Main Page
CV
 Publications
 Research
Kaan Gokcesu
M.S. Student, Department of Electrical and Electronics Engineering
Bilkent University, Ankara 06800, Turkey
E-mail: gokcesu@ee.bilkent.edu.tr		  

Kaan conducts research broadly, but not limited to, in the following topics:

Machine Learning

In many real life applications, the volume of data to be handled has significantly increased due to the recent developments in information technologies. Aside from the difficulties of storing such massive amounts of data, the processing of data usually has to be done in real-time. Hence, the need for machine learning algorithms that can process the data on-the-fly with significantly low computational complexities is steadily growing. In this context, there exist significant practical and theoretical difficulties to sequential learning, since there is usually no or little knowledge about the statistical properties of the underlying signals or systems involved. Furthermore, the classical robust methods that are resilient to such challenges are overly conservative and usually static such that they provide profoundly inferior results on the average, deeming them practically useless.

In order to provide robust adaptive methods that also perform satisfactorily in real life applications, we introduce sequential learning algorithms that are mathematically guaranteed to work uniformly for all possible signals without any explicit or implicit statistical assumptions on the underlying signals or systems. We construct adaptive algorithms that sequentially perform as well as the best batch algorithm, for any signal, that had the ability to observe the entire data in advance. Unlike the state-of-the-art methods in the literature, our algorithms can tackle problems with non-stationarity in data modeling by sequentially learning the model structure. In this manner, without introducing any ad-hoc assumptions or parameters, we introduce algorithms that directly minimize any desired loss measure in a strong deterministic sense. Hence, our results are guaranteed to hold, not on the average, but in an individual sequence manner.

Publications on this topic include:

  1. N. D. Vanli, K. Gokcesu, M. O. Sayin, H. Yildiz and S. S. Kozat, "Sequential Prediction Over Hierarchical Structures," IEEE Transactions on Signal Processing, , vol. 64, no. 23, pp. 6284-6298, Dec. 2016. (IEEEXplore)
        
  2. K. Gokcesu, and S. S. Kozat, "An Online Minimax Optimal Algorithm for Adversarial Multi-Armed Bandit Problem," submitted to IEEE Transactions on Neural Networks and Learning Systems 2016. (pdf)
        
  3. K. Gokcesu, and S. S. Kozat, "A Rate Optimal Switching Bandit Algorithm," submitted to 25th European Signal Processing Conference (EUSIPCO2017). (pdf)
        
  4. K. Gokcesu, and S. S. Kozat, "A General Framework for Adversarial Bandits," submitted to 25th European Signal Processing Conference (EUSIPCO2017). (pdf)
        
  5. K. Gokcesu, and S. S. Kozat, "An Efficient Asymptotically Optimal Algorithm for Adversarial Bandits with Multiple Plays", to be submitted to IEEE Transactions on Neural Networks and Learning Systems, 2016. (draft available with permission of supervisor)
     
  6. K. Gokcesu, and S. S. Kozat, "Adversarial Bandits with Universally Holding High Probability Bounds", to be submitted to IEEE Transactions on Neural Networks and Learning Systems, 2016. (draft available with permission of supervisor)
     
  7. K. Gokcesu, and S. S. Kozat, "Unification of Bandit and Expert Framework with Feedback Cost", to be submitted to IEEE Transactions on Signal Processing, 2016. (draft available with permission of supervisor)
     

Big Data Processing

Due to outstanding developments in sensor technologies, wide spread usage of smart phones and Internet, we now have the opportunity and capability to gather huge amounts of data in different real life signal processing applications (which was not possible in the past). Efficient and effective processing of this big data can significantly improve the performance of many signal processing algorithms. However, this big data has dimensions and volumes unseen before in signal processing problems, comes in different varieties and its quality, quantity and statistics rapidly change in time and among elements. To accommodate these problems, the big data should be adaptively processed by signal processing methods since the adaptive methods (1) can process the data online, i.e., instantly, without any storage requirement, (2) can constantly adapt to the changing statistics or quality of the data, hence can be robust and prone to variations and uncertainties.

In order to effectively process the big data, we aim to develop novel adaptive signal processing algorithms specifically tuned to effectively operate on large dimensional data and construct the big data adaptive signal processing framework, as the first time, by defining new cost measures and design methodologies.

