Select a calendar:
Filter April Events by Event Type:
SUNMONTUEWEDTHUFRISAT
Conferences, Lectures, & Seminars
Events for April
-
CS Colloquium: Guy van den Broeck (KU Leuven) - Scalable Inference and Learning for High-Level Probabilistic Models
Thu, Apr 02, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Guy van den Broeck, KU Leuven
Talk Title: Scalable Inference and Learning for High-Level Probabilistic Models
Series: CS Colloquium
Abstract: Probabilistic graphical models are pervasive in AI and machine learning. A recent push, however, is towards more high-level representations of uncertainty, such as probabilistic programs, probabilistic databases, and statistical relational models. This move is akin to going from hardware circuits to a full-fledged programming language, and poses key challenges for inference and learning. For instance, we encounter a fundamental limitation of classical learning algorithms: they make strong independence assumptions about the entities in the data (e.g., images, web pages, patients, etc.). These assumptions fail to hold in a global view of the data, where all entities are related. We also encounter a limitation of existing reasoning algorithms, which fail to scale to large, densely connected graphical models, consisting of millions of interrelated entities.
In this talk, I present my research on efficient algorithms for high-level probabilistic models, called lifted inference and learning algorithms. I begin by introducing the key principles behind exact lifted inference, namely to exploit symmetry and exchangeability in the model. Next, I discuss the strengths and limitations of lifting. Building on results from database theory and counting complexity, I identify classes of tractable models, and classes where high-level reasoning is fundamentally hard. I conclude by showing the practical embodiment of these ideas, in the form of approximate inference and learning algorithms that scale up to big data and big models.
The lecture will be available to stream HERE
Biography: Guy Van den Broeck graduated summa cum laude with a Ph.D. in Computer Science from KU Leuven, Belgium, in 2013. He was a postdoctoral researcher at UCLA and KU Leuven. His research interests are broadly in machine learning, artificial intelligence, knowledge representation and reasoning, and statistical relational learning. His work was awarded the ECCAI AI Dissertation Award 2014, Scientific Prize IBM Belgium for Informatics 2014, and Alcatel-Lucent Innovation Award 2009. He is the recipient of the best student paper award at ILP 2011 and a best paper honorable mention at AAAI 2014. For more information, see http://guyvandenbroeck.com
Host: Computer Science Department
Webcast: https://bluejeans.com/442226528Location: Olin Hall of Engineering (OHE) - 132
WebCast Link: https://bluejeans.com/442226528
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
CS Colloquium: Steve Checkoway (Johns Hopkins) - Revealing Reality Through Reverse Engineering
Mon, Apr 06, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Steve Checkoway, Johns Hopkins
Talk Title: Revealing Reality Through Reverse Engineering
Series: CS Colloquium
Abstract: Insecure computer systems in the wild can enable consequences ranging from crime to mass surveillance to (in the case of cyberphysical systems) physical destruction or even death. But how can anyone know if a particular computer system is insecure? One can rely on the representations of the system designers or manufacturers; however, the history of computers is replete with examples of claims that products are secure which are subsequently proven false. This is, in part, because computer systems tend to exhibit unanticipated, unintended, or poorly-understood behaviors that have complex interactions. As a result, the best way to learn about the security of a system is to take a detailed look at the hardware and software that comprise the system, and their interactions. In the common case where hardware designs and software source code are not available, reverse engineering the system is often the best way to derive ground-truth data on how the system functions.
In this talk, I'll describe some of my recent research where reverse engineering played a key role, covering TLS implementations with backdoors as well as cyberphysical systems. I'll also describe the scientific nature of reverse engineering as well as the positive, real-world impact reverse engineering can have on security and safety.
The lecture will be available to stream HERE.
Biography: Stephen Checkoway is an Assistant Research Professor in the Department of Computer Science at Johns Hopkins University and a member of the Johns Hopkins University Information Security Institute. Checkoway's research focuses on the security of embedded and cyberphysical systems. He has demonstrated exploitable vulnerabilities in such embedded systems as electronic voting machines, laptop webcams, automobiles, and airport scanners. He received his Ph.D. in Computer Science from the University of California, San Diego in 2012.
