A survey of deep learning for diabetic retinopathy screening (2016–2025), spanning 50+ studies and 20+ datasets. The key takeaway: strong benchmark scores don't equal clinical readiness: the field needs reproducible code, external validation, calibration, and per-patient evaluation.
RL fine-tuning (GRPO) lifted our chest X-ray model's score 23% on one benchmark but dropped it 19% on another: it learned to predict dataset-specific labels, not to read X-rays. The same pattern appears at 50x the budget, so the problem is the recipe, not the resources. Carefully curated SFT generalized better across institutions.
Sensitive documents make collecting real training data hard, so we generate realistic synthetic data instead: template-based documents filled by Faker and LLMs. A BERT classifier trained purely on synthetic data reaches 88% page-wise precision on a real-world test set it never saw during training.
To automate the labor-intensive expansion of nutritional databases, we built a two-stage information retrieval system that matches food items by text and nutrient similarity. Adding SVM-based food category prediction (99% accuracy) boosted retrieval precision to 80%, and we outline an LLM-plus-optimization pipeline for simulating unmatched recipes.
Welcome to the official blog of the Applied Machine Learning Lab at the University of Bonn: our space for research highlights, course announcements, lab news, and tutorials.