13 May 15.00
Dr. Derya Malak
Massachusetts Institute of Technology, MA, USA

ABSTRACT: This talk focuses on the central problem of coordinating computation and caching in networks, using some recent results in stochastic geometry and information theory. Our goal is to provide a FAST, RELIABLE, and CHEAP design for 5G mobile networks. The first part of the talk focuses on decentralized caching by utilizing the redundancy across multiple, geo-dispersed, and mobile sources of data. In order to leverage proximity-based communications such as peer-to-peer systems or device-to-device communications, we exploited the spatial diversity of the content and the topology as a proxy for optimizing cache placement. We proposed novel decentralized and spatial exclusion-based cache placement policies. These policies promote diversity and reciprocation (FAST); provide guarantees on the cache hit probability (RELIABLE); and offload traffic from congested base stations, and are promising for proximity-based applications (CHEAP). The second part of the talk concerns with the limits of reliability with imperfect feedback when coding, and development of scalable and robust routing solutions for connectivity in wireless mesh networks. This approach utilizes coding for optimizing the tradeoff between in-order delivery delay and throughput, which is promising for computing systems such as the Internet of things, and ultra-reliable and low-latency communications e.g. mission-critical communications, and connected vehicles in 5G networks (FAST). It also provides robustness and delay guarantees (RELIABLE); and has very low complexity in terms of coding overhead, and is cost effective via the use of multi-hop WiFi links (CHEAP). Finally, this talk describes a new perspective on cloud/fog computing, by coordinating caching and computation in order to handle the large volume of data with growing computational demand. Our goal is to devise coding techniques for functional compression, and coordinating computation and caching in networks, by employing the concepts of graph entropy and function surjectivity. These techniques suit different applications such as caching, classification, federated learning, quantization, and compressed sensing. Our unified insights suggest to cache at the edge (FAST); distribute storage by exploiting geographic diversity and paths (RELIABLE); and distribute computation by making use of underlying redundancy both in data and functions, in order to recover a sparse representation, or labeling (CHEAP).



Derya Malak is a Postdoctoral Associate at the Massachusetts Institute of Technology and Northeastern University, where she has been working with Prof. Muriel Médard and Prof. Edmund Yeh, respectively. She received a Ph.D. in Electrical and Computer Engineering at the University of Texas at Austin under the supervision of Prof. Jeffrey G. Andrews, in August 2017, where she was affiliated with the Wireless Networking & Communications Group (WNCG). Previously, she received an M.S. degree in Electrical and Electronics Engineering at Koc University, Istanbul, Turkey, in February 2013. She received a B.S. in Electrical and Electronics Engineering (with minor in Physics) at Middle East Technical University, Ankara, Turkey, in June 2010. Derya has held summer internships at Huawei Technologies, Plano, TX, and Bell Laboratories, Murray Hill, NJ. She was awarded the Graduate School fellowship by the University of Texas at Austin between 2013-2017. She was selected to participate in the Rising Stars Workshop for women in EECS, MIT, in October 2018.

