2018-19

Abstract

In this talk we will discuss model compression in general and an algorithm to achieve it, and its specific case for neural network pruning. Pruning a neural net consists of removing weights with the goal of minimally degrading its performance. This is an old problem of renewed interest because of the need to compress ever larger nets so they can run on mobile devices. We formulate pruning as an optimization problem of finding the weights that minimize the loss while satisfying a pruning cost condition. We give a generic algorithm to solve this which alternates "learning" steps that optimize a regularized, data-dependent loss and "compression" steps that mark weights for pruning in a data-independent way. Using a single pruning-level user parameter, we achieve state-of-the-art pruning in LeNet-s and ResNet-s of various sizes.

Biography

Yerlan Idelbayev is a 3rd year PhD student of EECS department studying under the supervision of Prof. Miguel Carreira-Perpinan. Before UC Merced he studied Information Systems in International IT University in Almaty, Kazakhstan, and Computer Science in UC San Diego. His research interests consist of nonlinear optimization, neural networks and their compression.

Abstract

Recently, there has been a tremendous growth in the amount of big spatial data that are acquired by different sources such as satellites, IoT sensors, smartphones, autonomous cars, and others. For decades, end-users were familiar with an interactive exploratory interface that allows them to apply spatial operations and explore the results in real-time. However, the increasing volume of the data makes it unpractical to provide the desired exploratory and real-time interface. This talk presents a new system paradigm that overcomes the limitation of the existing system by providing an approximate and incremental query processing for big spatial data. The system consists of three modules, synoptic computation, incremental indexing, and interactive visualization. The synoptic computation module scales up the query processing by providing a real-time approximate answer over small-size synopses of the data such as samples and histograms. The incremental indexing module works in the background and incrementally organizes the data over a cluster of machines to speed up the query processing. Finally, the interactive visualization module presents the results in a visual format which allows the users to inspect the query answers. Preliminary results on the proposed system show that it can bridge the gap between the user requirements for interactivity and the increasing volume of big spatial data.

Biography

Ahmed Eldawy is an Assistant Professor in Computer Science at the University of California Riverside. His research interests lie in the broad area of databases with a focus on big data management and spatial data processing. Ahmed is the main inventor of SpatialHadoop, the most comprehensive open source system for big spatial data management. Ahmed has many collaborators in industrial research labs including Microsoft Research and IBM Watson. He was awarded the best poster award in SIGSPATIAL 2017, Quality Metrics Fellowship in 2016, Doctoral Dissertation Fellowship in 2015, and Best Poster Runner-up award in ICDE 2014.

Abstract

The number of everyday smart devices (such as smart TV, Samsung SmartThings, Nest, Google Home) is projected to grow to the billions in the coming decade. The Cyber-Physical Systems or Internet of Things systems that consist of these devices are used to obtain human information for various smart building applications. Different sensing approaches have been explored, including vision-, sound-, RF-, mobile-, and load-based methods, to obtain various indoor human information. From the system perspective, general problems faced by these existing technologies are their sensing requirements (e.g., line-of-sight, high deployment density, carrying a device) and intrusiveness (e.g., privacy concerns).
My research focuses on non-intrusive indoor human information acquisition through ambient structural vibration, which is referred to as "structures as sensors". People's interaction with structures in the ambient environment (e.g., floor, table, door) induces those structures to vibrate. By capturing and analyzing the vibration response of structures, we can indirectly infer information about the people and their actions that cause it. However, challenges remain. Due to the complexity of the physical world (in this case, both structures and people), sensing data distributions can change significantly under different sensing conditions. Therefore, from the data perspective, accurate information learning through a pure data-driven approach requires a large amount of labeled data, which is costly and difficult if not impossible to obtain in real-world sensing applications. My research addresses these challenges by combining physical knowledge and data-driven approaches. Specifically, my system can robustly learn human information from limited labeled data distributions by iteratively expanding the labeled dataset. With insights into the relationship between changes of sensing data distributions and measurable physical attributes, the iterative algorithm guides the expansion order by measured physical attributes to ensure a high learning accuracy in each iteration.

Biography

Dr. Shijia Pan is a postdoctoral researcher at Carnegie Mellon University. She received her Bachelor's degree in Computer Science and Technology from University of Science and Technology of China and her Ph.D. degree in Electrical and Computer Engineering at Carnegie Mellon University. Her research interests include cyber-physical systems, Internet-of-Things (IoT), and ubiquitous computing. She worked in multiple disciplines and focused on indoor human information acquisition through ambient sensing. She has published in both top-tier Computer Science ACM/IEEE conferences (IPSN, UbiComp) and high-impact Civil Engineering journals (Journal of Sound and Vibration, Frontiers Built Environment). She is the recipient of numerous awards and fellowships, including Rising Stars in EECS, Nick G. Vlahakis Graduate Fellowship, Google Anita Borg Scholarship, Best Poster Awards (SenSys, IPSN), Best Demo Award (Ubicomp), Best Presentation Award (SenSys Doctoral Colloquium), and Audience Choice Award (BuildSys) from ACM/IEEE conferences.

