Research

My research focuses on mathematical approaches to machine learning with applications in computer vision and autoencoders. I’m particularly interested in scaling laws for contrastive learning and efficient optimization algorithms for accelerator hardware.

Research Areas

Research Overview

Primary Research Focus

My work addresses fundamental questions in machine learning and computational optimization by developing mathematically principled methods to understand scaling behavior and solve computationally intensive problems. The interdisciplinary nature of my research combines optimization theory, information theory, and high-performance computing to tackle complex challenges in computer vision and self-supervised learning.

Research Philosophy

I believe in open, reproducible science that advances both theoretical understanding and practical applications. My approach emphasizes:

Mathematical Rigor: Employing robust theoretical foundations and computational methods
Reproducibility: Ensuring all research is transparent with open-source implementations
Scalability: Developing solutions that work at unprecedented scales
Practical Impact: Bridging theory and real-world applications
Open Science: Contributing to the broader research community through shared resources

Current Research Themes

🧠 Scaling Laws in Contrastive Learning

Understanding how contrastive autoencoders scale with dataset size, model parameters, and computational budget.

Key Questions:

How do contrastive autoencoders scale with various system parameters?
What are the fundamental limits of representation learning through contrastive objectives?
How do architectural choices affect scaling efficiency in self-supervised learning?

Methodological Approach:

Theoretical Analysis: Mathematical characterization of power-law relationships
Information Theory: Understanding representation capacity and efficiency bounds
Empirical Validation: Large-scale experiments across multiple domains

Recent Work:

PICTAR framework for adversarial robustness through contrastive training
8-bit contrastive learning for efficient representation
Scaling studies on large-scale vision datasets

⚡ Mathematical Optimization on Accelerator Hardware

Efficient algorithms for NP-hard problems in computer vision and scientific computing applications.

Motivation: Traditional computer vision pipelines require solving assignment problems millions of times, creating computational bottlenecks that limit real-time applications and large-scale processing.

Technical Contributions:

Parallel Assignment Algorithms: Novel GPU-accelerated approaches to the Hungarian algorithm
HPC Optimization: High-performance computing solutions for scientific workflows (6DIMCOCO, JERICHO)
Memory-Efficient Implementations: Algorithms designed for accelerator hardware constraints

Impact:

Significant speedups in multi-object tracking and correspondence problems
Open-source HPC tools adopted by research groups
Integration into scientific computing workflows for plasma physics and molecular dynamics

🤖 Machine Learning for Scientific Computing

Applying modern ML techniques to solve complex problems in physics and engineering.

Innovation: This research direction leverages machine learning to accelerate traditional scientific computing methods, particularly in plasma physics simulations and molecular dynamics.

Current Projects:

6DIMCOCO: 6-dimensional phase space analysis for plasma physics
JERICHO: Large-scale molecular dynamics with GPU acceleration
ML-SLURM integration for intelligent HPC resource management

Applications:

Plasma physics simulation and analysis
Molecular dynamics acceleration
Scientific workflow optimization

Research Impact

Software Contributions

6DIMCOCO: Advanced plasma physics analysis toolkit
PGDVisualisation: Interactive visualization for adversarial attack analysis
ML-SLURM-Template: HPC workflow optimization framework
JERICHO: High-performance molecular dynamics simulator

Publications

Conference Papers: ICLR, NLDB contributions on contrastive learning and adversarial robustness
Preprints: Active research on scaling laws and optimization algorithms
Technical Reports: Documentation of open-source research software

Community Impact

Open Source Projects: Multiple research software packages with active user communities
Reproducible Research: All publications accompanied by open-source implementations
HPC Education: Templates and frameworks for scientific computing workflows

Current Research Infrastructure

Computational Resources

High-Performance Computing: Access to GPU clusters for large-scale experiments
Cloud Computing: Distributed training for scaling studies
Specialized Hardware: GPU acceleration for optimization algorithms
SLURM Integration: Automated workflow management for scientific computing

Software Development

Version Control: Git-based collaborative development
Documentation: Comprehensive documentation for all research software
Testing: Automated testing and continuous integration
Distribution: PyPI and Conda packages for easy installation

Future Directions

Short-term Goals (1-2 years)

Complete scaling laws analysis for contrastive autoencoders
Publish GPU-accelerated assignment algorithm research
Expand 6DIMCOCO capabilities for advanced plasma physics analysis
Develop new visualization tools for adversarial robustness analysis

Medium-term Vision (3-5 years)

Establish comprehensive framework for scaling laws in self-supervised learning
Create next-generation optimization algorithms for emerging accelerator hardware
Bridge machine learning and scientific computing communities
Develop production-ready tools for industrial applications

Long-term Impact (5-10 years)

Transform understanding of scaling behavior in machine learning systems
Enable real-time processing of large-scale computer vision applications
Accelerate scientific discovery through optimized computational tools
Train next generation of researchers in ML-HPC intersection

For specific current projects, see Current Projects. For detailed research interests, see Research Interests.

Last updated: Sep 16, 2025