CS Colloquium (BMAC) Speing 2015: Titles/Abstracts/Bios

Certifying Algorithms and the Undergraduate C.S. Curriculum

Ross Mc Connell
Department of Computer Science, Colorado State University

The primary economic cost of software bugs is not that a program fails, but that it convinces people and systems to act based on phoney outputs, compromising decision-making and safety.

Algorithms researchers must prove that an algorithm is correct. However, this does not mean that a software implementation of one has no bugs. A traditional approach to mitigating the impact of bugs is careful software design and extensive testing. However, this never produces certainty that outputs have not been compromised.

A viable alternative in many cases is to design an algorithm so that it produces, along with each output, an easy-to-check certificates that the output is a correct one for the given instance of the problem. If a certificate fails to confirm the output, a bug has undermined either the output or the generation of the certificate, and the user rejects the result. Otherwise, the user can proceed with the certainty that the output has not been compromised, sidestepping the difficult question of whether the software contains latent bugs.

The talk explore how of many of the algorithms taught in the standard undergraduate curriculum can be modified to be certifying.

Speaker Bio: Ross McConnell has been a professor at CSU since 2002. He grew up in Denver and got his Ph.D. at C.U. Boulder.

The Role of Automation in Undergraduate Computer Science

Chris Wilcox
Department of Computer Science, Colorado State University

Increasing enrollments in undergraduate computer science courses are overloading faculty, teaching assistants, labs, and other resources. To cope with resource problems, many universities have automated key processes such as program grading. Before replacing a manual process with an automated tool, we believe that two questions should be answered. First and most importantly, does the proposed automation contribute to or detract from student learning? Second, do the benefits outweigh the costs of the automation? In this talk we present results from an introductory programming course in which we have incorporated high levels of automation. We find that carefully designed and managed automation can improve student performance while realizing a significant savings of scarce teaching resources.

Speaker Bio: Chris Wilcox worked in industry for companies including Evans and Sutherland, Hewlett-Packard, and nVidia Corporation before returning to school at CSU in 2006. After finishing his doctorate in 2012 he transitioned to a teaching position at CSU, where he teaches primarily freshman and sophomore level courses. His current research interests are in Computer Science education.

Regularization of the Inverse Scattering Problem using Shearlet Frames

Gitta Kutyniok
Einstein Chair, Institute for Mathematics
Technical University, Berlin

The scattering problem analyzes how incident waves, radiation, or particles, which are transmitted in a medium, are scattered at inhomogeneities of this medium. The associated inverse problem aims to determine characteristics of the inhomogeneities from the asymptotic behavior of such scattered waves. This problem appears in various flavors in different application areas, e.g. non-destructive testing, ultrasound tomography, and echolocation.

In this talk, our focus will be on regularization techniques for the numerical solution of inverse scattering problems in two space dimensions. Assuming that the boundary of a scatterer is its most prominent feature, we exploit as model the class of cartoon-like functions. Since functions in this class are asymptotically optimally sparsely approximated by shearlet frames, we consider shearlets as a means for regularization. We then analyze two approaches, namely the nonlinear problem and a linearized problem obtained by the Born approximation technique. For the first class, we study the acoustic inverse scattering problem, whereas for the second class, we consider the inverse scattering problem of the Schrodinger equation. Besides a theoretical analysis of our numerical approaches, we will also provide numerical examples that highlight the effectiveness of our schemes.

This is joint work with Volker Mehrmann and Philipp Petersen (TU Berlin).

Speaker Bio: Gitta Kutyniok completed her Diploma in Mathematics and Computer Science in 1996 at the Universitat Paderborn in Germany. She received her Ph.D. degree in the area of time-frequency analysis from the same university in 2000. She completed her Habilitation in Mathematics in 2006 and received her venia legendi. In 2007, she was awarded a Heisenberg Fellowship by the DFG-German Research Foundation. From 2001 to 2008 she held visiting appointments at several US institutions, including Princeton University, Stanford University, Yale University, Georgia Institute of Technology, and Washington University in St. Louis. After returning to Germany in October 2008, she became a full professor of mathematics at the Universitat Osnabrueck, and headed the Applied Analysis Group. Since October 2011, she has an Einstein Chair at the Technical University of Berlin and is head of the Applied Functional Analysis Group (AFG).