Host: CS Department
Webcast: https://bluejeans.com/774936978Location: Olin Hall of Engineering (OHE) - 132
WebCast Link: https://bluejeans.com/774936978
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
CS Colloquium: Matt Fredrikson (University of Wisconsin-Madison) - Inference Attacks: Understanding Privacy in the Era of
Tue, Apr 07, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Matt Fredrikson, University of Wisconsin-Madison
Talk Title: Inference Attacks: Understanding Privacy in the Era of
Series: CS Colloquium
Abstract: As data from far-reaching sources is collected, aggregated, and re-packaged to enable new and smarter applications, confidentiality and data security are at greater risk than ever before. Some of the most surprising and invasive threats to materialize in recent years are brought about by so-called inference attacks: successful attempts to learn sensitive information by leveraging public data such as social network updates, published research articles, and web APIs.
In this talk, I will focus on two of my research efforts to better understand and defend against these attacks. First I will discuss work that examines the privacy risks that arise when machine learning models are used in a popular medical application, and illustrate the consequences of applying differential privacy as a defense. This work uncovered a new type of inference attack on machine learning models, and shows via an in-depth case study how to understand privacy "in situ" by balancing the attacker's chance of success against the likelihood of harmful medical outcomes. The second part of the talk will detail work that helps developers correctly write privacy-aware applications using verification tools. I will illustrate how a wide range of confidentiality guarantees can be framed in terms of a new logical primitive called Satisfiability Modulo Counting, and describe a tool that I have developed around this primitive that automatically finds privacy bugs in software (or proves that the software is bug-free). Through a better understanding of how proposed defenses impact real applications, and by providing tools that help developers implement the correct defense for their task, we can begin to proactively identify potential threats to privacy and take steps to ensure that they will not surface in practice.
The lecture will be available to stream HERE.
Biography: Matt Fredrikson is a Ph.D. candidate in the department of Computer Sciences at the University of Wisconsin-Madison. His research interests lie at the intersection of security, privacy, and formal methods, covering topics in software security, privacy issues associated with machine learning models, and applied cryptography. His work has been profiled by Reuters, Technology Review, and New Scientist, and received the best paper award at USENIX Security 2014. He is a recipient of the Microsoft Research Graduate Fellowship Award.
Website:
http://pages.cs.wisc.edu/~mfredrik
Host: Computer Science Department
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
Electrical Engineering Seminar
Tue, Apr 07, 2015 @ 10:30 AM - 11:30 AM
Thomas Lord Department of Computer Science, Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Dr. Yanzhi Wang, University of Southern California
Talk Title: Energy Efficiency Enhancement Techniques in Energy Generation, Storage, and Consumption Systems
Abstract: Abstract: Power and energy consumption pose serious economic, societal, and environmental concerns in various scales of information processing applications and cyber-physical systems, ranging from battery-powered embedded systems, handheld smartphones, desktop computers and household appliances, to data centers. A joint optimization framework is necessary for energy efficiency enhancement in all the sides of energy generation, energy storage, and energy consumption.
In this presentation, I will first present my work on energy efficiency enhancement in photovoltaic (PV) energy generation and hybrid electrical energy storage systems, including (i) modeling, optimal control, and reconfiguration for combating partial shading and PV cell faults in a PV system, (ii) optimal design and control of hybrid electrical energy storage systems which can exploit the benefits of its constituent energy storage elements, and (iii) joint optimization and control.
Next I will present our work on near-threshold computing for emerging devices, which will be highly useful for future embedded and heterogeneous computing. We have proposed a device-circuit-architecture cross-layer optimization framework. At the device level we design and optimize deeply-scaled FinFET devices using accurate device simulators. At the circuit level we develop robust logic cells and SRAM cells based on these devices. At the architecture level we optimize datapath structure and cache memories to achieve low power and high robustness.
