This page is created to inform any interested participants regarding the seminars to be conducted for the EE690 Graduate Seminars course. The schedule and the content will be updated throughout the semester.
2024-2025 Fall Graduate Seminars
| Date | Time | Location | Speaker | Institution & Country | Topic |
|---|---|---|---|---|---|
| Oct 18 | 12:40 | D-231 | Aykut Koç | Bilkent University, Ankara, Türkiye | Devising Transformers as an Autoencoder for Unsupervised Multivariate Time Series Imputation |
| Nov 8 | 17:30 | Online | Kaan Sel | Massachusetts Institute of Technology, Cambridge, MA, USA | Digital Medicine for Cardiovascular Health |
| Nov 15 | 12:40 | D-231 | Halil Ersin Söken | METU, Ankara, Türkiye | Fault-tolerant Attitude Determination and Control System Design for Small Satellites |
| Nov 29 | 12:40 | D-231 | Metin Aktaş | Aselsan, Ankara, Türkiye | Robustness of Artificial Intelligence Systems |
| Dec 5 | 12:40 | D-231 | Çağrı Çetintepe | Aselsan, Ankara, Türkiye | A Quick Primer on Frequency Diverse Arrays |
| Dec 12 | 15:40 | D-231 | Güner Çelik | Scotiabank, USA | Advanced Education and Its Value in Quantitative Finance and Technology Careers |
| Dec 13 | 12:40 | Online | Uğur Teğin | Koç University, İstanbul, Türkiye | Tunable and Scalable Optical Computing with Linear and Nonlinear Optical Dynamics |
Title: Devising Transformers as an Autoencoder for Unsupervised Multivariate Time Series Imputation
Abstract: Time-series data processing is essential across various fields, including healthcare, transportation, and weather forecasting. Multivariate time-series data, in particular, exhibit a correlation pattern over a common independent variable. This is illustrated by concurrent sensor readings in applications like autonomous driving or multiple channels in data collection devices used in medical diagnoses. However, the increasing incidence of data acquisition failures, including sensor malfunctions and human errors, results in gaps and substantial loss of information. We propose a novel method called Multivariate Time-Series Imputation with Transformers (MTSIT) to tackle these challenges. This method employs an unsupervised autoencoder model with a transformer encoder to leverage unlabeled observed data for simultaneous reconstruction and imputation of multivariate time series. The MTSIT strategy presents an input sequence with gaps (missing patterns) to the transformer encoder. The final encoder block produces an output sequence that is linearly transformed into the imputed sequence. The Mean Squared Error (MSE) is subsequently computed between the missing values and their predicted imputations, guiding the network’s training toward minimizing the MSE.
Bio: Aykut Koç received the B.S. degree from the Electrical and Electronics Engineering Department, the M.S. and Ph.D. degrees in electrical engineering from Stanford University, Stanford, CA, USA, in 2005, 2007, and 2011, respectively. He is a Faculty Member with the Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey. He has authored or co-authored more than 80 research papers and one book chapter and issued five patents. His research interests are machine learning and signal processing, extending into natural language and graph signal processing. He currently serves as an Associate Editor of IEEE SIGNAL PROCESSING LETTERS, IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, and IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY. He is the current Chair of IEEE Signal Processing Society (SPS) Turkiye Chapter. He received the 2023 Science Academy Young Scientists Award (BAGEP).
Title: Digital Medicine for Cardiovascular Health
Abstract: Over the past decade, the concept of real-time sensing has seen unprecedented advances, where a myriad of individuals has now access to wide choices of smart commercial off-the-shelf (COTS) wearables that can easily monitor vital biometrics such as heart rate on a continuous basis. Despite these advancements, extracting complex cardiovascular parameters from wearable device measurements for precision medicine remains a challenge, due to several unmet needs; limited availability of advanced sensing paradigms, substantial physiological heterogeneity among individuals, limited access to ground truth data at personalized levels for disease states, the data-intensiveness of artificial intelligence (AI)-based modeling of complex input (sensor measurements) and output (complex cardiovascular parameters) relationships, the disparity between computational model parameters and sensor measurements.
This seminar presents several topics towards developing advances sensors and algorithms coupled with human physiology to enable continuous and unobtrusive monitoring of complex cardiovascular health parameters. We will present simulations on electrical models of biological tissues to drive the design, optimization, and placement of sensors. Additionally, we will discuss novel sensing paradigms utilizing novel form-factors: rings, electronic tattoos, distributed patches, harnessing the deep tissue sensing capabilities of bioimpedance, achieving medical-grade accuracies in blood pressure (BP) estimation. Furthermore, we will introduce the concept of physiology-driven AI modeling, which leverages our existing knowledge of human physiology and real-time measurements to uncover hidden and complex BP information, reducing the dependence on ground truth data.