Her research and teaching have been recognized by various awards, including the von Kaven Prize by the German Research Foundation, awards by the University Paderborn and the Justus-Liebig University Giessen for Excellence in Research, as well as the Weierstrass Prize for Outstanding Teaching. She will also deliver the Noether-Lecture at this year's annual meeting of the Mathematical Societies from Germany and Austria. She is an Associate Editor and also Corresponding Editor for several journals in the area of applied mathematics. She is also a board member of the Berlin Mathematical School, a member of the council of the MATHEON "Mathematics for key technologies" in Berlin, and the chair of the GAMM activity group on "Mathematical Signal- and Image Processing".

Is Anything Safe Today? Vulnerabilities: Discovery and Exploitation

Yashwant Malaiya
Computer Science Department, Colorado State University

Software that runs our banking, commerce, infrastructure, even cars, baby monitors and pacemakers can be hacked by exploiting vulnerabilities. How much risk do they represent?

We consider some key questions - What are the components of the security risk and how they can possibly integrated. What are the challenges that need to be addressed before a quantitative security risk measure can be evaluated and validated?

We start by formally define risk, and then examine the partial risk measures. CVSS scores, Microsoft Exploitability Index and Attack Surface measurements measure some of the components of the overall risk. We need to examine the complete vulnerability life-cycle from discovery to potential exploitation.

While quantitative risk evaluation is common in some fields such as finance, attempts to quantitatively assess security are relatively new. There has been criticism of the quantitative attempts of risk evaluation due to the lack of data for validating the methods. Related data has now begun to become available. Some of the recent developments will be discussed and questions that need further research will be identified.

Speaker Bio:Yashwant K. Malaiya is a professor in the Computer Science Department at Colorado State University. He has published about 200 papers in the areas of fault modeling, software and hardware reliability and quantitative security risk evaluation. He has served as the General Chair of 1993 and 2003 IEEE International Symposium on Software Reliability Engineering (ISSRE). He has co-edited the IEEECS Technology Series book Software Reliability Models, Theoretical Developments, Evaluation and Applications. He is a recipient of the IEEE Third Millennium Medal, and the IEEE Computer Society Golden Core award.

Products of Random Stochastic Matrices

Behrouz Touri
Electrical, Computer and Energy Engineering Department, University of Colorado

Products of random stochastic matrices have applications in many engineering settings including coordination in multiagent networks, distributed optimization, and the study of random phenomena in time-varying environments. In this talk we will discuss some recent results on products of random stochastic matrices that relate their convergence properties to the study of certain types of flows over networks. Also, a family of Lyapunov functions will be introduced for the study of such products including a quadratic one. The quadratic Lyapunov functions enable us to characterize the ergodic properties of a general class of stochastic matrices and provide a unified approach to convergence analysis of distributed averaging dynamics.

Speaker Bio: Behrouz Touri is an Assistant Professor with the Electrical, Computer, and Energy Engineering Department at the University of Colorado Boulder. He received his B.S. degree in Electrical Engineering from Isfahan University of Technology, Isfahan, Iran in 2006, his M.S. degree in Communications, Systems, Electronics from Jacobs University, Bremen, Germany in 2008, and his Ph.D. degree in Industrial Engineering from the University of Illinois at Urbana-Champaign in 2011. Between 2011 and 2014, he was a postdoctoral researcher with the ECE department of the University of Illinois and Georgia Institute of Technology. His research interests include applied probability theory, distributed optimization, control and estimation, population dynamics, and evolutionary game theory. He is a 2014 IEEE Transactions on Automatic Control's outstanding reviewer.

Datacenter Computers: Modern Challenges in CPU Design

Richard Sites>
Google

Yue Lu
Harvard School of Engineering and Applied Sciences

Resolving individual photons in space and time is the holy grail of optical sensing. Recent advances in materials, devices and fabrication technologies have led to an emerging class of solid-state sensors with single photon sensitivity. Thanks to their sub-nanosecond time resolution, and rapidly improving spatial resolution, these new single-photon sensors (SPS) have been a key enabling technology behind recent breakthroughs in several domains, including fluorescence-based bio-imaging, time-of-flight 3D computer vision, LIDAR, high-speed videography, and astronomy.

In this talk, I will present models, theory, and algorithms in signal sampling and inference to address several challenges associated with the SPS. In particular, I will present our work on establishing the performance bounds of the SPS in acquiring light intensity fields; on time-sequential adaptive sensing schemes that allow one to push the imaging capabilities of SPS systems beyond the nominal limit imposed by current hardware; and on new image formation algorithms that can efficiently "decode" the massive bitstreams generated by the SPS.