Abstract

Collaboration and communication gaps, ranging from difficulties in expressing oneself or understanding another person, to failure in coordinating actions in teamwork, are prevalent problems to individuals and organizations. While improving personal communication skills continues to be important, designing digital communication channels to afford what group collaboration needs, can offer solutions with scalability and cost-efficiency. In this talk, I will conceptualize social computing interaction design as a meta-solution to shape group behaviors toward more desirable processes and outcomes. I will demonstrate the approach with our recent work tackling different tasks and contexts, such as creative brainstorming, cross-lingual conversation, and generating and understanding referential expressions in remote collaborative work.

Biography

Hao-Chuan Wang is an Acting Associate Professor in the Department of Computer Science, University of California, Davis. Before joining UC Davis, he was an Associate Professor in National Tsing Hua University, Taiwan (NTHU), Taiwan from 2012 to 2018. He's also affiliated with National Taiwan University (NTU)'s IoX Research Center as a Principal Investigator. He has formed international collaborations with peer researchers in North America and Asia, as well as industrial collaborations with Intel Labs, Microsoft Research, and Google. Dr. Wang's main research interest lies in the collaborative and social aspects of Human-Computer Interaction (HCI). His work integrates system design and the behavioral sciences of social computing research for problem solving and value creation. His recent projects include system designs for supporting multilingual collaboration, motion sensing-based analytics for studying non-verbal behaviors in mediated conversations, and studies of interpersonal knowledge transfer for augmenting human work in the future. Dr. Wang is an active member of international and regional HCI communities, including ACM SIGCHI, CSCW and Chinese CHI. He also served as a member in the Steering Committees of CSCW and Chinese CHI, and was a Subcommittee Chair for ACM CHI 2017 and 2018.

Abstract

As HPC systems scale in size and power, the danger of silent errors, i.e., errors that can bypass hardware detection mechanisms and impact application state, grows dramatically. Consequently, applications running on HPC systems need to exhibit resilience to soft errors. Previous work has found that, for certain codes, this resilience can come for free, i.e., some applications are naturally resilient. However, we still lacks fundamental understanding on the program constructs that result in such natural error resilience. Understanding such nature resilience is critical for error detection and recovery to avoid overprotecting regions of code that are naturally resilient.

In this talk, we will present our research efforts to capture and characterize application natural resilience, based on which we can quantify and model application resilience. This talk has two parts. In the first part, will discuss FlipTracker, a framework designed to extract resilience code patterns using fine-grained tracking of error propagation and resilience properties. The framework and patterns enable a deeper understanding of resilience properties of applications. We also show how we can guide application design towards natural resilience using resilience code patterns.

In the second part, we will discuss PARIS, a resilience prediction method that makes resilience predictions of fault manifestations using resilience code patterns and machine learning models. PARIS can predict the possibility of all fault manifestations, while the state-of-the-art resilience prediction model cannot. PARIS is also much faster (up to 450x speedup) than the traditional method (i.e., random fault injection).

Biography

Luanzheng Guo is a Ph.D. student of Computer Science at the University of California Merced. His study is under the supervision of Professor Dong Li. His research area is High Performance Computing System with a focus on fault tolerance in large-scale parallel systems. During his Ph.D. study, his poster was nominated as the best poster candidate in SC'16. He is a lead student volunteer in SC18. He is a reviewer of a couple of prestigious conferences and international journals. He was recognized as an outstanding reviewer by Elsevier in 2018. He was a summer intern at Lawrence Livermore National Laboratory in 2015-2018. Recently, his research is featured by HPCwire in its What's New in HPC Research. He is a student member of IEEE, ACM, and SIGHPC.