Biography: Biography: Yanzhi Wang graduated with Ph.D. degree from Ming Hsieh Department of Electrical Engineering at University of Southern California in Aug. 2014, under the supervision of Prof. Massoud Pedram. He graduated with B.S. degree with distinction from Tsinghua University in July 2009. Now he is a postdoctoral research associate and part-time lecturer at USC. His research interests include control and optimization of energy generation and energy storage systems, green and sustainable computing, and extremely low-power near-threshold computing and emerging technologies. He has published ~130 papers in major conferences and journals, including three best paper or top paper awards on top-tier conferences (ISLPED 2014, ISVLSI 2014, IEEE Cloud 2014), multiple best paper nominations and two IEEE Trans. on CAD popular papers.
Host: Prof. Viktor K. Prasanna
More Info: Hughes Aircraft Electrical Engineering Center (EEB) - 248
Location: Hughes Aircraft Electrical Engineering Center (EEB) - EEB 248
Audiences: Everyone Is Invited
Contact: Kathy Kassar
Event Link: Hughes Aircraft Electrical Engineering Center (EEB) - 248
-
CS Colloquium: Danai Koutra (Carnegie Mellon) - What’s in my data? Fast, principled algorithms for exploring large graphs
Thu, Apr 09, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Danai Koutra, Carnegie Mellon
Talk Title: Whatâs in my data? Fast, principled algorithms for exploring large graphs
Series: CS Colloquium
Abstract: Networks naturally capture a host of real-world interactions, spanning from friendships to brain activity. But, given a massive graph, such as the Facebook social network, what can be learned about its structure? Are there any changes over time? Where should people's attention be directed? In this talk I will present my work on scalable algorithms that help us to explore and make sense of large, networked data when we want to know âwhatâs in the dataâ. I will present how summarization and similarity analysis can help answer this question, and I will focus on two of my approaches âVoGâ and âDeltaConâ. VoG disentangles the complex graph connectivity patterns, and efficiently summarizes large graphs with important and semantically meaningful structures by leveraging information theoretic methods. DeltaCon is a well-founded, fast method that detects and explains changes in time-evolving or aligned networks by assessing their similarity. Both works are being used by industry, and give interesting discoveries in large real-world graphs.
The lecture will be available to stream HERE.
Biography: Danai Koutra is a Ph.D. candidate in the Computer Science Department at Carnegie Mellon. She earned her M.S. from CMU in 2013 and her diploma in ECE at the National Technical University of Athens in 2010. She works on large-scale graph mining and devises algorithms and methods for exploring, understanding, and learning from graph data when the nature of the problem is not known in advance. She holds one "rate-1" patent, and has six (pending) patents on bipartite graph alignment. She also has many papers (including 2 award-winning papers) and tutorials in top data mining conferences. Her work has been covered by media outlets, such as the MIT Technology Review, and is being taught in courses at top universities, including the Tepper School of Business at CMU and Rutgers University.
Host: Computer Science Department
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
Securing Cloud Databases
Thu, Apr 09, 2015 @ 02:30 PM - 03:30 PM
Thomas Lord Department of Computer Science, Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Ken Eguro, Embedded and Reconfigurable Computing Group, Microsoft Research
Talk Title: Securing Cloud Databases
Abstract: Despite the trend toward cloud services, many applications, such as very basic databases, are not yet in the cloud due to security concerns. This talk discusses how encryption requirements make cloud migration difficult and how fundamentally different cloud architectures are needed to support applications that handle sensitive data. The key challenge is to facilitate computation on encrypted data in an efficient manner. This talk provides an overview of MSR efforts working toward solving this problem, requiring a holistic approach combining crypto algorithms, secure hardware, distributed computation, and systems engineering.
Biography: Ken joined the Embedded and Reconfigurable Computing group at Microsoft Research in Redmond, Washington in 2008. He also holds an Affiliate Assistant Professor position in the Electrical Engineering Department at the University of Washington. Some of his past and present research interests include: applications of high-performance computing architectures, FPGA development and integration issues, and security concerns of hardware/security solutions using hardware. He is also an amateur enthusiast of cryptography & cryptanalysis.
