Several research opportunities and positions are available. If you are an undergraduate student or a prospective graduate student with an interest or a background in machine learning and curious about exploring the world of real-life research in this area, please send Prof. Kozat an e-mail stating your interest and CV. email: kozat@ee.bilkent.edu.tr

Benefits include highly competitive Salary, extensive Computing Resources, Funding for both international and national conferences and other benefits. For more information on the current and past research projects, current research group members (and their publications), alumni (and their current positions in US), other technical activities, open positions and announcements, please directly check this website regularly.

ACTIVE RESEARCH TOPICS

Prediction in Renewable Energy and e-commerce domains

We work on data where the data arrives in a sequential manner, highly correlated in time and space, rarely clean and highly non-stationary, e.g., the real life. That is the reason why high tech companies such as Google and Amazon still invest billions of dollars in this field. The promise of deep learning is not achieved here yet since there is not enough data to learn both feature extraction and algorithm building jointly at the same time.

Our applications include energy production prediction/regression/forecasting, residential and industrial natural gas consumption, financial markets and e-commerce. As an example, we predict how much a wind farm produces energy in the next 36 hours, how much natural gas will be consumed in the next 6 months, how many and how much a good, e.g., milk, dress, shoes, will be sold/shipped in the next 7 days, what will be the price of the bitcoin in the next month. We compete with the largest and the most successful international research groups both in universities and in high tech research labs. Along with many publications, our results will be patented.

ChatBots and NLP algorithms with cognitive and behavioral capabilities using fine tuned LLMs and Big Data

With the release of ChatGPT, we observe the recent advances in generative AI technologies. However, building exclusive ChatBots with cognitive, i.e., with human like understanding, and behavirol, i.e., with personality and domain knowlegde, capabilities is still an open problem. For this purpose, we construct specialized LLMs that are fine tuned with our extensive Big Data knowledge databases and computing resources. Our goal is to go one step further than the-state-of-the-art and seek to achieve the next generation ChatBots with adaptation capabilities to new domains such banking, e-commerce and IT. Unlike the current ChatBots that use LLMs trained over old or stale data, our algorithms are constantly updated with the most recent and reliable data.

To this end, our ChatBots are integrated to our Big Data collection platform, where the most recent reliable data are continuously collected from diverse resources including newspapers, online encyclopedias such as Wikipedia and/or Vikipedia, blogs, social media. However, this high volume of data should be rectified and fact-checked for LLM training and ChatBot integration. For this purpose, we construct advance NLP algorithms with specialized properties. Also, to provide human-like cognitive and behavirol abilities, our ChatBot ensembles information from different and diverse sources in a coherent manner with data clustering, advance retrieval and indexing abilities.

Big Data Processing and Anomaly Detection

We work on anomaly detection on big data. Here, the anomaly is defined as any non-regular pattern that is different than which is normal. The anomaly can be 1) just a point anomaly, where an unexpected increase in the temperature; 2) a pattern change in the course of a normal chain of events, e.g., a change in the behavior of a computer admin; 3) a collective contributions of small but a large number of sources that can cause big changes on the total behavior, e.g., a sudden ranking up of certain hashtags on Twitter coming from multiple sources. In all these cases, enormous amounts of data should be processed so that the normal behaviors are learned, say for an ordinary computer programmer or a group of Twitter users. Here, we develope machine learning algorithms that can process huge amounts of data on parallel, extract all the relevant statistics and relationships, model regular behavior(s) of a single or a group of entities, follow their ordinarily changing behavior(s) and alert when an unexpected behavior is present. In a summary, it is finding a needle in a haystack.