AI Model Dates Medieval Manuscripts with 5.4-Year Accuracy
An innovative AI model has achieved remarkable accuracy in dating medieval manuscripts, with a mean absolute error (MAE) of just 5.4 years. Utilizing a probabilistic approach, the model assesses visual features to predict dates, generating a full predictive distribution that accounts for uncertainties. The architecture employs an EfficientNet-B2 backbone paired with a Normal-Inverse-Gamma output head, trained with a joint negative-log-likelihood and evidence-regularization objective. Tested on the DIVA-HisDB benchmark, the model successfully analyzed 150 pages from three medieval codices, achieving a 93% accuracy within five years. The findings were published on arXiv under paper ID 2605.06475.
Key facts
- Probabilistic approach for dating historical manuscript pages from visual features alone.
- Poses dating as evidential deep regression over a continuous year axis.
- Outputs full predictive distribution with decomposed aleatoric and epistemic uncertainty.
- Architecture combines EfficientNet-B2 backbone with Normal-Inverse-Gamma (NIG) output head.
- Trained with joint negative-log-likelihood and evidence-regularization objective.
- Tested on DIVA-HisDB benchmark: 150 pages, 3 medieval codices, 151,936 patches.
- Achieves test MAE of 5.4 years, below 50-year century-label supervision granularity.
- 93% of patches within 5 years, 97% within 10 years.
- PICP of 92.6%, best calibration among compared methods.
- Published on arXiv under paper ID 2605.06475.
Entities
Institutions
- arXiv