IBM Research.

Abstract

Lawns make up the largest irrigated crop by surface area in North America, and carries with it a demand for over 9 billion gallons of freshwater each day. Despite recent developments in irrigation control and sprinkler technology, state-of-the-art irrigation systems do nothing to compensate for areas of turf with heterogeneous water needs. In this work, we overcome the physical limitations of the traditional irrigation system with the development of a sprinkler node that can sense the local soil moisture, communicate wirelessly, and actuate its own sprinkler based on a centrally-computed schedule. A model is then developed to compute moisture movement from runoff, absorption, and diffusion. Integrated with an optimization framework, optimal valve scheduling can be found for each node in the space. In a turf area covering over 10,000 square feet, two separate deployments spanning a total of 7 weeks show that MAGIC can reduce water consumption by 23.4% over traditional campus scheduling, and by 12.3% over state-of-the-art evapotranspiration systems, while substantially improving conditions for plant health. In addition to environmental, social, and health benefits, MAGIC is shown to return its investment in 16-18 months based on water consumption alone.

Biography

Daniel Winkler received his BS in Computer Science Engineering with honors from UC Merced in 2013. An ACM member, he since has been pursuing his PhD under advisement of Dr. Alberto Cerpa in UC Merced's ANDES Lab. Although his current research focuses on intelligent design and management of turf irrigation systems through the use of embedded devices, Daniel also has a growing interest in general resource management applications.

Abstract

In this talk, I will introduce our latest research in image scene understanding and interactive technologies. Our first line of research aims at rapid image scene understanding based on visual attention mechanism (IEEE TPAMI 2015, IEEE CVPR 2014 Oral). This is an area where people often have diverse feelings: some researchers believe that it is a principled research direction, while others might doubt its robustness. Instead of specific algorithm design, I would like to highlight how these algorithms can be robustly used in various applications, including image composition, photo montage, image retrieval, object detection, semantic segmentation, and even deep learning. Our second line of research aims at intelligent image manipulation mechanism. We try to explore smart image manipulation techniques for easily obtaining annotated data during users’ nature interaction with the real world (ACM TOG 2014, ACM TOG 2015), which is partially motivated by the growing requirement of high quality labeled training data (expensive to be collected) for scene understanding.

Biography

Ming-Ming Cheng is an associate professor with Department of Computer Science, Nankai University, China. He received his PhD degree from Tsinghua University, China, in 2012 under supervise of Prof. Shi-Min Hu, and working closely with Prof. Niloy Mitra. Then he spent 2 years as research fellow in UK, working with Prof. Philip Torr in Oxford. Dr. Cheng’s research primarily focus on algorithmic issues in image scene understanding, including image segmentation, salient object detection, image editing, objectness proposal, etc. He has published several highly cited papers in ACM TOG, IEEE TPAMI etc. See also: http://mmcheng.net

Abstract

One key open problem in robotics is the question of how a robot can reason about the space of possible trajectories. In particular, how can a robot determine not just a single shortest path between two points, but develop an understanding of continuous deformation classes (homotopy classes) of trajectories in configuration space? Over the last 5 years, computational geometry techniques for computing topological information from data have advanced dramatically. We will discuss our recent work on using persistent homology to determine a collection of homotopy inequivalent trajectories in robot configuration spaces and will present recent work on topologically clustering large databases of trajectories into consistent clusters with applications to the learning of robot control policies and motion primitives.

Biography

Florian Pokorny received a BSc (Honours) Mathematics from the University of Edinburgh, UK in 2005. He then obtained a Master of Advanced Studies in Mathematics (Part III of the Mathematical Tripos) from the University of Cambridge, UK before completing his PhD in pure mathematics under supervision of Prof. Michael Singer at the University of Edinburgh in 2011 in the field of differential geometry. Following his PhD, he has refocused his research on Robotics and Machine learning problems, and particularly topological methods, robotic manipulation and motion planning and joined the Center for Autonomous Systems at KTH Royal Institute of Technology, Stockholm, Sweden working with Prof. Danica Kragic. Since May 2015, he has joined the AMPLab & Automation Lab at UC Berkeley conducting postdoctoral research with Prof. Ken Goldberg and his group.

Abstract

Robotics is currently undergoing a dramatic transformation. High-performance networking and cloud computing has radically transformed how individuals and businesses manage data, and is poised to disrupt the state-of-the-art in the development of intelligent machines. This talk explores the long-term prospects for the future evolution of robot intelligence based on search, distributed computing, and big data. Ongoing research on autonomous cars and humanoid robots will be discussed in the context of how cloud-connectivity will enable future robotic systems to be more capable and useful.