Speaker Bio: Yue M. Lu attended the University of Illinois at Urbana-Champaign, where he received the M.Sc. degree in mathematics and the Ph.D. degree in electrical engineering, both in 2007. He was a Research Assistant at the University of Illinois at Urbana-Champaign, and has worked for Microsoft Research Asia, Beijing, and Siemens Corporate Research, Princeton, NJ. From September 2007 and September 2010, he was a postdoctoral researcher at the Audiovisual Communications Laboratory at Ecole Polytechnique Federale de Lausanne (EPFL), Switzerland. He is currently an Assistant Professor of Electrical Engineering at Harvard University, directing the Signals, Information, and Networks Group (SING) at the School of Engineering and Applied Sciences. His work received several awards, including the Most Innovative Paper Award of IEEE International Conference on Image Processing (ICIP) in 2006, the Best Student Paper Award of IEEE ICIP in 2007, the Best Student Presentation Award at the 31st SIAM SEAS Conference in 2007. Student papers supervised and coauthored by him won the Best Student Paper Award of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) in 2011, and the Best Student Paper of IEEE Global Conference on Signal and Information Processing (GlobalSIP) in 2014.

On Abstraction, Information hiding, and Crosscutting Contract Modularity

Henrique Rebelo
Universidade Federal de Pernambuco
Brazil

In this talk, I will give a literature perspective on abstraction, information hiding, modular reasoning, crosscutting contract modularity, and their relation to Design by Contract languages (DbC), specially JML. I will than critically show how existing DbC languages fall short in addressing these concepts and I'll explain/demonstrate how our DbC language, called AspectJML, can fixe/improve the modularity issues in those DbC languages (using JML as example). Finally, I'll conclude with a speculative outlook at the future of crosscutting contract modularity.

Speaker Bio: Henrique Rebelo is an assistant professor at the Universidade Federal de Pernambuco, Brazil. A primary theme of his work is focused on improving the productivity of programmers by allowing them to write better programs that, as much as possible, look like their design and constraints. He believes that programs that clearly express the interfaces they implement are more fun to develop and maintain. Currently, he is leading the development of the aspect-oriented extension to the JML design by contract language, called AspectJML. While at Microsoft Research (as a research intern), he was involved to the development of SymDiff, a language-agnostic tool for equivalence checking to reason about imperative program changes.

Bridging the Gap between Reat-Time Inverter Control and Steady State Oprimization in Distribution Systems

Emiliano Dall'Anese
NREL
Golden, CO

In an effort to bridge the temporal and spatial gap between real-time control and network-wide steady-state optimization, this talk outlines one way to addresses the synthesis of feedback inverter controllers that provably steer the inverter output powers towards solutions of relevant optimization problems for example, the controllers continuously seek solutions of steady-state AC optimal power flow (OPF) renditions that encapsulate economic objectives and voltage-regulation constraints. The guiding motivation is to ensure that renewable-energy systems operation and control strategies are adaptable to changing ambient conditions and loads, and enable seamless end-user participation without compromising system efficiency. The design of the controller is grounded on dual epsilon-subgradient method, whereas semidefinite programming relaxations are advocated to bypass the NP-hardness of AC OPF problems. Throughout the talk, global convergence of inverter output powers will be demonstrated under mild technical conditions, for the case where affordable computational limits and communications constraints involve asynchronous updates of the controller signals. A distribution-system-specific rendition of the OPF, which accommodates ancillary-service-oriented objectives, will be also outlined.

Speaker Bio: Emiliano Dall'Anese received the B.Sc Degree M.Sc Degree from the University of Padova, Italy, in 2005 and 2007, respectively, and the Ph.D. in Information Engineering from the Department of Information Engineering, University of Padova, Italy, in 2011. From January 2009 to September 2010, he was a visiting scholar at the Department of Electrical and Computer Engineering, University of Minnesota, USA. From January 2011 to November 2014, he was a Postdoctoral Associate at the Department of Electrical and Computer Engineering and Digital Technology Center, University of Minnesota, within the group directed by Prof. Georgios B. Giannakis. Since December 2014 he has been a Senior Engineer within the Distributed Energy Systems Integration (DESI) group at the National Renewable Energy Laboratory (NREL). His research is centered around the analysis, algorithms, and application of Optimization and Statistical Signal Processing methods to energy management in power systems, cyber-physical systems, and grid informatics.

Video Content Analysis at YouTube

Manfred Georg
Google

Theresa Wellington
Google

Julien Langou
University of Colorado
Denver, CO

We consider communication lower bounds for matrix-matrix multiplication in the sequential case (number of messages and total volume of messages). Our new proof technique improves the known lower bound for the number of reads and writes in the sequential memory case. This lower bound can be adapted to hierarchical memory and parallel distributed memory cases.