Host: Viktor Prasanna
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248
Audiences: Everyone Is Invited
Contact: Kathy Kassar
-
CS Colloquium: Anca Dragan (Carnegie Mellon) - Interaction as Manipulation
Tue, Apr 14, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Anca Dragan, Carnegie Mellon
Talk Title: Interaction as Manipulation
Series: CS Colloquium
Abstract: The goal of my research is to enable robots to work with, around, and in support of people, autonomously producing behavior that reasons about both their function -and- their interaction with humans. I aim to develop a formal understanding of interaction that leads to algorithms which are informed by mathematical models of how humans interact with robots, enabling generalization across robot morphologies and interaction modalities.
In this talk, I will focus on one specific instance of this agenda: autonomously generating motion for coordination during human-robot collaborative manipulation. Most motion in robotics is solely functional: industrial robots move to package parts, vacuuming robots move to suck dust, and personal robots move to clean up a dirty table. This type of motion is ideal when the robot is performing a task in isolation. Collaboration, however, does not happen in isolation, and demands that we move beyond solely functional motion. In collaboration, the robot's motion has an observer, watching and interpreting the motion - inferring the robot's intent from the motion, and anticipating the robot's motion based on its intent. My work develops a mathematical model of these inferences, and integrates this model into motion planning, so that the robot can generate motion that matches people's expectations and clearly conveys its intent. In doing so, I draw on action interpretation theory, Bayesian inference, constrained trajectory optimization, and interactive learning. The resulting motion not only leads to more efficient collaboration, but also increases the fluency of the interaction as defined through both objective and subjective measures. The underlying formalism has been applied across robot morphologies, from manipulator arms to mobile robots, and across interaction modalities, such as motion, gestures, language, and shared autonomy with assistive arms.
The lecture will be available to stream HERE.
Biography: Anca Dragan is a PhD candidate at Carnegie Mellon's Robotics Institute, and a member of the Personal Robotics Lab. She was born in Romania and received her B.Sc. in Computer Science from Jacobs University in Germany in 2009. Her research lies at the intersection of robotics, machine learning, and human-robot interaction: she works on algorithms that enable robots to seamlessly work with, around, and in support of people. Anca's research and her outreach activities with children have been recognized by the Intel Fellowship and by scholarships from Siebel, the Dan David Prize, and Google Anita Borg. Anca served as General Chair in the Quality of Life Technology Center's student council, as associate editor for ARSO'14, and as program chair for three workshops on collaborative manipulation at RSS, ICML, and HRI.
Host: CS Department
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
CS Colloquium: Varun Kanade (Fondation Sciences Mathématiques de Paris) - Distributed Online Learning
Thu, Apr 16, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Varun Kanade, Fondation Sciences Mathématiques de Paris
Talk Title: Distributed Online Learning
Series: CS Colloquium
Abstract: The massive amount of data involved in modern information processing systems necessitates the use of new paradigms to effectively handle data. One such paradigm is the design of machine learning algorithms that interact with data in an online fashion, i.e., data used to make predictions is received little at a time. In addition, such algorithms may be implemented on distributed systems, resulting in a tradeoff between communication cost and prediction accuracy. In this talk, I will present a classical question in the context of online learning, the so-called experts problem. The goal in this problem is to design a prediction strategy over a set of actions in the face of uncertainty that is guaranteed to perform almost as well as the best single action in hindsight. This is a fundamental question with several applications such as drug testing, network routing, and online advertising. I will discuss the new challenges that arise when implementing algorithms for this problem in a distributed setting and present a novel algorithm that achieves a non-trivial tradeoff between prediction accuracy and communication.
The lecture will be available to stream HERE.
Biography: I am now a postdoctoral fellow at ENS through the Fondation Sciences Mathématiques de Paris (FSMP). Before this I was at UC Berkeley as a Simons Fellow. I completed my Ph.D. at Harvard University, and was extremely fortunate to have had Leslie Valiant as my adviser. Before joining Harvard, I was a graduate student at Georgia Tech, where I was working with Adam Kalai (now at Microsoft Research). I obtained a B.Tech at the Indian Institute of Technology, Bombay in Mumbai, India.