Abstract

Walnut is under constant attack by soil-borne plant pathogens including crown gall, root rots, and plant-parasitic nematodes. Because of the lifetime expectancy of walnut orchards of at least three to four decades, a high level of sustainable management and mitigation strategies for these soil-dwelling nematodes are paramount. Using rootstocks with elevated resistance and tolerance to all of these damaging organisms is an environmentally friendly and sustainable approach to reduce reliance on costly and possibly environment impacting management practices. Built on previous successes, a group of researchers from several UC campuses, USDA-ARS, the California State University of Fresno, and UCANR has formed to investigate the potential of walnut germplasm (Juglans spp.) to generate such rootstocks. Interspecific crosses within Juglans have been made, taken into tissue culture by embryo rescue, and regenerated into clonal plants. Recent efforts have focused on two breeding populations with ca. 300 genotypes of clonal offspring. These are characterized for responses against different soil-borne pathogens including Crown gall, Phytophthora root rots, and plant-parasitic nematodes. In parallel, each genotype is sequenced to create a genotypic map. As soon as phenotypic maps become available, these will be overlaid on the genotypic map to identify quantitative trait loci (QTL). Depending on the time necessary for the pathogen testing, progress varies among pathogen systems. Goal of these efforts are to improve breeding strategies, release new superior rootstocks, and to convey information on plant utility and economics to the walnut stakeholders.

Biography

Andreas Westphal is a native to Germany. He completed his College and early University training in Germany before pursuing his Ph.D. in the US. For two decades, he has been working in several nematode-host plant systems. His research endeavors encompass nematode management in several annual crops including maize, potato, small grains, soybean, sugar beet, watermelon and others. After a scientist role at the German resort research institute "Julius Kühn-Institut", he focused his research emphasis on host plant resistance and tolerance research in perennial crops. Since his employment with UC Riverside in 2015, he directs selection efforts for nematode resistance and tolerance in Walnut, Prunus, Pistachio, and grape. He also conducts management research in these crops.

Abstract

Recent advances in new memory technologies and packaging options has focused attention on computer memory system design and evaluation. Examples include high bandwidth memories such as Hybrid Memory Cube and HBM, 3DXpoint non-volatile memory, STT-MRAM, and ReRAM. Emerging memories display a wide range of bandwidths, latencies, and capacities, making it challenging for the computer architect to navigate the design space of potential memory configurations, and for the application developer to assess performance impact of complex memory systems.

The Logic in Memory Emulator (LiME) is an FPGA-based hardware/software tool specially designed for memory system evaluation and experiment. LiME uses the Xilinx Zynq UltraScale+ Multi Processor System on Chip (MPSoC) to capture any/all memory access, either from the CPU (Processing System or PS) or the FPGA (Programmable Logic or PL). LiME employs novel loopback circuitry in conjunction with address map relocation to pass memory references from the PS into the PL side. The memory request is looped back into the PS DRAM memory controller and concurrently processed by LiME.

We have demonstrated three high value use cases: non-intrusive memory access logging, emulation of multiple memory systems by passing the memory request through delay registers before entering the PS memory subsystem, and emulation of acceleration engines that can independently access memory. In this talk, I will describe this novel application of state-of-the-art FPGA embodied by the LiME framework and highlight its uses.

Biography

Maya Gokhale is Distinguished Member of Technical Staff at the Lawrence Livermore National Laboratory, USA. Her career spans research conducted in academia, industry, and National Laboratories. Maya received a Ph.D. in Computer Science from University of Pennsylvania. Her current research interests include data intensive architectures and reconfigurable computing. Maya's Streams-C programming language and compiler was adoptd by Impulse Accelerated Technologies as the basis for Impulse C. Maya is co-recipient of an R&D 100 award for the Trident C-to-FPGA compiler, co-recipient of four patents related to memory architectures for embedded processors, reconfigurable computing architectures, and cybersecurity, and co-author of more than one hundred technical publications. Maya is a member of Phi Beta Kappa and a Fellow of the IEEE for contributions to reconfigurable computing technology.

Abstract

Deep learning algorithms often require solving a highly non-linear and nonconvex unconstrained optimization problem. Generally, methods for solving the optimization problems in deep learning are restricted to the class of first-order algorithms, like stochastic gradient descent (SGD). SGD methods has several drawbacks such as undesirable effect of not escaping saddle-points and requirement for tuning so many hyper parameters. Using the second-order curvature information to find the search direction can help with more robust convergence for the non-convex optimization problem. However, computing the Hessian matrix for the large-scale problems in not computationally practical. Alternatively, quasi-Newton methods construct an approximate of Hessian matrix to build a quadratic model of the objective function. Quasi-Newton methods, like SGD, require only first-order gradient information, but they can result in superlinear convergence, which makes them attractive alternatives for solving the non-convex optimization problem in deep learning. In this talk, I will introduce limited-memory quasi-Newton optimization methods that are efficient for deep learning problems such as classification and regression of big data.