The integration of next-generation wearables and AI will have a significant impact on precision medicine, revolutionizing traditional medical practices that heavily rely on outdated, bulky, and invasive systems. By motivating the use of AI and advanced wearable sensors, in view of human physiology, we aspire to usher in a new era of personalized, effective, and accessible medical care. With this seminar we will cover key challenges of extracting clinical parameters from wearable device measurements and potential solutions through bioimpedance sensors, physiology-driven sensing paradigms, and AI modeling, to provide an understanding of the advancements in cardiovascular health monitoring and the potential for personalized, accessible medical care.
Bio: Dr. Kaan Sel is a postdoctoral associate at the Laboratory for Information & Decision Systems, Massachusetts Institute of Technology, Cambridge, MA. He earned his Ph.D. degree in computer engineering from Texas A&M University, as well as his B.Sc. and M.Sc. degrees in electrical and electronics engineering, and a B.Sc. (double major) degree in physics from Middle East Technical University in Ankara, Turkey. His research focuses on wearable and mobile systems, computational cardiovascular modeling, and scientific machine learning for digital health. Dr. Sel has authored over 20 papers published in reputable journals and conferences and holds multiple patent applications. Dr. Sel has served as a reviewer for prestigious journals including Nature, ACM, and IEEE, as well as for various IEEE conferences. He has received the runner-up award for the IEEE Engineering in Medicine and Biology Society student paper competition in 2019. Additionally, Dr. Sel has been recognized for his academic excellence, receiving the Charles Bowman and Lynn Holleran International Endowment Academic Excellence Award in 2022 and the Shibata International Memorial Scholarship in the same year.
Title: Fault-tolerant Attitude Determination and Control System Design for Small Satellites
Abstract: Today, we are seeking faster progress in space activities. New mission concepts, led by cheap and affordable small satellites, are expanding the possibility of space research to more people. As a consequence, spacecraft attitude determination and control (ADC) have become an even more attractive research field. Despite the shrinking sensors and actuators, we need to propose solutions for ADC systems that are as accurate as those for larger spacecraft. One of the major challenges for small satellite ADC systems is surely the vulnerability of the system to faults. A small satellite is more open to system faults due to its small, commercial off-the-shelf equipment, which is used without comprehensive testing procedures. The aim of this talk is to give an overview of the challenges we face when designing the ADC system for small satellites and then review the approaches for designing a fault-tolerant ADC system. Specifically, we will discuss the multi-algorithmic hybrid approaches that integrate the ADC algorithms with the Fault Detection Isolation and Recovery (FDIR) schemes.
Bio: Personal webpage.
Title: Robustness of Artificial Intelligence Systems
Abstract: For Artificial Intelligence (AI), robustness describes the ability of a system to maintain its level of performance under a variety of circumstances. Developing and verifying high quality models through machine learning faces particular challenges. Generally recognised conditions that most AI might need to be robust to include:
- Uncertainty in training and operational data;
- Inputs that are different from the training set, yet consistent with the training population statistically or semantically;
- Inputs that are outside the training population;
- Adversarial action.
Over the past few decades, it has been shown that machine learning models based on deep learning techniques can achieve and even surpass human-level performance in a variety of tasks. On the other hand machine learning models are often vulnerable to perturbation of their input and can easily be fooled to yield incorrect output. These types of manipulations are referred to as adversarial attacks and the performance of machine learning models against these attacks are measured as adversarial robustness. Adversarial robustness is investigated in two different camps. In the first camp, the researchers try to find a method to generate adversarial attacks to decrease the robustness of the models the most. The researchers in the second camp try to find better training or defensive methods that make the network architectures more robust to such adversarial attacks. In this talk, we survey the methods for adversarial attacks and defences and we define the metrics and measurement methods of the adversarial robustness from current literature.
Bio: Metin Aktaş is a Senior Chief Engineer at Aselsan Inc., where he has been since 2005. From 2005 to 2014 he served as avionic software engineer at Microelectronics, Guidance and Electro-Optics Division of Aselsan. He is currently working at Defence Systems Technologies Divison in the same company. He received a B.Sc. in Electrical and Electronics Enginnering Department from Dokuz Eylul University, Izmir, Turkey in 2002. He received his M.Sc and Ph.D. in Electronics Enginnering Department from Middle East Techinal University, Ankara, Turkey in 2004 and 2012. His research interests are array signal processing, computer vision, machine learning, deep learning, remote sensing, security systems.
Title: A Quick Primer on Frequency Diverse Arrays
Abstract: Frequency diverse array (FDA) concept is one particular manifestation of waveform diversity, a topic which gains popularity among the research community in response to pressing demands of modern radar and communication systems such as multiple tasks, capacity and security. Introduction of a set of waveforms to the elements of an array, with slight shifts in the carrier frequencies so to establish frequency diversity, leads to far-field patterns that are remarkably different from conventional phased arrays, and opens up new possibilities like simultaneous missions, automatic beam scanning, range filtering, and flexible beamforming.
Being a relatively new topic and lacking experimental demonstration in practical scenarios, the FDA concept lends itself to research opportunities from both theory and implementation perspectives. This talk will cover the FDA concept in a nutshell, and will touch upon sample applications which would benefit from aforementioned potential.
