Bridging Domains: Transfer Learning in AI-Powered Text Recognition

This presentation was given during the WiNoDa winter school, a five-day intensive course on the topic of Research with Natural Science Collections. Data, Quality, and Methods from 24-28 November 2025. Organization: German Federation for Biological Data e.V. (GFBio) with support from the Museum für Naturkunde Berlin (MfN), German Archaeological Institute (DAI), Vernetzungs- und Kompetenzstelle Open Access Brandenburg (VuK). Abstract: This interactive lecture demonstrates how AI methods developed for digitizing natural history specimen labels can be adapted for ancient text recognition, using two case studies: automated entomological label extraction (ELIE - https://github.com/MargotBelot/entomological-label-information-extraction) and hieroglyphs papyrus analysis (HieraticAI - https://github.com/MargotBelot/HieraticAI). Through live demonstrations and interactive discussions, participants will explore the methodological challenges of domain transfer in computer vision, examining how models trained on museum specimens perform on ancient manuscripts. The session covers practical applications of FAIR data principles, discusses ethical considerations in cultural heritage digitization, and provides interactive experience with validation workflows that combine AI predictions with expert knowledge.