Host: Computer Science Department
More Info: https://bluejeans.com/112351952
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
Event Link: https://bluejeans.com/112351952
-
CS Colloquium: Tintri Tech Talk - Architecture in the hidden world of enterprise infrastructure
Thu, Apr 16, 2015 @ 04:00 PM - 05:30 PM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Ken Klein, Tintri
Talk Title: Architecture in the hidden world of enterprise infrastructure
Series: CS Colloquium
Abstract: The CEO - Ken Klein, USC alumnus and trustee - and a key engineer from Tintri, a successful storage virtualization startup will be visiting us on Thursday at 4pm in EEB 248 to meet students and tell us about their company and the interesting technical challenges they face. Please join us!
Intro Talk Abstract
A new model for IT is here - where virtualized applications are the norm, public and private clouds are the new modes of data access, and traditional storage is 20 years overdue for a shakeup. Enterprises are still bound by the same tedious, time-consuming, blind, antiquated storage solutions that don't address the most important dynamics of IT - the dynamic nature of data in a virtualized world.
You probably know a lot about the software and hardware in your iPhone, and something about the software and hardware that runs Google search, but do you know what kind of infrastructure is used at Time Warner or Chevron?
Most of the world's infrastructure actually looks more like Time Warner and Chevron than an iPhone or Google search, but because of fierce competition and huge budgets, most enterprise infrastructure vendors keep details of this world hidden, perhaps only revealing them to bigger customers while under NDA. However, today's enterprise infrastructure is very different from either consumer or Google-like infrastructure and is changing more rapidly than ever before.
In this talk I'll give you a glimpse into the dynamic and less conspicuous world of the enterprise by talking about the architecture of Tintri's application centered storage system. Because Tintri was architected around the modern datacenter, it is impossible to discuss our architecture without also discussing three disruptions that are reforming the enterprise:
Flash is reshaping the way we think about and use storage.
Virtualization and containerization have changed our compute and IO paths and our management frameworks.
Private cloud is renegotiating the contract between infrastructure and applications.
Come get a rare peek into modern enterprise infrastructure architecture!
Biography: Ken Klein is Chairman of the Board and Chief Executive Officer of Tintri, the leader in smart, application-aware, storage for the enterprise and the cloud. A software industry veteran with over 25 years of experience, he was previously president of Wind River, an Intel subsidiary subsequent to the sale to Intel Corporation for $1B. Prior to that he was Chairman, CEO, and president of Wind River where he was responsible for the management of 1,900 employees and nearly $400M in revenues. Before that, Ken served as Chief Operating Officer and a Board member of Mercury Interactive for 12 years. Klein and his team built Mercury from a pre-revenue startup into a software powerhouse with a peak market capitalization of $15B, 2,150 employees, operations in 35 countries, and membership in the NASDAQ 100 and S&P 500. The team went on to grow the company to nearly $1B in annual revenue and sell to Hewlett-Packard for $5B. Before his tenure at Mercury, Klein held various engineering, marketing, and management roles at Interactive Development Environments, Daisy Systems, and Hughes Aircraft Company.
Mr. Klein earned a bachelor of science degree in electrical engineering and biomedical engineering from the University of Southern California. He is a USC Distinguished Alumnus, member of the USC School of Engineering Board of Councilors, founder of USC's Klein Institute for Undergraduate Engineering Life (KIUEL), and a USC Trustee.
Brandon Salmon Bio:
Brandon Salmon has been working in systems and storage for over twelve years. He has a Ph.D in computer engineering from Carnegie Mellon University and a bachelors in computer science from Stanford. He has worked at Microsoft, VMware and Intel in the past. As the 6th engineer at Tintri he designed and built significant portions of both the core Tintri filesystem and integrations with private cloud environments. He now works in the Office of the CTO investigating market changes and new technologies.
Tintri Mission:
Tintri has created a different world. A world of smart storage that sees and learns and adapts, removing the opacity of traditional storage products, anticipating changes and needs before they arise. A world where simplicity and transparency replace complexity and brute force, all while delivering the return on investment required by business leaders.
