Machine learning engineer & PhD researcher.
I build computer-vision systems, data pipelines, and software for research. I study how vision models relate to human brain activity.
A horror film recommender with mood search, an explorable map of 19,000 films and an illustrated history of the genre.
Designed and built it solo: a collaborative-filtering recommender, text-embedding search, the film atlas, an AI-assisted theme analysis, accounts and a self-hosted deployment.
Naturalness scores for all 2.1 billion images in ReLAION-2B, plus CLIP embeddings for the 500 million natural photos.
Built the GPU pipelines for classification, quality filtering and embedding extraction at billion-image scale. Downloaded over 30,000 times on Hugging Face.
A community challenge testing whether findings in visual neuroscience replicate on LAION-fMRI.
Designed the data splits the challenge runs on, replicated a 118-model brain-alignment benchmark for the consortium paper, and built the website and data services.
A deeply sampled 7T fMRI dataset: 5 people, 60 hours each, 25,000 natural images.
Designed the stimulus set, optimizing it for diversity to 2.5× the effective dimensionality of NSD and THINGS. Also built its Python package, website, 3D brain explorer and data access on AWS.
An open-source library for extracting and comparing features from vision models.
Third-largest contributor: added low-memory extraction, HDF5 dataset support and model-integration tests.
A lightweight dashboard for Slurm jobs, logs and resource use on HPC clusters.
An interactive QA and review tool for fMRI data snapshots.
Cross-validated MANOVA for multivariate fMRI pattern analysis.
Customisable searchlight analysis for Python.
Code and experiments for the CCN 2025 paper on sampling natural images.
A LAION-fMRI replication of a 118-model brain-alignment study.
At the Max Planck Institute and University of Giessen, I study how visual representations in models relate to the human brain.
Mahner*, Roth* et al. · Shared first author
We compared the object similarity structure of 162 vision models to distinguish universal from model-specific dimensions, and related that structure to image properties, neural activity, and human judgments.
Roth & Hebart · Oral
How does the choice of images affect what we can learn about the visual brain? We compared image-sampling strategies using simulations and fMRI data to study dataset coverage and model generalization.
CMBB Replication Award
Roth*, Duan* et al. · Shared first author
A practical guide to research software, covering project organization, testing, documentation, version control, and collaboration.
Badwal, Bergmann, Roth et al. · J Neuroscience
Roth, Miyawaki & Hebart · VSS poster
Roth et al. · MICCAI BrainLes 2021 proceedings
Roth, Seeliger, Schmid & Hebart · Cosyne poster
Max Planck Institute for Human Cognitive and Brain Sciences & University of Giessen
Independent research on vision models and human brain activity, with GPU-based data pipelines, fMRI experiments, and public research software.
ScaDS.AI Dresden/Leipzig
Deep-learning models for brain-tumor segmentation and predictive models for clinical data, including mortality prediction and uncertainty estimation.
CHECK24, Travel Vertical
Designed and deployed a production computer-vision service for 20 million hotel images. Improved recommendation-data pipelines, reducing update time by over 95%.
Kimetric UG
Delivered Django websites for Hebart Lab and the THINGS Initiative, from client requirements to deployment, Linux hosting, and maintenance.
Webdata Solutions GmbH (now Vistex)
Developed machine-learning models for product matching and image classification in e-commerce.
Mercateo Services GmbH & Co. KG
Data integration, reporting, and testing for an AWS-based data warehouse using Apache NiFi and Pentaho.
University of Giessen
Research focus: machine learning and computational neuroscience.
Leipzig University · Grade 1.2
Master’s thesis: generative deep learning for visual processing in the human brain (grade 1.1).
Leipzig University · Grade 1.5
What scares us changes. An illustrated analysis of 19,302 horror films, from atomic mutants to masked killers to ghosts, on Horrorverse.
A passage from Art and Fear and what it says about how we see other people's work versus our own.
A little less formal.
Here's what life looks like away from the screen: climbing, making music, reading, and having friends over for dinner.
My love for moving my body started with joining a breakdance crew at 17. These days I mostly boulder, lift weights, and do yoga. A day in the mountains with some climbing is hard to beat.
A few years ago, I asked some friends if they wanted to start a band. Everyone said yes. We mostly play pop punk covers at birthday parties, but the dream of writing our own songs is still alive.
I used to be an incredibly picky eater. Now I love throwing whole food ingredients together and having friends over for dinner. Hosting is a blast, even when it gets surprisingly stressful.
Meditation and I have had an on-off relationship for over a decade. A ten-day silent retreat in 2024 changed a lot for me. I walked this forest path probably a hundred times and kept discovering something new.
Books, music, and movies.
I listen to a bit of everything. Spotify probably knows my taste better than I do.
My wife and I watch horror movies with friends. I built a little recommender to help us choose.
The Obstacle Is the Way
Stolen Focus
The Wim Hof Method
Atomic Habits
Indistractable
Waking Up
Built to Move
Kitchen Confidential
The Comfort Crisis
The Shallows
What scares us changes. An illustrated analysis of 19,302 horror films, from atomic mutants to masked killers to ghosts, on Horrorverse.
A passage from Art and Fear and what it says about how we see other people's work versus our own.
For research, collaborations, or a good conversation.