Publications on this topic include:

  1. K. Gokcesu, and S. S. Kozat, "A Universally Optimal Low-Complexity Algorithm under Exp-Concave Losses", to be submitted to IEEE Transactions on Signal Processing, 2016. (draft available with permission of supervisor)
     
  2. M. Neyshabouri, K. Gokcesu, S. Ciftci and S. S. Kozat, "A Nearly Optimal Contextual Bandit Algorithm," submitted to IEEE Transactions on Signal Processing, 2016. (pdf)
        
  3. K. Gokcesu, and S. S. Kozat, "Multimodel Density Estimation with Growing and Decaying Trees", to be submitted to IEEE Transactions on Signal Processing, 2016. (draft available with permission of supervisor)
     

Deterministic Signal Processing

Signal processing field is by far considered as the power horse and the deriving force behind all recent advances from telecommunications, Internet, green communications to smart cities. However, there exist significant practical and theoretical difficulties to apply signal processing methods to these newly emerging real life systems since these systems show high degrees of non-stationarity, unpredictability and in many cases even chaotic, as an example, consider the uncontrollable Internet. This forces us to fundamentally change our perspective on how we design signal processing methods and introduce a completely radical approach to signal processing. Deterministic signal processing is envisioned to mitigate these challenges by removing all unrealistic assumptions on real life systems and provide algorithms that in a strong sense guaranteed to work and to dramatically improve performance.

Publications on this topic include:

  1. K. Gokcesu, and S. S. Kozat, "Online Density Estimation of Nonstationary Sources Using Exponential Family of Distributions," submitted to IEEE Transactions on Neural Networks and Learning Systems, 2016. (pdf)
        
  2. K. Gokcesu, and S. S. Kozat, "Universal Estimation of Time-Varying Distributions," to appear in 42nd IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP2017). (pdf)
        
  3. K. Gokcesu, and S. S. Kozat, "Minimax Optimal Algorithms for Expert Selection under Convex and Non-convex Loss Functions", to be submitted to IEEE Transactions on Signal Processing, 2016. (draft available with permission of supervisor)
     
Anomaly Detection

Recent developments in information and detection technologies, and increasing use of intelligent mobile devices and internet have bolstered the capacity and capabilities of data acquisition systems beyond expectation. Today, many sources of information from shares on social networks to blogs, from intelligent device activities to security camera recordings are easily accessible. In our project, we will design state-aware anomaly detection and event prediction systems based on data from diverse sources acquired rapidly under unstationary conditions and quality. The problem of anomaly (or irregularity) detection and subsequent event prediction is a critical topic of national importance, particularly due to increasing border violations and cyber attacks in recent years. However, traditional techniques show insufficient performance in real-life applications.

In this project, we design novel algorithms that process text on social media, usage statistics of mobile devices and surveillance camera recordings in real time, and subsequently leverage these data for context-dependent anomaly detection and event prediction. The main aims of the project are: 1) Design of techniques that will scan text in Turkish on social media platforms, and that will extract text features for event detection and prediction; 2) Design of systems that will extract visual features from surveillance camera recordings using deep-learning methods; 3) Design of techniques that will process communication data as well as user information (such as location and time) from mobile devices to perform anomaly detection and event prediction; 4) Design of systems that will perform context-dependent processing and interpretation of data acquired at various dimensions, times and quality from written, audio and visual sources, as well as extracted features; 5) Development of a theoretical foundation for anomaly detection in big data applications.

Publications on this topic include:

  1. K. Gokcesu, and S. S. Kozat, "Online Anomaly Detection with Minimax Optimal Density Estimation," submitted to IEEE Transactions on Signal Processing, 2016. (pdf)
        
  2. I. Delibalta, K. Gokcesu, M. Simsek, L. Baruh and S. S. Kozat, "Online Anomaly Detection With Nested Trees," IEEE Signal Processing Letters, vol. 23, no. 12, pp. 1867-1871, Dec. 2016. (IEEEXplore)
        
  3. I. Delibalta, K. Gokcesu, M. Simsek, L. Baruh and S. S. Kozat, "Online Anomaly Detection With Nested Trees," to appear in 42nd IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP2017). (pdf)