Tintri builds smart storage that sees, learns, and adapts, enabling IT organizations to focus on virtualized applications and business services instead of managing storage infrastructure. Named a "Visionary" in Gartner's 2014 Magic Quadrant for General-Purpose Disk Array, Tintri is the only storage platform in the industry that provides virtual machine-level storage management, data protection, analytics, quality of services, and automation, with all flash performance. Addressing a TAM which IDC estimated to reach $17 billion in 2017, Tintri is fundamentally changing how companies deploy virtualized workloads in their data centers and in the cloud.
Tintri eliminates the need to overprovision storage for performance to meet SLA or QoS levels - a common practice with conventional storage which dramatically increases CAPEX, OPEX, and storage footprint. Every step in the Tintri experience is designed to be profoundly simple - what used to take days and hours to accomplish now takes minutes.
Host: Wyatt Lloyd
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
CS Colloquium: Joyce Ho (UT-Austin) - Extracting medically interpretable concepts from complex health data
Mon, Apr 20, 2015 @ 11:00 AM - 12:00 PM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Joyce Ho , UT-Austin
Talk Title: Extracting medically interpretable concepts from complex health data
Series: CS Colloquium
Abstract: Electronic health records (EHRs) are an increasingly important source of patient information. However, a major challenge is how to transform EHR into meaningful concepts so domain experts can act on the information in an appropriate manner. In this talk, I will discuss two approaches to extract concise, meaningful concepts from certain types of health datasets. First, I will describe a dynamic time series model that tracks a patient's cardiac arrest risk based on physiological measurements. Our algorithm is inspired by financial econometric and yields interpretability and predictability of a cardiac arrest event. Next, I will present sparse, nonnegative tensor factorization models to obtain multiple medical concepts with minimal human supervision. Tensor factorization utilizes information in the multiway structure to derive concise latent factors even with limited observations. Experimental results on real EHRs demonstrate the effectiveness of our models to extract medically interpretable concepts from complex health data.
Biography: Joyce Ho is a PhD Candidate in the Electrical and Computer Engineering Department at the University of Texas at Austin co-advised by Dr. Joydeep Ghosh and Dr. Sriram Vishwanath. Her research involves the development of novel data mining and machine learning algorithms to address problems in healthcare. Joyce has also co-founded a healthcare data analytics company, Accordion Health, which was awarded an NSF Small Business Innovation Research (SBIR) grant.
Host: Prof. Yan Liu
Location: Henry Salvatori Computer Science Center (SAL) - 322
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
CS Colloquium: Yasin Abbasi-Yadkori (Queensland University of Technology) - Planning and Learning in Sequential Decision ProblemsPlanning and Learning in Sequential Decision Problems
Tue, Apr 21, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Yasin Abbasi-Yadkori, Queensland University of Technology
Talk Title: Planning and Learning in Sequential Decision Problems
Series: CS Colloquium
Abstract: Many decision problems have an interactive nature; the decision maker executes an action, receives feedback from the environment, and finally uses the feedback to improve the next decision. For instance, an Internet news recommendation system must make a recommendation based on the current visitor. The system then observes the click patterns of the visitor and can change its future recommendations. Such sequential decision problems are particularly challenging when the decision and state spaces are large, which is often the case in modern applications.
In this talk, I will present my research in planning and learning in large sequential decision problems. I will consider three fundamental decision problems: problems with linear dynamics and quadratic losses (LQ problem); linear optimization with limited feedback (bandit problems); and policy optimization for large scale Markov decision processes. I will demonstrate a data-efficient adaptive controller and show the first finite-time performance guarantee for the LQ problem. For bandit problems, I will present an algorithm that can exploit sparsity in data. The improvement stems from the construction of smaller confidence sets. In particular, I will show the first sparsity confidence set for the linear regression problem. Finally, I will discuss convex optimization reductions for very general Markov decision (planning) problems. The reductions allow us to design computationally efficient algorithms that enjoy strong performance guarantees.