17 Dec 15.00
Zafer Dogan, PhD
Postdoctoral Research Associate, Harvard University


We are experiencing a data-driven revolution at the moment with data being collected at an unprecedented rate. In particular, there is an increasing excitement toward systems with learning capabilities. Several data-driven applications have already shown practical benefit of having access to more data. However, large acceptance of such systems heavily depends on stability, tractability and reproducibility features, where current systems fall inadequate to deliver. The scale and complexity of these modern datasets often render classical data processing techniques infeasible, and therefore, several new algorithms in signal processing (SP), optimization, and machine learning (ML) are required to address new technical challenges associated with the nature of the data. In this talk, I will present computational methods in data analytics using sparse signal models and my recent work on analyzing the exact dynamics of iterative algorithms in nonconvex setting that can address aforementioned challenges with examples from low level SP to non-convex ML problems. In the first part of my talk, I will focus on novel computational methods for sampling and reconstruction of signals and images using sparse signal models. In particular, foundations of SP are heavily based on powerful sampling theorems for acquisition, representation and processing. However, increasing evidence shows that the requirements imposed by sampling theorems are too conservative for many naturally-occurring signals, which can be accurately characterized by sparse representations that allows to operate at lower sampling rates closer to the signal’s intrinsic information rates. Finite rate of innovation (FRI) is a new theory that allows to extract underlying sparse signal representations while operating at a reduced sampling rate. In this part, I will present data-driven reconstruction techniques for FRI signals from both theoretical and practical points of view. Specifically, the presentation will cover applications that involve temporal and spatial localization of events in neuroimaging, and in nonlinear tomography for radiating fields. In the second part of my talk, I will present my recent work on analyzing, in the high-dimensional limit, the exact dynamics of iterative algorithms for solving non-convex optimization problems that arise in signal estimation. More specifically, for solving a broad range of learning problems in SP and ML, non-convex optimization problems receive increased interest due to their favorable computational and statistical efficiency. Fortunately, there exists broad range of iterative algorithms (e.g., stochastic-gradient descent (SGD)) with great empirical success. However, there is a lack of true characterization of the dynamics of these methods. Here, I will show how the very high-dimensional settings allow one to apply powerful asymptotic methods to obtain precise characterization. For concreteness, I will focus on the prototypical problem of principal component analysis in an online and distributed setting. I will show that the time-varying empirical measures of the estimates given by the algorithms converge weakly to a deterministic limiting process in the high-dimensional limit. Moreover, this limiting process can be characterized as the unique solution of a nonlinear ODE, and it provides exact information regarding the asymptotic performance of the algorithms. For example, performance metrics such as the MSE, and the cosine similarity can be obtained by examining the deterministic limiting process. A steady-state analysis of the ODE also reveals interesting phase transition phenomena related to the performance of the algorithm. I will conclude my talk by briefly presenting the application of similar analysis techniques to other nonconvex optimization problems.



Dr. Dogan is currently a postdoctoral research associate at the John. A. Paulson School of Engineering and Applied Sciences at Harvard University. His main research interests are at the intersection of data analytics, optimization, inverse problems, and machine learning. Apart from the theoretical aspects, his research goals bear practical features in sensor networks, emerging imaging modalities, mining in high dimensional data, and applications of neural networks in imaging and inverse problems. He received his PhD and MSc degrees in electrical engineering from EPFL, Switzerland, in 2015 and 2011, respectively, and his BSc degree in electrical and electronics engineering from METU, Turkey, in 2009.

12 Apr 15.00
Sedat Nizamoğlu
Electrical and Electronics Engineering, Koç University
Abstract: I will present a wide variety of novel optoelectronic devices for energy harvesting, neural interfaces, light generation and wound closure. Colloidal quantum dots offer attractive electronic, optical and surface properties. In the last decade indium-based colloidal quantum dots have attracted significant attention as a biofriendly alternative to cadmium-based ones with their tunable electrical and optical properties. I will demonstrate biofriendly quantum dots based luminescent solar concentrators for low-cost, large-area and high-efficiency energy harvesting and neural interfaces for nervous system diseases. Moreover, I will discuss the formation and characteristics of self-assembled biopolymer- based lasers that are formed via coffee stain effect. In addition, I will discuss bioabsorbable optical waveguides for wound closure.
About the Speaker: Sedat Nizamoglu received his B.Sc. degree in Electrical and Electronics Engineering (EEE) in 2005, M. Sc. degree in Physics as a Valedictorian in 2007 and his Ph.D. in EEE in 2011 at Bilkent University. Immediately after graduation, he continued as a research fellow with a joint affiliation with Harvard Medical School and Wellman Center for Photomedicine, Massachusetts General Hospital in USA. Before joining Koç University, he was a faculty member at Özyeğin University. His research focuses on the demonstration of innovative devices and interfaces for the applications to energy, medicine, and environment. He has published more than 40 research papers in prestigious journals including Nature Communication, Nature Photonics, Advanced Materials and Nano Letters. He was recognized by MIT Technology Review as Innovator Under 35 Turkey, he recevied Outstanding Young Scientist Award by Turkish Academy of Sciences and Science Academy. In addition, he was awarded an ERC (European Research Council) Starting Grant.
12 Mar 15.00
Dr. Ahmet C. Durgun