Biography

James Kuffner is a Roboticist at the Toyota Research Institute and an Adjunct Associate Professor at the Robotics Institute, Carnegie Mellon University. He received a Ph.D. from the Stanford University Dept. of Computer Science Robotics Laboratory in 1999, and was a Japan Society for the Promotion of Science (JSPS) Postdoctoral Research Fellow at the University of Tokyo working on software and planning algorithms for humanoid robots. He joined the faculty at Carnegie Mellon University's Robotics Institute in 2002. He has published over 125 technical papers, holds more than 40 patents, and received the Okawa Foundation Award for Young Researchers in 2007. In 2009, James joined Google as part of the initial engineering team building Google’s self-driving car. He is known for introducing the term "Cloud Robotics" in 2010 to describe how network-connected robots could take advantage of distributed computation and data stored in the cloud. In 2014, he was appointed head of Google’s Robotics division, which he co-founded along with Andy Rubin. In 2016, he joined the newly created Toyota Research Institute as the Area Lead for Cloud Computing.

Abstract

Maker Media is a global platform for connecting Makers with each other, with products and services, and with our partners. Through media, events and ecommerce, Maker Media serves a growing community of Makers who bring a DIY mindset to technology. Whether as hobbyists or professionals, Makers are creative, resourceful and curious, developing projects that demonstrate how they can interact with the world around them. The launch of Make: magazine in 2005, followed by Maker Faire in 2006, jumpstarted a worldwide Maker Movement, which is transforming innovation, culture and education. Located in San Francisco, CA, Maker Media is the publisher of Make: magazine and the producer of Maker Faire. It also develops “getting started” kits and books that are sold in its Maker Shed store as well as in retail channels.

Biography

Dale Dougherty is the founder and Executive Chairmen of Maker Media Inc. In 2005, Maker Media launched Make Magazine and Maker Faire, which held its first events in San Francisco in 2006. He has developed a maker ecosystem, serving the needs of makers as they seek out product support, startup advice, and funding avenues. His idea for Make Magazine came from his experiences with the Hacks Books and then recognized that hackers were playing with hardware and more broadly, they were looking at how to hack the world, not just computers.

Abstract

Big Data analytics has been approached exclusively from a data-parallel perspective, where data are partitioned to multiple workers – threads or separate servers – and model training is executed concurrently over different partitions, under various synchronization schemes that guarantee speedup and/or convergence. The dual – Big Model – problem that, surprisingly, has received no attention in database analytics, is how to manage models with millions if not billions of parameters that do not fit in memory. In this talk, I will introduce the first secondary storage array-relation dot-product join operator between a set of sparse arrays and a dense relation. The dot-product join operator incurs minimal overhead when sufficient memory is available and gracefully degrades when memory resources become scarce. Overall, the dot-product join operator achieves an order of magnitude reduction in execution time for Big Model analytics over alternative in-database solutions.

Biography

Chengjie Qin is a PhD candidate in EECS department advised by Florin Rusu. His research focuses on supporting large-scale data analytics in databases. He received his Bachelor of Science degree in Computer Science in 2011 from Fuzhou University, China.

Abstract

After nearly 20 years in academia featuring several papers with "real-world" in the title, I recently left academia to found a company that commercializes one of our robots. The small, autonomous watercraft have the ability to dramatically change how water data is collected. In this talk, I will describe some of the challenges and issues I've encountered as we take cutting edge technology from this community and put it in the hands of end users. In the course of the effort, I've learned that research and commercial success are not always compatible, but that some of the same creative skills and ability to deal with failure are essential. Over time, we've found a balance between research and commercialization that helps both move forward in parallel, as well as finding different business models that work for technology on the edge of research.

Biography

Dr. Scerri is co-Founder President of Platypus, LLC. and the Director of the Perceptronics Solutions Robotics Lab. Prior to this he was an Associate Research Professor Carnegie Mellon University Robotics Institute. The focus of his research while at CMU was multi-agent and multi-robot systems. In his current roles in industry, the emphasis is on taking state-of-the-art research and applying it to real problems, with a specific focus on making the collection of important data about an environment less expensive, more reliable and more accessible.

Abstract

Researchers must navigate big data. Current scientific knowledge includes 50 million published articles. How can a system help a researcher find relevant documents in their field? We introduce SciNet, an interactive search system that anticipates the user's search intents by estimating them from the user's interaction with the interface. The estimated intents are visualized on an intent radar, a radial layout that organizes potential intents as directions in the information space. The system assists users to direct their search by allowing feedback to be targeted on keywords representing the potential intents. Users can provide feedback by moving the keywords on the intent radar. The system then learns and visualizes improved estimates and corresponding documents. The resulting user models are explicit open user models curated by the user during the interactive information seeking. SciNet has been shown to significantly improve users' task performance and the quality of retrieved information without compromising task execution time. We also show how user models learned in SciNet can be used to help cold-start recommendation in another system, the CoMeT talk management system, by cross-system user model transfer across the systems.