The lecture will be available to stream HERE
Host: Fei Sha
More Info: https://bluejeans.com/866147590
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
Event Link: https://bluejeans.com/866147590
-
CS Colloquium: Mark Zhandry (Stanford) - The Surprising Power of Modern Cryptography
Thu, Apr 23, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Mark Zhandry, Stanford
Talk Title: The Surprising Power of Modern Cryptography
Series: CS Colloquium
Abstract: Modern cryptography is surprisingly powerful, yielding capabilities such as secure multiparty computation, computing on encrypted data, and hiding secrets in code. Currently, however, some of these advanced abilities are still too inefficient for practical use. The goals of my research are two-fold: (1) continue expanding the capabilities of cryptography and its applications, and (2) bring these advanced capabilities closer to practice.
In this talk, I will focus on a particular contribution that addresses both of these objectives: establishing a shared secret key among a group of participants with only a single round of interaction. The first such protocols required a setup phase, where a central authority determines the parameters for the scheme; unfortunately, this authority can learn the shared group key and must therefore be trusted. I will discuss how to remove this setup phase using program obfuscation, though the scheme is very impractical due to the inefficiencies of current obfuscators. I will then describe a new technical tool called witness pseudorandom functions and show how to use this tool in place of obfuscation, resulting in a significantly more efficient protocol.
the lecture will be available to stream HERE.
Biography: Mark Zhandry is a Ph.D. candidate at Stanford University advised by Dan Boneh. He studies cryptography and computer science theory and is currently focusing on developing new cutting-edge cryptographic capabilities and improving the efficiency of these applications. He is visiting Microsoft Research New England and MIT for the 2014-15 academic year.
Host: Computer Science Department
More Info: https://bluejeans.com/638649971
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
Event Link: https://bluejeans.com/638649971
-
CS Student Colloquium Series: Kuan Liu & Alireza Bagheri Garakani
Thu, Apr 23, 2015 @ 04:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Kuan Liu & Alireza Bagheri Garakani, USC Computer Science
Talk Title: Similarity Learning for High Dimensional Sparse Data; A Distributed Frank-Wolfe Algorithm for Communication-Efficient Sparse Learning
Series: Student Seminar Series
Abstract: Similarity Learning for High Dimensional Sparse Data
Kuan Liu
A good measure of similarity between data points is crucial to many tasks in machine learning. Similarity and metric learning methods learn such measures automatically from data, but they do not scale well respect to the dimensionality of the data. In this talk, we describe a method that can learn efficiently similarity measure from high dimensional sparse data. The core idea is to parameterize the similarity measure as a convex combination of rank-one matrices with specific sparsity structures. The parameters are then optimized with an approximate Frank-Wolfe procedure to maximally satisfy relative similarity constraints on the training data. Our algorithm greedily incorporates one pair of features at a time into the similarity measure, providing an efficient way to control the number of active features and thus reduce overfitting. It enjoys very appealing convergence guarantees and its time and memory complexity depends on the sparsity of the data instead of the dimension of the feature space. Our experiments on real world high-dimensional datasets demonstrate its potential for classification, dimensionality reduction and data exploration
A Distributed Frank-Wolfe Algorithm for Communication-Efficient Sparse Learning
Alireza Bagheri Garakani
Learning sparse combinations is a frequent theme in machine learning. In this paper, we study its associated optimization problem in the distributed setting where the elements to be combined are not centrally located but spread over a network. We address the key challenges of balancing communication costs and optimization errors. To this end, we propose a distributed Frank-Wolfe (dFW) algorithm. We obtain theoretical guarantees on the optimization error and communication cost that do not depend on the total number of combining elements. We further show that the communication cost of dFW is optimal by deriving a lower-bound on the communication cost required to construct an epsilon-approximate solution. We validate our theoretical analysis with empirical studies on synthetic and real-world data, which demonstrate that dFW outperforms both baselines and competing methods. We also study the performance of dFW when the conditions of our analysis are relaxed, and show that dFW is fairly robust.
