AI Researcher · Generative Models · Foundation Models
David Zimmerer
building robust generative models.
benchmarking the unknown.
building vision language models.
detecting anomalies that matter in medical imaging.
Senior AI researcher at the German Cancer Research Center (DKFZ) working on generative models, unsupervised anomaly detection, and foundation model adaptation — bridging theoretical innovation and industrial-grade engineering.
"Specialized in advancing un-/self-supervised and generative machine learning techniques for medical image anomaly detection. I'm passionate about leveraging technology for better healthcare."
Research Highlights
All publications →Foundation Model Adaptation
Adapting Large Vision-Language Models (Qwen-2.5 VL 8B, CLIP) to specialized, dense medical domains. Novel embedding approaches for high-fidelity radiology report generation and visual prompt engineering for VLMs in clinical settings.
Learn more →Anomaly Detection & Localization
Context-encoding VAEs (ceVAE) and gradient/score-based anomaly localization on 3D medical images — introducing inspectability to unsupervised models by visualizing why a model flags an input as anomalous.
Learn more →Benchmarking, Robustness & OOD
Founded the MOOD Challenge — the first international benchmark for medical anomaly localization (80+ teams, Docker-based, running since 2020). Comparative analysis of failure detection, confidence aggregation, and contrastive representations for OOD detection on medical images.
Learn more →Skills & Languages
Programming
Languages
ML / Deep Learning
Infrastructure & DevOps
Frameworks & Tools
From the Blog
All posts →Demystifying Variational Autoencoders (VAEs): From Variational Inference to Deep Generative Models
A from-scratch walkthrough of the math behind VAEs — from ELBO and KL divergence to the reparameterization trick — explained the way I wish someone had explained it to me.
2024-11-20AI-Sciantist: An Autonomous Research Loop That Never Sleeps
I built a closed-loop AI research system that ideates, implements, trains, evaluates, and iterates on ML experiments — with 8 expert personas, HPC cluster integration, and live human steering.
2024-09-15Paper Reader: Listening to Science, One Word at a Time
How I built a web app that reads scientific papers aloud with word-level highlighting, converts LaTeX math to spoken English, and runs on Kubernetes with neural TTS.
Connect
Interested in collaboration, research, or just want to chat about generative AI? Find me on these channels.