Saitama, Japan
I’m a machine learning researcher and engineer with a PhD in theoretical physics, currently working at Saitama University (Uchida Lab) on sequence models and neural architectures inspired by dynamical systems. My work focuses on building and evaluating machine learning systems for time-series prediction and sequence modeling. I design new architectures, run controlled large-scale experiments, and analyze model behavior using dynamical systems diagnostics. Recent work includes: • Training GPT-style language models (7M–120M parameters) on OpenWebText (~18B tokens) • Developing transformer variants with causal convolution pathways improving model performance by ~5% • Designing reservoir/attention hybrid models achieving ~7× reduction in forecasting error on chaotic dynamical systems • Building adaptive sensing models for PDE-based diffusion systems achieving ~5× accuracy improvement Technical focus • PyTorch / deep learning systems • Transformers and sequence models • Time-series forecasting • Neural architecture experimentation and benchmarking • Dynamical systems analysis of neural networks Tools PyTorch, Python, C++, CUDA/Linux, mixed precision training, experiment pipelines and reproducible ML workflows. Open-source work includes CTM-CIFAR Bench, LAERC, Causal Convolution Transformer (CCT), and a C++ implementation of NEAT. I’m interested in roles in machine learning engineering, applied AI research, time-series modeling, and quantitative modeling.
Research and development of machine learning architectures for sequence modeling and time-series prediction at the intersection of dynamical systems and modern deep learning. • Designed hybrid reservoir–attention architectures for chaotic time-series prediction, achieving ~7× reduction in forecasting error (NRMSE) across benchmark dynamical systems. • Built PyTorch training and benchmarking pipelines for neural architecture experiments, training models from 7M to 120M parameters across 100+ controlled runs. • Trained GPT-style language models on OpenWebText (~9B tokens, ~18B processed tokens) and implemented training optimizations including RoPE positional embeddings. • Developed adaptive sensing models for PDE-based diffusion systems using 128 sensors/observation points, achieving ~5× accuracy improvement over fixed sensing strategies. • Implemented and benchmarked transformer variants including causal convolution front-ends, improving model performance by ~5% for a 120M parameter language model. • Built reproducible ML experimentation frameworks (configs, logging, checkpoints, benchmark scripts) enabling controlled ablations and architecture comparisons. • Maintained open-source ML research repositories and collaborated with lab members and students on model development and experimentation.
Research on machine learning approaches for time-series prediction using nonlinear dynamical systems and reservoir computing. • Developed delay-based reservoir computing models for forecasting chaotic time-series and signal prediction tasks. • Analyzed stability, memory capacity, and information-processing properties of dynamical systems used as machine learning models. • Built analytical and numerical tools to quantify memory capacity and computational limits of time-delay reservoir systems. • Designed hybrid modeling approaches combining physical dynamical systems with data-driven machine learning for improved prediction performance. • Implemented reproducible experimentation pipelines in Python/C++ (simulation, model training/evaluation, hyperparameter sweeps) using Linux-based research workflows. • Published results in peer-reviewed journals (Physical Review Applied, IEEE TNNLS) and presented work in academic seminars and research conferences. • Supported teaching and student supervision during the doctoral program.
Supported research projects in nonlinear dynamics and machine learning, including simulation, modeling, and evaluation workflows implemented in Python/C++. • Teaching assistant for courses in Nonlinear Dynamics and Statistical Physics, leading tutorials and problem-solving sessions. • Mentored Bachelor’s and Master’s students on research projects and theses, providing guidance on modeling, analysis, and scientific programming. • Prepared instructional materials, graded assignments/exams, and supported course organization with faculty and teaching staff.
Conducted visiting research on machine learning approaches for forecasting complex non-equilibrium dynamical systems. • Built time-series modeling pipelines in Python, including data preprocessing, training/evaluation loops, and baseline comparisons. • Analyzed model behavior and robustness across different architectures and training setups, identifying stability and sensitivity effects. • Collaborated with researchers to translate domain-specific physics problems into machine learning forecasting tasks and evaluation metrics. • Contributed to research reports, presentations, and manuscript preparation.
Worked on computational simulations of pathogen spread using numerical modeling and C++. • Implemented simulation pipelines including parameter sweeps and experiment setup. • Performed basic statistical analyses and validation checks on simulation outputs. • Collaborated with researchers to translate modeling assumptions into implementable algorithms.
Supported undergraduate laboratory courses in physics, assisting students with experimental setup, measurements, and data analysis. • Guided students through calculations, error analysis, and documentation of experimental results. • Reviewed lab reports and provided feedback on methodology and clarity.