Learning from Scarce Information: Using Synthetic Data to Classify Roman Fine Ware Pottery.

Area of Science:

• Computer Science
• Archaeology
• Digital Humanities

Background:

• Machine learning models often require large datasets for effective generalization.
• Small datasets, common in specialized fields like archaeology, pose a significant challenge for AI model training.
• Roman pottery classification is hindered by limited, often fragmented, available data.

Purpose of the Study:

• To develop and evaluate a transfer learning approach for improving AI model generalization on small, specialized datasets.
• To address the challenge of learning from limited data in the context of Roman pottery classification.
• To integrate expert knowledge into synthetic data generation for AI model training.

Main Methods:

• A hybrid transfer learning strategy was employed, involving initial training on a synthetic dataset.
• Synthetic data was generated using features from published profile drawings of Roman pottery forms, incorporating expert knowledge.
• The model was subsequently fine-tuned using a smaller dataset of real smartphone photographs of Roman pottery vessels.
• Experiments were conducted using various deep learning architectures and considering factors like image viewpoint and vessel damage.

Main Results:

• The proposed hybrid approach significantly improved classifier generalization performance compared to training solely on original data.
• The method demonstrated robustness across different deep learning architectures and under varying experimental conditions (viewpoint, damage).
• The integration of expert knowledge via synthetic data generation proved effective in enhancing model training.

Conclusions:

• Transfer learning, particularly using expert-informed synthetic data, offers a promising solution for training AI models on small datasets.
• This approach effectively alleviates the fundamental issue of poor generalization in data-scarce domains like archaeological artifact analysis.
• The developed method holds potential for broader applications in classifying historical artifacts and other specialized visual recognition tasks.