Biography

Jacob Rafati is a Ph.D. candidate in the Electrical Engineering and Computer Science program at the University of California, Merced. He is also a member of Dr. David C. Noelle’s Computational Cognitive Neuroscience Lab. His research focus is on Optimization, Machine Learning and Reinforcement Learning. This talk is based on his recent collaborative research work that involves investigating and implementing alternative optimization methods for large-scale machine learning problems, such as deep learning and deep reinforcement learning. For more details about this project visit his website at http://rafati.net.

Abstract

Artificial Intelligence (AI) technologies are often included as product features (e.g., facial and voice recognition, autonomous driving) that drive product and service innovation. However, such innovations increase software complexity, leading to security and privacy issues. Customer reactions to security or privacy failures may affect product demand; customer demand reaction to the security or privacy implications of new features, such as AI-driven technology, plays a role in regulating the rate of innovation. This work examines the trade-off between product innovation and the increased risk of security breaches in AI-enabled products and services.

Biography

Dr. Lisa Yeo is an Assistant Professor in the Ernest & Julio Management program at UC Merced who works to help organizations understand how to safely govern the data and information they need to compete. By focusing on people and process, Lisa believes that organizations can design and build information systems that make it easy to protect privacy and prevent security breaches without requiring extensive investments in security layers after the fact. Lisa has worked in information security for 15 years as both a technical specialist and a business advisor. During this time, she wrote the book Personal Firewalls, protected the infrastructure for the Alberta Legislature, and guided the secure connection of all public libraries in Alberta as part of the Alberta SuperNet project. Lisa holds a B. Math in Applied Math from the University of Waterloo and an MBA and PhD (Operations & Information Systems) from the University of Alberta.

Abstract

Extracting information of interest from remote sensing images, such as taken from satellites or aircraft, is a long-standing problem, yet one which has benefited significantly over the past few years from advances in deep learning. In this talk, we cover some of the advances in remote sensing image analysis based on deep learning and present our recent research in the field. In particular, we describe our work on cloud and cloud shadow detection in Landsat multi-spectral imagery, and our work on extracting geo-objects of interest from high resolution satellite and aerial imagery.

Biography

Dr. Dengfeng Chai is an associate professor at the Institute of Spatial Information Technique, Zhejiang University, China. He received his PhD from the State Key Lab of CAD&CG at Zhejiang University. Before that, he received his Master’s degree from the State Key Lab of Information Engineering in Surveying, Mapping and Remote Sensing at Wuhan University and his Bachelor’s degree from Wuhan University. He was a postdoctoral fellow at Department of Photogrammetry, University of Bonn, Germany. He is currently a visiting scholar at the University of California, Merced where he is being hosted by Prof. Shawn Newsam. His research interests include photogrammetry and remote sensing, computer vision and pattern recognition, mainly focusing on statistical approaches and spatial models for object recognition and extraction from remote sensing images, especially deep learning based approaches. He is a principle investigator of two projects supported by National Natural Science Foundation of China and one project supported by Zhejiang Provincial Natural Science Foundation of China. He has published at ICCV, CVPR and Pattern Recognition among other venues.

Abstract

Social communication is one key component to success and happiness. Our ability to express our needs and wants as well as understand others is central to our connection to one another and our availability to teach and learn. Challenges with social communication puts learning on hold and youth at risk for bullying, social isolation, and potentially serious mental-health concerns. Thus, supporting social skills of people with autism could have a positive impact on both the social and mental wellbeing of individuals with autism. Although much researched has focused on supporting social skills broadly, little attention has been paid to developing effective nonverbal behaviors-which are necessary to initiate, maintain, and gracefully terminate a social interaction. The talk describes the design and evaluate the effect of realtime visualizations of prosody and proximity through three lab-based experiments as well as interviewing the participants and family members about their experience with these novel AR and VR technologies. The results from the interviews with participants and parents about their experiences highlight issues of usability, learnability, and comfortability of the systems culminate in an assistive technology design concept-Sensory Accommodation Framework-which provides four technical mechanisms for supporting sensory perception differences through computation.

Biography

Dr. LouAnne Boyd is an Assistant Professor of Software Engineering and Computer Science Department in Chapman's Schmid College of Science and Technology. Her current research interests in Human-Computer Interaction include designing, developing, and evaluating novel assistive and accessible technologies for neurodiverse users. LouAnne holds a B.A. in psychology from Washington University in St. Louis, a M.A. in psychology from Towson University, and a Ph.D. in Informatics from UC Irvine. She also is a Board Certified Behavior Analyst with over 20 years of professional clinical experience working with neurodiverse people in hospital, school, home, and community settings. Her overarching goal is to promoting diversity and inclusion. To that end, her current HCI research explores technical mechanisms to support sensory accommodation for assistive technology users.

Spring 2019