Timur Oner

Master’s holding Data Scientist with a solid mathematical background. Currently working on Multimodal LLMs and their performance on cognitive tasks. Focuses on bridging cutting edge research in multimodal AI, cognitive science and industry applications.

My CV

Educational Background

Currently 2nd year Data Science Master’s Student at University of Padua. Completed a theoretically comprehensive courses like Vision and Cognitive Systems and Statistical Methods. Currently working on improving and understanding how Multimodal LLMs perform downstream tasks related to human cognition like counting with Prof. Alberto Testolin and Dr. Sina Shafizadeh.

With my colleagues from Cognitive Computational Neuroscience Lab.

What I do?

Programming Expertise

I have worked with Python and its main data science related libraries like PyTorch and TensorFlow. Most importantly I am fluent in modern programming practices and can learn/adapt to new tools/frameworks very fast. In addition I describe myself as a creative problem solver.

Solid Teoretical Background

I am in the stage of completing of Master’s Degree in Data Science@University of Padova. The curriculum included a wide range of theoretical and practical material and gave me a very solid academic background on artificial intelligence and data science.

Focus on Multimodal AI

I specialize in AI systems that work with different modalities. I believe innovations in the multimodal AI will have ground-breaking industrial applications and will be an important step towards reaching artificial general intelligence (AGI).

Taken from: https://vocal.media/futurism/the-agi-testing-crisis-why-current-ai-evaluation-fails-to-detect-cognition

My Professional Experience

I am currently doing a research internship at Cognitive Computational Neuroscience Lab@University of Padova. Our work is focused on understanding and improving performance of Multimodal Large Language Models on tasks as counting and mathematical reasoning. My role involves diving deep into nuts and bolts of MMLLMs (particularly LLaVA) and performing specialized finetuning procedures to improve counting performance on complex real-world datasets. This work has a lot of real-world application potential from counting number of products from surveillance recording to development of multi-modal chatbots that can reason based the image provided by the user better.

From November 2022 until June 2023 I worked in FinTech, building and deploying a credit risk classifier based on gradient boosting decision trees. As a team we reached a validation accuracy of 78 percent. This experience was very valuable as it taught me how to make sense of noisy and incomplete data that constitutes the majority of the real world data.

In addition, I have developed a simple end-to-end sentiment analyser web-application to demonstrate that light-weight models can still achieve a good performance in an age when LLMs dominate tasks such as sentiment analysis and text classification. You can find it HERE.