Marios Koulakis

Machine Learning Researcher · Mathematician

Marios Koulakis

Portrait of Marios Koulakis

I am a machine learning researcher with a background in mathematics, computer vision, and applied AI. My work focuses on the geometry and structure of data, learned representations, and data-generating processes, with particular interest in scientific machine learning, manifold learning, generalization, and model reliability.

I aim to develop measurable, mathematically grounded methods that connect theory with empirical analysis, open-source software, and real-world applications. I am especially interested in collaborations involving scientific data, simulations, imaging, and structured transformations.

Scientific Machine Learning Data Geometry Representation Learning Manifold Learning Dimensionality Reduction Reliable AI

CAREER

2022–Present

Senior Machine Learning Scientist · DeepHealth

Working on deep learning for lung disease detection, including model development, data selection, and reliability- and regulation-oriented workflows.

2021–2023

Postdoctoral Researcher · Karlsruhe Institute of Technology

Research in computer vision and machine learning, including h-NNE, supervision, and construction-related vision projects.

2019–2021

Data Scientist / Senior Data Scientist · Royal Schiphol Group

Built computer vision and activity-recognition solutions for aircraft handling processes and operational analytics.

2016–2018

Software Engineer / Data Scientist & Engineer · Nulogy

Worked on predictive modeling, optimization, data infrastructure, and AWS-based systems in manufacturing and supply-chain settings.

2015–2016

Software Engineer · ASD Personalinformationssysteme

Full-stack development for scheduling and time-tracking software in an agile environment.

2011–2015

PhD Researcher · University of Münster

Research in mathematical logic and set theory as a Marie Skłodowska-Curie fellow.

RESEARCH

Geometry, dimensionality reduction, and structure in data

The Data Manifold under the Microscope

ICML 2026

Benchmarking datasets and geometric tools for studying simple but real-life-like data manifolds, with a focus on analysis and experimentation.

Paper Code

Hierarchical Nearest Neighbor Graph Embedding for Efficient Dimensionality Reduction

CVPR 2022

A hierarchical dimensionality reduction method designed for efficient and structured visual exploration of high-dimensional data.

Paper Code

Applied AI, anomaly detection, and industrial computer vision

Position: Quo Vadis, Unsupervised Time Series Anomaly Detection?

ICML 2024

A critical perspective on evaluation practices and benchmarking in time series anomaly detection.

Paper Code

Domain-independent Detection of Known Anomalies

CVPRW VAND 2.0 2024

Approaches for anomaly detection across previously unseen objects in industrial inspection settings.

Paper Code

Bildbasierte Baufortschrittsüberwachung

Book chapter 2024

Image-based construction progress monitoring in a practical engineering context.

Chapter

Reliable perception and representation learning

Is My Driver Observation Model Overconfident? Input-Guided Calibration Networks for Reliable and Interpretable Confidence Estimates

IEEE T-ITS 2022

Work on calibration and confidence estimation for driver observation models under real-world deployment shifts.

Paper

Affect-DML: Context-Aware One-Shot Recognition of Human Affect using Deep Metric Learning

FG 2021

Context-aware one-shot affect recognition using deep metric learning.

Paper Code

Mathematical logic

Coding into Inner Models at the Level of Strong Cardinals

BSL Thesis Abstract 2018 PhD Thesis 2015

PhD thesis, accompanied by a thesis abstract in the Bulletin of Symbolic Logic.

Abstract Thesis

PROJECTS

Manifold Microscope

Code

A benchmarking and experimentation toolkit for studying geometric properties of synthetic and real-life-like data manifolds.

  • Benchmark datasets for manifold analysis
  • Geometry-oriented evaluation tools
  • Public codebase for reproducing the framework and experiments

CVPR 2025 tutorial

Webpage

Co-organized tutorial on dimensionality reduction, clustering, and structure in data for computer vision and machine learning.

  • Conceptual overview and practical materials
  • Bridges theory, visualization, and applications
  • Part of a broader interest in structure-aware ML

Awesome Dimensionality reduction and Clustering

Dimensionality Reduction Clustering

Curated reading and implementation lists for two closely related toolboxes: dimensionality reduction and clustering.

  • Surveys and landmark methods with paper and code links
  • Frameworks, metrics, and datasets for practical comparison
  • Clustering resources also include theory and cluster-number estimation

QuoVadisTAD

Code

An open-source toolkit accompanying our ICML 2024 position paper, with baselines and evaluation tools for time-series anomaly detection.

  • Supports reproducible benchmarking and consistent evaluation
  • Includes simple and neural-network baselines
  • Provides quick-start notebooks for custom datasets

h-NNE

Code

A hierarchical dimensionality reduction method related in spirit to t-SNE and UMAP, but designed for speed, simplicity, and structured coarse-to-fine exploration.

  • Built around hierarchical structure in the data
  • Supports interpretable zoom-like exploration
  • Useful for visualization and data analysis