Impact of Fiber Weaves On the Electrical Performance of Glass Epoxy Packages

Today’s microelectronics manufacturing industry is facing several challenges as the market marches towards smaller form factor and higher bandwidth products. One of the key challenges associated with these devices is to achieve the necessary electrical performance with sufficiently low-cost packaging solutions. In order to minimize packaging costs, engineers continuously attempt to shrink the size of the package, which imposes a set of challenges on maintaining the electrical performance. Furthermore, the ever-increasing bus speeds drive increasingly stringent requirements on the current and next-generation high-speed communication channels, both at the system and component levels. Some of the low-cost packaging technologies utilize glass epoxy substrates that are composed of glass fiber bundles and epoxy resin, which have different electrical properties. These differences result in variations in characteristic impedance and propagation speeds, which may be detrimental at high data rates. The insertion loss, within-pair skew, differential to common mode conversion ratio, and crosstalk of transmission lines may drastically increase due to the fiber weave effect. Consequently, the link budget of high-speed communication channels may be significantly hindered. In this seminar, the package level impact of fiber weaves on the electrical performance of high-speed buses will be addressed, in the light of current microelectronic demands, and some mitigation techniques will be discussed.



Ahmet Cemal Durgun received the B.S.E.E. and M.S.E.E. degrees from Middle East Technical University, Turkey, Ankara, in 2005 and 2008, respectively, where he completed the double major program in mathematics and received the B.S. degree, in 2006. He completed his Ph.D. studies, related to flexible antennas and high impedance surfaces, at Arizona State University, Tempe, AZ, USA, in 2013. Currently, he is with the Intel Corporation, Assembly and Test Technology Development Department, working as a Sr. Analog Engineer. His research interests include microelectronic packaging, design of high-speed communication channels, high routing density interconnects, signal integrity, and applied electromagnetics.

05 Mar 15.00
Dr. Eren Balevi
The University of Texas at Austin, Electrical and Computer Eng. Dept.

Designing and managing large-scale wireless networks using stochastic geometry and machine learning are discussed for one intriguing network architecture, which is composed of cloud and fog nodes, and dubbed as cloud-fog-thing network architecture, that is under consideration for 5G. Some important points of this architecture that are the optimum number of fog nodes and their locations are determined in an attempt to increase the average data rate and decrease the transmission delay using stochastic geometry and machine learning. Interestingly, these results may be adapted to increase the understanding of the role of spinal cord plasticity in learning when brain and spinal cord are treated as a cloud and fog network, respectively. This may ultimately suggest new means of treating central nervous system disorders associated with the spinal cord plasticity. Based on this duality, a modified coded caching policy is proposed as well for wireless networks inspired from the central nervous system, i.e., from the cooperation between brain and spinal cord. Lastly, the importance of machine learning in wireless communications and their application areas in large-scale wireless networks are emphasized in this seminar.



Eren Balevi received his B.Sc., M.Sc. and Ph.D degree from the Electrical and Electronics Engineering Department of Middle East Technical University. He worked as an instructor in the Electrical Engineering Department of University of South Florida. Now, he is working as a post-doctoral researcher in the University of Texas, Austin in the wireless networking and communications group.  


26 Feb 15.00
Dr. Alper Koz
ODTÜ, İmge Analiz Merkezi

26 Feb 15.00
Dr. Alper Koz
ODTÜ, İmge Analiz Merkezi

21 Feb 9.00
Günay Turan

Tez Danışmanı     : Prof. Dr. Z. Yasemin Kahya
Tez Eş Danışmanı: Prof. Dr. Mehmet Akar

Yer: Yorgo Istefanopulos Seminer Salonu