Biography

Jaakko Peltonen is an Associate Professor of statistics (data analysis) at the School of Information Sciences, University of Tampere, Finland where he leads the Statistical Machine Learning and Exploratory Data Analysis group; he is also currently an academy research fellow at Aalto University, Finland, where he is a PI of the Probabilistic Machine Learning research group. He is an associate editor of Neural Processing Letters and an editorial board member of Heliyon. He has served in organizing committees of seven international conferences and one international summer school, has served in program committees of 31 international conferences/workshops and has performed referee duties for numerous international journals and conferences. He is an expert in statistical machine learning methods for exploratory data analysis, visualization of data, and learning from multiple sources.

Abstract

In recent years, deep neural networks have proven to be highly effective for acoustic modeling in speech recognition. However, the input to the acoustic model consists of hand-crafted features, namely, logarithm of the energy of the Mel-weighted filter bank (log-mel). Mel-weighted filters were developed about 4 decades ago and were inspired by human perception. Apart from the possibility that they may not be optimal features for automatic speech recognition, the log-mel features strip information from the speech signal that may be useful especially for jointly modeling de-reverberation and beam-forming within the neural networks. As an alternative to log-mel features, we investigate using complex-valued frequency transform of the speech frames directly as inputs to the acoustic models and to utilize the complex-valued inputs we employ complex-valued linear layers whose parameters are learned jointly with the rest of the acoustic model. In this talk, we will discuss the properties of these complex-valued layers and demonstrate their advantage on a large speech recognition task.

Biography

Izhak Shafran is a speech and machine learning researcher, who has been working on acoustic modeling for speech recognition. Before joining Google, he was an Associate Professor and a member of the Center for Spoken Language Processing at OHSU, where he also focused on medical application of spoken language technology. He graduated from University of Washington in Seattle in 2001 and subsequently worked at AT\&T Research Labs at Florham Park with the speech algorithms group. In summer of 2006, he was a visiting professor at University of Paris-South, working at LIMSI. Subsequently, he was a research faculty at the Center for Language and Speech Processing (CLSP) in Johns Hopkins University. He received an NIH Career Development Award in 2010.

Abstract

Mechanizing the hand harvesting of fresh market crops constitutes one of the biggest challenges to the sustainability of the U.S. fruit and vegetable industry. Depending on the crop, labor contributes up to 60% of the variable production cost, and recent labor shortages have led to loss of production and reduction of planted acreage in several crops. Innovation is desperately needed in the design of mass – shake-and-catch - harvesters, and selective fruit-picking robotic harvesters. This seminar will present the challenges related to mechanized harvesting and how concepts and tools from model-based design and robotics can be used to provide solutions. Regarding robotic fruit harvesters, most developed prototypes utilize multiple-degree-of-freedom arms, often kinematically redundant. The hypothesis is that as branches constrain fruit reachability, redundancy is necessary to navigate through branches and reach fruits inside the canopy. Modern commercial orchards increasingly adopt trees of SNAP architectures (Simple, Narrow, Accessible, and Productive). In this seminar results will be presented from a recent simulation study on linear fruit reachability (LFR) on high-density, trellised pear trees, when linear only motion was used to reach the fruits. Results based on digitized geometric tree models and fruit locations showed that 91.1% of the fruits were reachable after three “harvesting passes” with proper approach angles. This implies that for trees of SNAP-type architectures fruit reachability may not require complex and expensive arms with many degrees of freedom. For shake-and-catch harvesting, results based on a physics-based simulation of falling fruits will be shown, which suggest that when fruit-intercepting rods are inserted optimally into the tree canopies during shaking, the percentage of fruits hitting branches can be lowered by more than 50%. Such designs could enable mass - harvesting with low fruit damage, and, hence, provide mechanized harvesting solutions for some crops.

Biography

Dr. Stavros Vougioukas is an Assistant Professor of Biological and Agricultural Engineering at the University of California, Davis. He joined the Department in 2012 and his research group focuses on the development of robotic and automation systems for agricultural applications, with emphasis on mechanized harvesting of specialty crops. Dr. Vougioukas earned his Diploma in Electrical Engineering (1989) at Aristotle University, Greece. He undertook graduate studies in the US under a Fulbright Fellowship. He completed his MS (1991) at SUNY Buffalo and PhD (1995) at Rensselaer Polytechnic Institute, in Electrical, Computers and Systems engineering. His PhD thesis addressed force-guided assembly and robotic fine motion planning. He was a post-doctoral researcher for one year at the University of Parma, Italy. After his army service he became faculty at Aristotle University, Greece, where he worked on agricultural automation for 10 years.