Tim Hallyburton
I am a doctoral researcher in the Pattern Recognition in Embedded Systems group at TU Dortmund University, advised by Prof. Dr.-Ing. Gernot A. Fink. I build recognition systems that work with minimal supervision in scarce label settings, such as historical archives, rare alphabets and scripts that only a handful of people can still read (for instance: my own handwriting from last week).
Most machine learning assumes data comes labelled. The collections I care about do not: an expert can annotate a few characters per class, not a few thousand. My work asks how far that handful of examples can be stretched through prototype-anchored learning, self-supervised representations, and models that combine what a page looks like with how its strokes are structured.
Before the PhD I spent five years building production software, most recently as a backend and database architect at MotionMiners, and earlier at CERN on grid computing for the ALICE experiment. That background shapes how I work: I like research that runs, scales, and ends up in someone’s hands.
Open to a research internship. I am looking for a 3–6 month research internship starting in January 2027, with flexibility for a May–July 2027 start. Expected PhD completion: December 2027. Get in touch — I am always glad to hear about interesting problems.
What I work on
Label-efficient learning
Prototype-guided and few-shot classification, deep clustering, pseudo-label expansion and open-set recognition — getting from a handful of exemplars to a usable model.
Representation learning
Masked autoencoders, contrastive objectives and domain-specific handwriting encoders that learn from unlabelled pages.
Geometric deep learning
Graph neural networks and hybrid image–graph architectures, combining visual appearance with the structure of a written trace.
Document analysis
Handwriting and character recognition, writer retrieval and archival information extraction — especially in low-resource settings and rare alphabets.
Selected highlights
- Best Paper Award at the HIP workshop, ICDAR 2026, for PAUL, which turns one to five expert-chosen exemplars per class into a full pseudo-labelled training set.
- Doctoral Consortium Award, ICDAR 2026, Vienna.
- Showed that feeding whole pages to vision-language models beats line segmentation for historical handwriting recognition — cutting Gemini 2.5 Pro’s character error rate on Cyrillic from 22.6% to 3.9%, at lower cost.
- Co-authored X-VLAD, the aggregation module behind writer retrieval at scale: 78.8% mAP on HisFragIR20, +21.6% over prior published work.
- Scholar of the German Academic Scholarship Foundation (Studienstiftung des deutschen Volkes).
Read more about my research See all publications Full CV
Recent news
- Sep 2026Best Paper Award at the HIP workshop and a Doctoral Consortium Award at ICDAR 2026 in Vienna; two further ICDAR papers presented.
- Aug 2026Presented Talking as a Drug at the ETHICAIA workshop, IJCAI 2026 in Bremen.
- Jun 2026Journal manuscript on segmentation and vision-language models submitted to Informatica.
- 2026Attended the Summer School on Document Analysis (SSDA) in Vall de Núria.
- 2025Best Presentation Award and Best Student Paper runner-up at the International Semantic Web Research Summer School (ISWS), Bertinoro.