CSE

Speaker

Elias S. Manolakos
Room: ITEB336
Time: 4/3/13 2-3 pm

Abstract

Rapid advances in Systems Biology call for the development of efficient methods and tools for analyzing the complex dynamics of large-scale biomolecular networks to unravel the “epigenetic landscape” of cells and predict how it can be manipulated by intervention strategies (e.g. combinatorial drug design for personalized medicine, stem cell therapies etc). Stochastic simulation is the method of choice for analyzing the dynamics of biological systems while accounting for their inherently stochastic behavior due to intrinsic and extrinsic “noise”. However, Stochastic Simulation Algorithms (SSA) are very demanding computationally and only small biomolecular networks can be analyzed in a reasonable amount of time today using a modern computer. Inferring from high throughput data “biological circuits” of increasing complexity and simulating their dynamical behavior under different conditions motivates the need for new flexible and efficient computation strategies that exploit modern hardware technologies to deliver scalable performance per unit power, without sacrificing simulation accuracy. Such solutions will allow us to analyze efficiently systems such as cross-talking metabolic pathways, whole-species gene regulatory networks, biofilm formation mechanisms etc. We present the design of a scalable Multiprocessor Systems-on-Chip architecture implementing Gillespie’s SSA in reconfigurable hardware. Our MPSoC architecture can deliver performance (Mega-Reactions/sec) and throughput (MReaction cycles/sec) that is increasing linearly with the number of processors in the SoC. It can handle the simulation of very large biomolecular networks with up to m = 16K reactions (of up to the 3rd order) using a moderate size FPGA. The MPSoCs can be configured to use the available processors to split the reactions of a Single Simulation run and execute them In Parallel (SSIP mode), or to execute Multiple independent Simulation runs In Parallel (MSIP mode). We have synthesized and verified several MPSoC instances with up to N=8 Processing Elements for Xilinx Virtex 5 and Virtex 6 FPGAs, reaching clock frequencies up to 320 MHz and delivering performance that exceeds by 2 orders of magnitude that of software based simulators running on Intel Core 2 and i7 CPUs at frequencies higher than 2GHz. In addition, we have developed a configurable fully parametric soft IP core of the architecture, expressed in VHDL, and a completely automated design flow that can be used to synthesize the most appropriate MPSoC instance for a given biomodel of any complexity (captured in SBML). Moreover, we have developed a Hardware Abstraction Layer API in Python which allows user applications running in a host PC to view the SoC running in the FPGA just as a component for efficient stochastic simulations. We are in the process of developing a StochSoCs based flexible simulation platform for systems biology modeling studies accessible over the internet.

Biography

Elias S. Manolakos is a Visiting Scholar at the Wyss Institute for Biologically Inspired Engineering, Harvard University and the Director of the Multidisciplinary Program "Information Technologies in Medicine and Biology", University of Athens. Before returning to Greece he was with the faculty of the ECE Dept. of Northeastern University, Boston, where he directed the Communications and DSP Center for Research and graduate studies, promoting student-centered innovation through academia-industry collaboration.

Speaker

Yong-Jun Shin
Room: ITEB336
Time: 1-2pm

Abstract

Although biological or living systems show unique features, they are fundamentally governed by the same physical laws that rule non-living systems. For example, molecular dynamics (statistical mechanics) and density functional theory (quantum mechanics) can be used to model not only carbon nanotubes but also DNA or protein molecules. However, what is often neglected is that some engineering tools, especially digital signal processing (DSP) techniques, can complement conventional physics-based approaches when modeling unique biological features, such as adaptation or robustness, that exhibit intelligence. In this talk, I suggest that various DSP techniques, including adaptive filtering (e.g., the Kalman filter) and digital feedback control, can be used to model intelligent features of biological systems. Considering biological complexity, computational models should always be validated using relevant experiments. Digital modeling approaches also benefit our integrative modeling/experimental efforts as they can nicely integrate with experimental data, which are mostly digital these days. Digital microfluidics (DMF) is a relatively new technology that involves manipulating discrete and independently controllable micro/nano liter droplets and its suitability as a true lab-on-a-chip platform has been recently proposed. In this talk, I present DMF as an integrative platform for studying intelligent features of biological systems digitally modeled.

Biography

Dr. Yong-Jun Shin is a postdoctoral associate in the School of Electrical and Computer Engineering at Cornell University. His current research interests include biological applications of estimation and control theory, multi-scale modeling of biological tissues and digital microfluidics/bioMEMS. He received his M.D. from the Seoul National University College of Medicine in Korea. He was a research associate at the Samsung Biomedical Research Institute (Seoul, Korea) before joining the Micro/Nano Devices and Systems (MiNDS) Lab at the University of Texas at Dallas (UTD) for his graduate work. He earned his M.S. and Ph.D. in electrical engineering from UTD, specializing in computational biology and digital microfluidics.