Host: CS Department
Location: Henry Salvatori Computer Science Center (SAL) - 101
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
-
Enery informatics distinguished seminar
Fri, Apr 24, 2015 @ 10:30 AM - 11:30 AM
Thomas Lord Department of Computer Science, Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Dr. Christos Faloutsos, Carnegie Mellon University
Talk Title: Large Graph Mining: Patterns, Cascades, Fraud Detection, and Algorithms
Series: Energy Informatics Distinguished Seminar Series
Abstract: Given a large graph, like who-calls-whom, or who-likes-whom, what behavior is normal and what should be surprising, possibly due to fraudulent activity? How do graphs evolve over time? How does influence/news/viruses propagate, over time? We focus on three topics: (a) anomaly detection in large static graphs (b) patterns and anomalies in large time-evolving graphs and (c) cascades and immunization.
For the first, we present a list of static and temporal laws, including advances patterns like 'eigenspokes'; we show how to use them to spot suspicious activities, in on-line buyer-and-seller settings, in FaceBook, in twitter-like networks. For the second, we show how to handle time-evolving graphs as tensors, how to handle large tensors in map-reduce environments, as well as some discoveries such settings.
For the third, we show that for virus propagation, a single number is enough to characterize the connectivity of graph, and thus we show how to do efficient immunization for almost any type of virus (SIS - no immunity; SIR - lifetime immunity; etc)
We conclude with some open research questions for graph mining.
Biography: Christos Faloutsos is a Professor at Carnegie Mellon University. He has received the Presidential Young Investigator Award by the National Science Foundation (1989), the Research Contributions Award in ICDM 2006, the SIGKDD Innovations Award (2010), twenty "best paper" awards (including two "test of time" awards), and four teaching awards. Five of his advisees have attracted KDD or SCS dissertation awards. He is an ACM Fellow, he has served as a member of the executive committee of SIGKDD; he has published over 300 refereed articles, 17 book chapters and two monographs. He holds eight patents and he has given over 35 tutorials and over 15 invited distinguished lectures. His research interests include data mining for graphs and streams, fractals, database performance, and indexing for multimedia and bio-informatics data.
Host: Prof. Viktor Prasanna and Dr. Charalampos Chelmis
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 132
Audiences: Everyone Is Invited
Contact: Annie Yu
-
CS Colloquium: Ilias Diakonikolas (University of Edinburgh) - Algorithmic Approaches in Unsupervised Learning
Tue, Apr 28, 2015 @ 09:45 AM - 10:50 AM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Ilias Diakonikolas, University of Edinburgh
Talk Title: Algorithmic Approaches in Unsupervised Learning
Series: CS Colloquium
Abstract: The growing scale of modern data sets and our increasingly ambitious inferential goals have highlighted new algorithmic challenges. In this talk, I will discuss recent progress in this vein that lies at the interface of computer science and statistics. I will highlight how the algorithmic perspective brings novel insights and leads to computationally efficient methods for classical statistical problems.
In this talk, I will focus on a core problem in unsupervised learning: how to infer information about a distribution based on random samples. An important goal in this context is understanding the structure in the data without making strong assumptions on its form. I will describe a unified algorithmic framework that yields new, provably efficient estimators for several natural and well-studied statistical models, including mixtures of structured distribution families (e.g., gaussian, log-concave, etc.). This framework provides a fairly complete picture of the sample and computational complexities for fundamental inference tasks, including density estimation and hypothesis testing.
I will also briefly describe some of my other work on learning, including supervised learning with missing and noisy data, as well as connections between these questions and seemingly unrelated problems in game theory and complexity theory.
The event will be available to stream HERE
Biography: Ilias Diakonikolas is an Assistant Professor in the School of Informatics at the University of Edinburgh. He holds a diploma in electrical and computer engineering from the National Technical University of Athens, and a Ph.D. in computer science from Columbia University (2010) where he was advised by Mihalis Yannakakis. He received a best thesis award for his doctoral dissertation and an honorable mention in the 2009 George Nicholson competition from the INFORMS society. Before moving to Edinburgh he spent two years (2010-2012) as the Simons postdoctoral fellow in theoretical computer science at the University of California, Berkeley. Ilias has worked in several areas of algorithms, including optimization, computational learning, and computational economics. His research focus is on the algorithmic foundations of massive data sets, in particular on designing efficient algorithms for statistics and machine learning.
Host: Computer Science Department
Location: Olin Hall of Engineering (OHE) - 132
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
