Archaeological Classification of Small Datasets Using Meta- and Transfer Learning Methods: A Case Study on Hittite Stele Fragments

In recent years, the increasing application of machine learning (ML) techniques to solve archaeological problems has garnered substantial interest. The relevant issue is that limited-sample, heterogeneous, and non-standard structures of datasets in archaeology affect the quality of deep learning (DL) model efficiency in training and testing. This study explores the effectiveness of artificial neural networks (ANNs) in addressing the unique challenges posed by the classification of Hittite stelae fragments, given a reduced set of well-known samples for training and testing. A hybrid model combining Model-Agnostic Meta-Learning (MAML) and Few-Shot Learning (FSL) is compared to a transfer learning approach (ResNet18 architecture), as well as conventional classificatory methods. This hybrid approach has been successful in other research domains for training ML and DL models with small datasets. This paper critically evaluates these approaches, addressing key challenges based on small training and testing datasets, the effects of fragmented and altered archaeological data in automated learning, and variable documentation quality, thereby underscoring the advantages of the proposed method over conventional ML and statistical techniques. Additionally, insights from a Hittite art expert were incorporated to assess the practical benefits of our models in this study on the same classification task. Our preliminary results on the provenance of stelae from four different Hittite cities indicate that effective classification and origin prediction are achievable with only 208 samples enabling systematic analysis of how different algorithms and human expertise respond to increased data state. Gradient-weighted Class Activation Mapping (GRAD-CAM) analysis confirmed the performance of models in base classification decisions on archaeologically meaningful iconographic features rather than photographic artifacts, validating their authenticity for scholarly applications. The code for programs used in the paper can be accessed on https://github.com/dkayikci/ML-Arch.