About

I’m Rinor Cakaj, a PhD student at the University of Stuttgart in the Institute of Signal Processing and System Theory, where I am advised by Prof. Dr.-Eng. Bin Yang. I’m working on my PhD in cooperation with Robert Bosch GmbH, supervised by Dr. rer. nat. Jens Mehnert. My research focuses on improving the performance of Convolutional Neural Networks (CNNs) for computer vision tasks.

Feel free to reach out to me via email. You can also check out my CV.

Research Interests

My current research focuses on developing new regularization techniques and innovative CNN architectures to improve their efficiency and performance. As part of my PhD, I have published five papers and authored five patents related to these novel methods, with one already published and four currently under review. Additionally, I’ve gained a deep understanding of other deep learning architectures, including Transformers (such as Vision Transformers (ViT) and Large Language Models (LLMs)).

Background

Before starting my PhD, I completed a master’s degree in Mathematics interdisciplinary Informatics and a bachelor’s degree in Business Mathematics at the Technical University of Darmstadt. I also gained work experience as a working student in CIO Consulting - Digital Strategy at KPMG and as a mathematician in a multi family office at Segura & Jesberger GmbH.

Selected Publication

• CNN Mixture-of-Depths
  Rinor Cakaj, Jens Mehnert, Bin Yang
  To appear in the Asian Conference on Computer Vision (ACCV) 2024, Hanoi, Vietnam.
  [ArXiv]

Personal Projects

Stock Index Trend Prediction Tool

In addition to my academic research, I’ve been working on a personal project to predict stock index trends using deep learning. I developed an end-to-end pipeline, starting with automatic data collection through web scraping and adding a plausibility check for data integrity. The data is processed using rolling statistics and data augmentation techniques to enhance robustness, followed by splitting it into training, validation, and test sets.

The deep learning model uses multi-scale convolutions to capture diverse patterns and BiLSTM layers to handle temporal dependencies. Implemented in Python and PyTorch, this project allowed me to gain practical experience with various tools and libraries, categorized as follows:

• Development Environment: JupyterLab, VS Code, Anaconda
• Data Processing and Visualization: Pandas, NumPy, Plotly, Matplotlib, Scipy
• Automation and Testing: Selenium, Pytest, Optuna
• Model Tracking and Deployment: Weights & Biases, TensorBoard, Docker, Git
• Documentation: Sphinx

Electronic Access Control System

For my local swimming club, I developed an electronic access control system using a Raspberry Pi 4 and Python. The system features a GUI, an RFID reader with cards, and a touch display, providing an efficient way for members to access the club.

Early App Development

During my school years, I created simple Android and iOS apps, such as a calculator and a game called “Bouncing Ball,” where the goal was to keep the ball in the air. The apps were implemented in Java for Android and Swift for iOS. These early projects sparked my interest in software development and problem-solving.