Speaker

Abstract

We consider achieving security in strong adversarial models that capture complex, realistic computing environments. In particular, we consider two settings that go beyond the scope of traditional cryptography.

In the first setting, we consider adversaries who gain physical control of a device with a secret stored on it (such as a smartcard or an iphone) and continuously tamper with the wires of the device, while observing the outputs. We would like to ensure that the secret state of the circuit is protected even in the face of such an attack. We present a compiler that converts any circuit into one that remains secure even if a constant fraction of its wires are continuously tampered with. We consider adversaries who may choose an arbitrary set of wires to corrupt, and may set each wire to 0 or to 1, or toggle with the wire. We prove that such adversaries can learn at most logarithmically many bits of secret information.

In the second setting, we continue the study of non-malleable cryptography, initiated by Dolev et al. (SIAM J. Comput., 2000). Here, we consider active adversaries who control and manipulate network traffic. We study non-malleable encryption schemes, which, in addition to traditional security against eavesdropping adversaries (called semantic security), guarantee that an active adversary cannot maul a ciphertext to create a new ciphertext encrypting a related message. We present the first black-box construction of a non-malleable encryption scheme from any semantically secure one. We thus resolve a complexity-theoretic question while achieving a more efficient construction that avoids the inherent inefficiencies of non-black-box techniques.

Biography

Speaker

Abstract

Geometry, topology, and algorithms combine to give new and interesting ways of understanding data. A growing field called topological data analysis (TDA) attempts to extract information about the shape of the distribution underlying a data set. Persistent homology is the main tool of TDA, and it gives a robust, multiscale view of the shape underlying data that has been applied successfully in biology, imaging, sensor networks, materials science, and machine learning. I will show how combining geometric algorithms with persistent homology yields dramatic improvements in efficiency. For example, many popular techniques in TDA that required data structures of size n^d (for d-dimensional data) can now be efficiently approximated with linear size data structures. Along the way, I will touch on several new results in computational geometry and topology and discuss the future of this growing area.

Speaker

Abstract

Biography

Speaker

Shuiwang Ji
Time: February, 21 : 1:15pm to 2:15pm
Room: ITE336
 

Abstract

The mammalian brain controls cognition, emotion, and perception and is one of the most complex yet least understood biological systems. It is known that there are at least several hundreds of distinct types of cells in the mammalian brain. These cell types are arranged into complex circuits, which ultimately are responsible for generating brain function.

In this talk, I will present our work on mining the gene expression and brain connectivity data from the Allen Brain Atlas (ABA), which provides genome-scale, cellular-resolution, three-dimensional gene expression and connectivity patterns in the mouse brain. I will discuss how unsupervised learning of the spatiotemporal gene expression data can lead to results that are consistent with classical neuroanatomy. I will then present our work on correlating and predicting the brain connectivity using gene expression signatures. Finally, I will describe our work on modeling the brain networks.

Biography

Shuiwang Ji received the Ph.D. degree from Arizona State University (ASU) in 2010 and is currently an Assistant Professor of Computer Science at Old Dominion University (ODU). His research interests include machine learning, data mining, and computational biology. He received the Outstanding Ph.D. Student in Computer Science Award from ASU in 2010 and the Early Career Distinguished Research Award from ODU's College of Sciences in 2012. His team received the first prize from the TRECVID Video Surveillance Evaluation in 2009 and the second prize on Demographic Prediction from the Nokia Mobile Data Challenge in 2012.

Speaker

Logan Everett
Time: February, 19 : 1:15pm to 2:15pm
Room: ITE336
 

Abstract

Metabolic homeostasis in mammals critically depends on the regulation of fasting-induced genes by the transcription factor CREB in the liver. Previous genome-wide analysis has shown that only a small percentage of CREB target genes are induced in response to fasting-associated signaling pathways. The precise molecular mechanisms by which CREB specifically targets these genes in response to alternating hormonal cues remain to be elucidated. Using data from chromatin immunoprecipitation coupled to high-throughput sequencing (ChIP-seq), I quantitatively compared the extent of CREB-DNA interactions in livers from both fasted and fed mice. This analysis demonstrated that CREB remains constitutively bound to its target genes in the liver regardless of the metabolic state, revealing a role for additional factors in specifying fasting-induced CREB target genes. Additionally, I developed a novel, robust framework for filtering ChIP-seq peaks, termed the ‘single sample independence’ (SSI) test that greatly reduced the number of false positive peaks, while still identifying thousands of novel CREB target genes. Integration of the CREB ChIP-seq data with expression microarray data and additional ChIP-seq data sets revealed that fasting-induced genes are specifically associated with complex regulatory elements bound by both CREB and additional transcription factors. Our results support a model in which CREB is constitutively bound to thousands of potential target genes and combinatorial interactions between DNA-binding factors are necessary to achieve the specific transcriptional response of the liver to fasting. The methods applied in this work can be used more broadly to explore the role of transcription factor cooperation in shaping the overall transcriptional regulatory network of the cell.