Bio: Cagri Cetintepe received the B.Sc., M.Sc., and Ph.D. degrees (Hons.) in electrical and electronics engineering from Middle East Technical University (METU), Ankara, Turkey, in 2007, 2010, and 2015,respectively. From 2007 to 2009, he was a Research Assistant with the METU-MEMS Application and Research Center, and focused on RF MEMS research. He was a Teaching and Research Assistant with the Electrical and Electronics Engineering Department, METU, from 2010 to 2016, where he extended his focus to radar and communication systems within the scope of his Ph.D. studies, which were also supported with a scholarship from ASELSAN Inc., Turkey. From 2016 to 2018, he was a Post-Doctoral Researcher at the School of Electrical and Electronic Engineering, University College Dublin (UCD), Dublin, Ireland, where he conducted research on mm-wave integrated circuits, chip/package-integrated antennas, and quantum computing applications.
Dr. Cetintepe joined ASELSAN Inc. in 2019. He initially worked as a lead research engineer at ASELSAN Research Center, and participated in various R&D projects for defense applications, with particular focus on phased arrays, antenna design, advanced radar algorithms, alternative navigation methods, and magnetic anomaly detection. Since 2023, he has been a chief antenna design engineer at Electronic Hardware Design Directorate, Communication and Information Technologies Business Sector, ASELSAN.
Dr. Cetintepe's scientific interests include RF MEMS switches and phase shifters, surface micromachined lumped components, antennas implemented with various technologies, frequency diverse and phased arrays, RF/mm-wave CMOS transceivers, microwave and mm-wave instrumentation, and radar/communication applications.
Title: Advanced Education and Its Value in Quantitative Finance and Technology Careers
Abstract: Choosing a rewarding and fulfilling career in today’s fast-changing social and technological landscape is more challenging than ever. Making the right educational, professional, and financial investments can significantly accelerate progress in this pursuit. As a fellow METU graduate who obtained advanced degrees from MIT and went on to take leadership roles in Quantitative Finance and High Tech as well as founding multiple technology-driven startups, I’d like to share some of my personal experiences and professional choices that lead to defining moments in my career. Furthermore, I’d like to describe some high-profile occupations within these fields from the perspective of an insider, including different techniques and approaches used in practice.
Techniques used in quantitative trading range from traditional techniques in fundamental fields such as Probability Theory, Linear Algebra, and Statistical Methods to newer and cutting-edge models in Machine Learning and broader Artificial Intelligence. Commonly utilized methods involve Dynamic Programming and Optimal Control, Stochastic Differential Equations, Utility Maximization and Portfolio Optimization, Markov Decision Theory, Boosted Regressions and Ensemble Tree Methods, and various Deep Learning algorithms.
Bio: Guner Celik is Head of Quantitative Trading in Fixed Income at Scotiabank where he leads the quantitative research and algorithmic trading efforts for Rates, Credit, and Derivatives.
Guner joined Scotiabank from Cantor Fitzgerald where he led Rates Algorithmic Trading and Research and prior to that he was a co-founder and head of research and trading for Springtech Capital, a systematic trading hedge fund. Guner joined Springtech from Goldman Sachs where he was a VP of Rates Systematic Market Making, and prior to Goldman, he was a Senior Mathematical Modeling and Algorithms Specialist at Oracle. He has extensive experience in quantitative and algorithmic trading strategies, mathematical modeling, predictive signal generation, electronic trading, market making, and a strong background in Probability theory, Algorithms, Optimization, and Machine Learning.
Guner holds both Ph.D. and M.Sc. degrees in Electrical Engineering and Computer Science from MIT and he has a B.Sc. in Electrical and Electronics Engineering from METU, where he was ranked 2nd in the engineering department.
Title: Tunable and Scalable Optical Computing with Linear and Nonlinear Optical Dynamics
Abstract: Optical computing, a branch of analog computing, uses photons rather than traditional electronic signals to perform computing operations. Optical systems excel at parallel processing, allowing multiple operations to co-occur, which makes them ideal for heavy data processing. Propagation of photons through optical elements performs computation intrinsically by offering fundamental operations like Fourier transform, matrix multiplication, and convolution. Thus, optical computing has a rich history closely following the evolution of neural networks and AI. In this seminar, I will present our past and present studies to develop scalable, energy-efficient, and nonlinear optical computing tools for machine learning studies.
Bio: Uğur Teğin is an Assistant Professor of Electrical and Electronics Engineering at Koç University. He holds a BSc in Physics (2015) and an MSc in Material Science and Nanotechnology (2018) from Bilkent University in Ankara, Turkey. He earned his PhD in Photonics (2021) from École Polytechnique Fédérale de Lausanne (EPFL), Switzerland. He continued his postdoctoral research at the California Institute of Technology (Caltech) in the Departments of Electrical Engineering and Medical Engineering. His research spans nonlinear optics, optical computing and machine learning, fiber optics and lasers, ultrafast optics, and imaging. He has contributed to 18 journal publications and holds three patents in these areas. He received the Optica Foundation's 2024 Challenge Prize in Information.