Kwangryeol Park

Summary

Highly accomplished AI/ML Engineer with a proven track record in developing robust and memory-efficient solutions for Time Series Analysis and Embedded AI systems. My research focuses on enhancing the reliability and performance of Deep Learning models, evidenced by four publications accepted at top-tier conferences (NeurIPS, ICLR, CIKM, AAAI) during my Ph.D. program. Expertise includes Time Series Forecasting, Self-Supervised Learning, Model Optimization, and practical C/C++ deployment on resource-constrained hardware.

Publications

• AliO: Output Alignment Matters in Long-Term Time Series Forecasing NeurIPS 2025 (Poster) Authors: Kwangryeol Park, Jaeho Kim, Seulki Lee*
  • AliO YouTube (Generated by NotebookLM): Link
  • AliO NeurIPS Poster: Link
  • Problem Identification: Solved the critical Output Alignment Problem in Long-Term Time Series Forecasting (LTSF), where predictions lack consistency across lagged inputs.
  • Methodology: Developed AliO (Align Outputs), a novel loss function that ensures prediction consistency in both time and frequency domains.
  • Key Metric: Proposed the Time Alignment Metric (TAM) to quantitatively measure prediction consistency and data-model robustness, a factor previously unaddressed by MSE.
  • Performance Highlights: Achieved substantial performance gains across LTSF models:
    • MSE Improvement: Up to 27.5%
    • \text{TAM} Improvement: Up to 58.2%
  
• PPT: Patch Order Do Matters In Time Series Pretext Task ICLR 2025 (Poster) Authors: Jaeho Kim†, Kwangryeol Park†, Sukmin Yun, Seulki Lee*
  • Problem Identification: Identified that existing patch-based models often fail to capture essential time- and channel-wise patch order dependencies in time series data.
  • Methodology: Introduced PPT (Patch order-aware Pretext Task), a novel self-supervised learning task that exploits intrinsic sequential order information through channel-wise patch permutations.
  • Key Components: Proposed two core learning methods: Patch Order Consistency Learning (quantifies correctness) and Contrastive Learning (distinguishes permutation strength).
  • Key Metric: Proposed ACF-COS (Autocorrelation Function with Cosine Similarity), a metric to quantify the "orderness" of a time series dataset.
  • Performance Highlights: Enhanced model performance over mask-based learning by:
    • Supervised Task: Improving up to 7% accuracy for cardiogram tasks.
    • Self-Supervised Task: Outperforming mask-based learning by 5% in human activity recognition.
  
• Smart ECU: Scalable On-Vehicle Deployment of Drivetrain Fault Classification Systems for Commercial Electric Vehicles CIKM 2025 (Oral) Authors: Jaeho Kim†, Kwangryeol Park†, Kyu Hwan Lee†, Jeongmin Oh, Dongjin Park, Hyunseok Oh, Youngrock Chung, Kyung-Woo Lee, Dae-Un Sung, Seulki Lee*
  • SmartECU Presentation Video: Link
  • Goal: Presented Smart ECU, the first on-vehicle drivetrain fault classification solution for motor-reducers on commercial EVs, designed for scalable mass production.
  • Data: Collected authentic fault and normal vibration data from seven different EV models (e.g., Hyundai IONIQ 5, KIA EV6) and over 19 drivetrains under diverse real-world driving conditions.
  • Deployment Challenge: Addressed the challenge of deploying on extremely resource-constrained ECU environments, facilitating Progonstic and Health Management (PHM) solutions on commercial EVs.
  • Efficiency & Results: Achieved strong efficiency and performance:
    • Model Size: Required only 3.28 MB of memory, a 99.24% reduction compared to prior cloud-based methods.
    • Generalizability: Demonstrated strong generalization performance in both intra-car (same vehicle type) and inter-car (different vehicle types) evaluations.
  
• SMMF: Square-Matricized Momentum Factorization for Memory-Efficient Optimization AAAI 2025 (Poster) Authors: Kwangryeol Park, Seulki Lee*
  • SMMF YouTube: Link
  • Problem Solved: Addressed the high memory consumption of widely-used adaptive optimizers (e.g., Adam) in training large-scale Deep Learning models.
  • Methodology: Proposed SMMF (Square-Matricized Momentum Factorization), a novel method utilizing a square-matricization and one-time single matrix factorization scheme.
  • Key Advantage: Enables flexible and efficient factorization of momentum tensors regardless of their arbitrary rank (shape), making it applicable to all layers (Linear, CNN, Transformer).
  • Theoretical Foundation: Provided a regret bound analysis to theoretically confirm that SMMF converges comparably to non-memory-efficient adaptive optimizers.
  • Performance Highlights: Achieved state-of-the-art memory efficiency:
    • Memory Reduction: Reduced the memory requirement by up to 96% compared to standard adaptive optimizers.
    • Competitive Performance: Maintained competitive optimization capability across various benchmarks (e.g., ResNet50 on ImageNet).
  

Professional Activities

(2026) - Reviewer, ICLR 2027 Main Track
• (2026) - Reviewer, KDD 2027 Data & Benchmarks Track
• (2026) - Reviewer, NeurIPS 2026 Main Track
• (2026) - Reviewer, FMTS 2026 (NeurIPS 2026 Workshop)
• (April 2026~) - Research Collaboration 3, Hyundai Motor Company [Ongoing]
• (April 2026~) - Research Collaboration 2, Hyundai Motor Company [Ongoing]
• (December 1-7, 2025) - Poster Presenter, NeurIPS 2025 (Conference on Neural Information Processing Systems), San Diego, USA (AliO)
• (November 25, 2025) - [Project] Conference Deadline Tracker, Local web server project , [Project Completed]
• (November 20, 2025) - Poster Presenter, JKAIA 2025 (Joint Korean Artificial Intelligence Association)
• (November 11, 2025) - Session Chair, CIKM 2025 (ACM International Conference on Information and Knowledge Management)
• (November 10-14, 2025) - Paper Presenter (Oral), CIKM 2025 (Smart ECU)
• (October 2025) - Research Collaboration 1, Hyundai Motor Company [Completed]
• (April 24-28, 2025) - Poster Presenter, ICLR 2025 (International Conference on Learning Representations)
• (February 25 - March 4, 2025) - Poster Presenter, AAAI 2025 (The 39th Annual AAAI Conference on Artificial Intelligence)

Technical Skills & Expertise

Research Interests

Education

Ulsan National Institute of Science and Technology (UNIST)
AI Graduate School (Embedded AI Lab by Seulki Lee)
Ph.D. Candidate (Integrated M.S. & Ph.D. Program, 7th Semester) (2022 - Present)

Sangmyung University
B.S. in Electronic Engineering (Embedded Systems Track)
Graduated as Valedictorian (Early Graduation) (~2022)