Online Learning - Multi-Armed Bandits
I have been conducting research on multi-armed bandits since 2015 under the supervision of Cem Tekin. Multi-armed bandit problem is a framework used to model and analyze online machine laerning problems, where the algorithms are trained on the fly, rather than on an offline dataset. These problems capture the explore-exploit dilemma, where the learner tries to strike a balance between benefiting from his imperfect knowledge (exploiting) and discovering unobserved actions (exploring). We have been analyzing different problem settings that capture various challenges such as combinatorial, active learning and multi-objective problems. We propose algorithms that solve these problems and present theoretical guarantees on te ir performance as well as presenting numerical results on their demonstrative applications.
Multi-Station Detection
During my MS, I have worked on the basic limits of multi-station detection problem. This is the task of identifying the subset of active stations from a large population. This problem is encountered in many large scale detection setting and finds applications in stock control, supply chain management and Internet of Things applications. We formulated an information-theoretic framework for evaluating detection algorithms and presented our analyses of a novel algorithm and state-of-the-art.
Applied Machine Learning - Computational Musicology
In addition to my aforementioned theoretical work, I am also interested in practical methods for applying machine learning on real-world problems from various domains. I have worked on applications on natural language processing, image classification, multi-user communications, online advertising and audio classification.
I had the opportunity to collaborate with Music Technology Group of Universitat Pompeu Fabra, Barcelona. I worked on the EU funded CompMusic project under the supervision of Xavier Serra. We have developed classification methods to classify non-Western music traditions, implemented a scientific computation library called MORTY and curated a classification dataset. Furthermore I was exposed to data curation and annotation process, as well as semantic data representation methods like knowledge-bases.