Sitemap
A list of all the posts and pages found on the site. For you robots out there, there is an XML version available for digesting as well.
Pages
Posts
Optimizing Vision-Language Models for Production: A Deep Dive into Quantization and Pruning
Published:
I benchmarked popular Vision-Language Models (VLMs) like LLaVA, Qwen-VL, and PaliGemma to see how they handle quantization and pruning. The takeaway: 4-bit quantization is a no-brainer for most use cases, offering massive memory savings with minimal quality loss. Here’s what I learned about VLM size reduction.
Curriculum Learning for Mathematical Reasoning: Why Design Matters More Than You Think
Published:
I trained two language models on math problems using curriculum learning with different difficulty estimation methods. The surprising finding? The wrong curriculum can actually hurt performance. Here’s what I learned about training LLMs for mathematical reasoning.
portfolio
C++ Software Developer LTE L1 at YADRO
Published:
4G eNodeB Development using AVX and C++ for real-time telecommunications components
BAGen: Automatic production and visualization of children songs
Published:
End-to-end generation of content for children with GUI and TUI support using Python and PyTorch
Feature Selection Experimental Code
Published:
Collection of Unsupervised Feature Selection algorithms and experiments used in recent publications
Curriculum Learning Fine-Tuning for Mathematical Reasoning
Published:
Fine-tuning SmolLM2 and Phi-3 on GSM8K using complexity-based curriculum learning, achieving 1.37% accuracy improvement on difficult problems
VLM Compression: Optimization Techniques for Vision-Language Models
Published:
Implementing quantization and structural pruning for BLIP-2 and LLaVA, achieving 50% model size reduction with less than 2% accuracy loss
publications
Recognition of road markings, signs and traffic lights using deep learning models
Published in Electrotechnical University "LETI" Bachelor Thesis, 2023
Bachelor thesis focusing on traffic sign recognition using deep learning approaches for autonomous driving applications.
Review of current approaches in the area of object detection for autonomous vehicles
Published in Institute of Electronics and Information Engineers (IEIE), 2024
This survey reviews and classifies object detection methods for autonomous driving, highlights state-of-the-art approaches, and presents a taxonomy diagram to capture current research trends and challenges, particularly in ensuring reliability under adverse weather conditions.
Recommended citation: Khairulov, Timur, Sanghyuck Lee, and Jaesung Lee. "Review of current approaches in the area of object detection for autonomous vehicles." 대한전자공학회 학술대회 (2024): 2836-2839.
Download Paper
Diffusion Model-Based Generative Pipeline for Children Song Video
Published in IEEE International Conference on Consumer Electronics (ICCE), 2025
We present an approach that leverages natural language processing and diffusion-based generative models to automatically create music videos for children’s songs from lyrics, with experiments on 20 prompts showing that the Cascade SD model outperforms four alternatives across multiple evaluation metrics.
Recommended citation: Lee, Sanghyuck, Timur Khairulov, and Jaesung Lee. "Diffusion Model-Based Generative Pipeline for Children Song Video." 2025 IEEE International Conference on Consumer Electronics (ICCE). IEEE, 2025.
Download Paper
ChainImputer: A Neural Network-Based Iterative Imputation Method Using Cumulative Features
Published in MDPI Symmetry, 2025
A neural network–based missing value imputation method that incrementally builds a cumulative feature set during training, avoiding reliance on naively imputed data and achieving superior performance across 25 benchmark datasets compared to conventional methods.
Recommended citation: Seo, Wangduk, et al. "ChainImputer: A Neural Network-Based Iterative Imputation Method Using Cumulative Features." Symmetry 17.6 (2025): 869.
Download Paper
A Study on Maximizing Joint Entropy and Pattern Discrimination for Unsupervised Feature Selection
Published in Chung-Ang University Master Thesis, 2025
Master thesis exploring advanced methods for automated feature selection using joint entropy maximization and pattern discrimination techniques.
teaching
Robotarium Assembly and Launch - Teaching Assistant
Undergraduate research assistance, Electrotechnical University 'LETI', Department of Computer Science, 2022
Assisted professor in assembling and launching Russia’s largest Robotarium facility. Gained hands-on experience with multi-robot systems, hardware integration, and remote laboratory infrastructure.
Computer Vision - Teaching Assistant
Graduate course, Chung-Ang University, Department of Artificial Intelligence, 2025
Teaching Assistant for Computer Vision course at Chung-Ang University. Prepared course materials, designed examinations, and supported students in understanding advanced computer vision concepts.