Speaker

Marco Molinaro
Time: 11am - noon
Room: ITE336

Abstract

This talk explores from different perspectives two main sources of difficulty in decision making: incomplete information and large dimensionality. In the first part, I will talk about optimization under uncertainty, more specifically resource allocation with item uncertainty. I will focus on the Online Packing IP model, where columns of the IP (i.e. items) come one-by-one in random order. This, and related models, have wide application in revenue management, e.g. airline booking and online advertisement allocation. Combining ideas from learning theory and geometric insights, we provide a strategy that is able to better cope with uncertainty and the first with guarantees that do not degrade as the number of items increases.

In the second part of the talk, I will address other perspectives on decision making. I will briefly discuss sublinear algorithms, which tradeoff the amount of information used to perform a computational task and the quality of the solution obtained. These are crucial in an increasing number of applications that involve massive data, ranging from biology to network analysis. Finally, I will discuss some of my work on Integer Programming, a classical tool for dealing with large decision spaces of combinatorial problems. Here, I will highlight our advances on the construction and analysis of cutting planes, a crucial piece of solving Integer Programs in practice, where we (partially) resolve several questions raised in the literature.

Speaker

Song Han
Room: ITE336
Time: Friday February, 8 1pm-2pm
 

Abstract

A cyber-physical system (CPS) is a system featuring a tight combination of, and coordination between, the system's computational and physical elements. A large-scale CPS usually consists of several subsystems which are formed by networked sensors and actuators, and deployed in different locations. These subsystems interact with the physical world and execute specific monitoring and control functions. How to organize the sensors and actuators inside each subsystem and interconnect these physically separated subsystems together to achieve secure, reliable and real-time communication is a big challenge.

In this talk, I will first present a TDMA-based low-power and secure real-time wireless protocol. This protocol can serve as an ideal communication infrastructure for CPS subsystems which require flexible topology control, secure and reliable communication and adjustable real-time service support. I will describe the network management techniques for ensuring the reliable routing and real-time services inside the subsystems and data management techniques for maintaining the quality of the sampled data from the physical world. To evaluate these proposed techniques, we built a prototype system and deployed it in different environments for performance measurement. I will also present a light-weight and scalable solution for interconnecting heterogeneous CPS subsystems together through a slim IP adaptation layer. This approach makes the the underlying connectivity technologies transparent to the application developers and thus enables rapid application development and efficient migration among different CPS platforms.

Biography

Song Han is currently a Postdoctoral Research Fellow in the University of Texas at Austin. He holds the B.S. degree in computer science from Nanjing University, the M.Phil. degree in computer science from City University of Hong Kong, and the Ph.D. degree in computer science from the University of Texas at Austin. His research interests include Cyber-Physical Systems, Real-Time and Embedded Sensing and Control Systems, Wireless Networks and Mobile Computing.

Speaker

Abstract

Large and complicated data sets have become ubiquitous in science and engineering, and have given rise to new challenges in Computer Science. In this new computational horizon, geometry has become an indispensable source of algorithmic tools, and ideas. A main catalyst behind this development has been the theory of metric embeddings. A metric embedding is a mapping from a metric space X into some metric space Y, which preserves all pairwise distances up to a small factor, called the distortion.

In this talk I will describe a general framework for obtaining low-distortion embeddings of complicated spaces into simpler ones. An immediate corollary is a general reduction from hard instances to corresponding simpler ones, for a large class of optimization problems on certain families of graphs. Moreover, these techniques yield improvements on the state of the art for problems such as Sparsest Cut, Treewidth, Crossing Number, Vertex Sparsification, 0-Extension, and the estimation of Laplacian eigenvalues for several interesting graph families.

Perhaps surprisingly, the above ideas can also be used to construct hard instances for certain problems. More precisely, guided by the above simplification machinery, we construct spaces that yield a nearly optimal integrality gap for a specific SDP relaxation of Sparsest Cut.

Biography