RESUME
Education
Ulsan National Institute of Science and Technology (UNIST)
Bachelor of Science in Industrial Engineering (Senior year, Second Semester)
AI Relevent Courses Taken: Computer Visioin, Principle of Deep Learning , Applied Machine Learning , Data Mining, Data Science Programming, Artificial Intelligence Programming I & II, Social Network Analysis
Research and Work Experiences
- Working on improving the accuracy of pedestrian detection by combining RGB and Thermal Images using GANs
- Implemented WGAN-GP for enhancing input for YOLOV3.
- Increased AP score by 1 of base YoloV3 model on KAIST multispectral pedestrian detection dataset.
- Developed database server for the experiment using MongoDB and node.js
- Added distance measurement methods using camera of phone for the existing Android App.
- Explored augmentation techniques to improve accuracy of 2D Human Pose Estimation model HRNet. (github link)
- Boosted 0.1% accuracy of baseline HRNet on COCO dataset by using the idea explained in the paper, “How Robust is 3D Human Pose Estimation to Occlusion?”.
- Programmed official demo of HRNet for human pose estimation using python. Code link.
- Reduced computational cost of news article recommendation system by parallelizing bandit algorithms(LinUCB and Thompson sampling) using python (numba, numpy). (github link).
- Acquired funding for the project from UNIST Innovative Education Center.
- Implemented the parallelized version of LinUCB.
- Achieved 3.48 times and 2 times acceleration in parallelized version of LinUCB and Thompson Sampling compared to their sequential counter part. (Experimented on Yahoo R6A dataset ).
- Presented the poster of the outcome of the project in a conference organized by UNIST.
Research Assistant
Vision and Learning Lab
UNIST, South Korea
Jan 23 - Present
Android Developer
Perception, Action & Learning Lab
UNIST, South Korea
Aug 22 - Dec 22
Research Assistant
Machine Learning & Vision Lab
UNIST, South Korea
Sep 21 - Feb 22
Research Assistant
Statistical Decision Making Lab
UNIST, South Korea
Dec 20 - Aug 21
PUBLICATIONS AND POSTERS
- K. Sayem1, M. Chowdhury1, E. Ismayilzada1, and G. Kim*, “GBOSE: Generalized Bandit Orthogonalized Semiparametric Estimation”. (Under Review), 2022.
- K. Sayem, E. Ismayilzada, M.T. Chowdhury and G. Kim, "Development of news article recommendation system via reinforcement learning". Poster presented at: UNIST, 2021 Dec 3, Ulsan, Republic of Korea.
Projects
- Portfolio construction of stocks using social network analysis.
- Course: Social Network Analysis, UNIST
- Duration: 2021.09 - 2021.12
- Keywords: Social Network, Centrality Measures, Networkx
- Tools: Python, Networkx, Pandas, Numpy, Matplotlib
- Abstract: In this work, we form a series of optimized portfolios of stocks out of companies listed in the S&P500 index fund, using social network theories of centrality. Centrality measures for each stock in the network is used to decide the proportion of portfolio budget to be allocated to different stocks. And then, the performance of the optimized portfolios is tested against the overall performance of the S&P500 index funds at the specified duration. The objective is to analyze how centrality property of a stock impacts it's price changes within a particular duration.
- Overview of CNN based Monocular depth estimation methods.
- Course: Principle of Deep Learning, UNIST
- Duration: 2022.04 - 2022.06
- Keywords: CNN, Monocular Depth Estimation, Depth Map
- Abstract: In this work, we present an overview of CNN based monocular depth estimation methods. We discuss the different architectures of CNN based depth estimation methods and their performance on different datasets. We also discuss the different loss functions used in these methods and their impact on the performance of the depth estimation methods. Finally, we discuss the different applications of depth estimation methods and their future scope.
SKILLS
- Programming Languages: Python, C++
- Frameworks: PyTorch, Tensorflow, Scikit-learn, Pandas, Numpy, Matplotlib, Numba