Speaker Bio: Julien Langou is Professor in the Department of Mathematical and Statistical Sciences at the University of Colorado Denver. He received a B.Sc. degree in propulsion engineering from Ecole National Superieure de l'Aeronautique et de l'Espace (SUPAERO), France, and a Ph.D. degree in applied mathematics from the National Institute of Applied Sciences (INSA), France. His research interest is in numerical linear algebra with application in high-performance computing.

Two Projects in BioInspired AI: (1) Evolving Neural Networks that are Modular, Hierarchical, and Regular, and (2) Robots that Recover from Damage

Jeff Clune
Computer Science Dept
University of Wyoming

This talk summarizes two different, but related research projects.

(1) Evolving Neural Networks that are modular, hierarchical, and regular. A major open question in biology is how evolution produced the complexity seen in nature. I study this question by attempting to evolve such complexity in computational simulations of evolution, specifically in computational brain models called neural networks. To date, neural network phenotypes evolved in silico tend to be simple and lack much of the structural organization found in animal brains. I will describe new approaches that combine computational abstractions of evolution and developmental biology to automatically produce neural networks that are modular, hierarchical, and regular. Moreover, I will show that these properties increase the speed of evolution and level of intelligence it generates.

(2) Robots that Recover from Damage. While animals can quickly adapt to injuries, current robots cannot "think outside the box" to find a compensatory behavior when they are damaged. I will introduce an intelligent trial and error algorithm that allows robots to adapt to damage in less than two minutes. Before deployment, a robot exploits a novel algorithm to create a detailed map of the space of high-performing behaviors: This map represents the robot's intuitions about what behaviors it can perform and their value. If the robot is damaged, it uses these intuitions to guide a trial-and-error learning algorithm that conducts intelligent experiments to rapidly discover a compensatory behavior that works in spite of the damage. Experiments reveal successful adaptations for a legged robot injured in five different ways, including damaged, broken, and missing legs, and for a robotic arm with joints broken in 14 different ways.

Speaker Bio: Jeff Clune is an assistant professor in the Computer Science Department at the University of Wyoming, where he directs the Evolving Artificial IntelligenceI lab (http://EvolvingAI.org). His lab focuses on creating artificial intelligence in neural networks, either via evolutionary algorithms or deep learning. His lab also studies how evolution produced the complex, intelligent, diverse life on this planet by trying to computationally recreate it. He has a bachelor's degree in Philosophy from the University of Michigan, a Master's degree in philosophy and a Ph.D. in computer science from Michigan State University, and was a Postdoctoral Fellow at Cornell University funded by an NSF Postdoctoral Fellowship.

Bayesian Hierarchical Topographic Clustering Method Motivated From the Self-Organizing Map

Wen Zhou
Statistics Dept
Colorado State University

The Kohonen's self-organizing map (SOM) has been widely applied as an unsupervised clustering method for a variety of advantages including the spontaneous vector quantization and topographic preservation. Inspired by the topographic preservation of SOM, this article establishes a topographic clustering method using Bayesian hierarchical models. Stationary Gaussian processes are proposed to model the core self organization process on latent grid on Z^2 and they are proposed as the prior distributions for the latent prototypes of data. To accommodate data with possible complex dependency, we model the data covariance structure by the principal component approximations. Using the distribution theory on special manifold, the heterogeneous part of the covariance matrices have the Bingham-von Mises-Fisher distribution as their conditional posterior distributions that reduces the numerical difficulties for sampling. Latent labeling variables bridge the data model and prototype model, and are projected onto the latent grid to provide topographic clustering of data. Markov Chain Monte Carlo (MCMC) methods are exploited for sampling the posterior distribution of the clustering rules, from which an optimal rule is defined by minimizing an empirical risk function that accounts for both data partition and topographic preservation. Numerical studies on simulated and empirical data demonstrate the satisfactory performances of the newly proposed topographic clustering method compared with other clustering algorithms. Convergence of the numerical algorithm is studied. The proposed method has intrinsic connection with Markov Random Fields, kernel smoothing and the original SOM, which are also extensively explored.

Speaker Bio: Wen Zhou is an assistant professor in Statistics. He obtained Ph.D. in statistics in May 2014 from Iowa State University and Ph.D. in applied mathematics in Aug 2010 in ISU too. His research focus on statistical machine learning, high dimensional inference, functional data analysis and applications of new statistical methods in genomic analysis and biology problems.