Archaeology is often romanticized as a field of grand discoveries—golden masks and lost cities. But for the working archaeologist, the reality is often a mountain of "commonware": the everyday, unglamorous kitchen pots and storage jars of antiquity. In a new paper, researchers from the University of Cambridge demonstrate how Variational Autoencoders (VAEs) can turn this overwhelming pile of broken ceramic fragments into a structured map of ancient trade and production.
TL;DR
Researchers developed an unsupervised machine learning pipeline to classify Roman potsherds using 2D sectional profiles. By training a Variational Autoencoder to learn "shape signatures" and applying hierarchical clustering, they identified hundreds of new morphological matches that had escaped decades of manual archaeological study.
What Problem Does This Paper Address?
The "big data" of Roman archaeology isn't stored in a cloud; it’s scattered across thousands of hand-drawn 2D profiles in printed excavation catalogs (corpora). While high-status "finewares" (like fancy dinner plates) are well-standardized and easy to track, "commonwares" are the opposite. They were produced locally, vary wildly in shape, and are usually found as small fragments (potsherds).
Manually matching these fragments across different sites is a Herculean task. Because commonware was the "baseline" of ancient life, failing to classify it means missing crucial data on local economies and site chronologies. The authors sought an automated, unbiased way to find similarities between these fragments without requiring a human expert to pre-define what features (like rim angle or thickness) are important.
Methods
The authors propose a two-step unsupervised workflow:
- Feature Extraction via VAE: They used a deep convolutional Variational Autoencoder (VAE). Unlike a standard autoencoder, a VAE maps inputs to a continuous latent space distribution. This forces the network to learn the underlying "manifold" of pottery shapes.
- Technical Detail: To prevent "KL-vanishing" (where the network ignores the latent space), they implemented a cyclical annealing schedule for the loss function, gradually increasing the weight of the regularization term.
- Hierarchical Clustering: Once the VAE was trained, each potsherd was represented by a 256-dimensional vector (the mean and variance of its latent representation). They used Agglomerative Hierarchical Clustering (specifically the UPGMA method) to group similar profiles. This is superior to methods like k-means because it doesn't require the user to guess the number of clusters in advance and produces a "dendrogram" (a tree-like map) that archaeologists can visually inspect.
Datasets
The study introduced the ROCOPOT (ROman COmmonware POTtery) database, a significant contribution to the field.
- Content: Over 4,000 black-and-white 2D profiles extracted from 25 different archaeological corpora.
- Focus: The current version (v1.0) focuses on rims (3,678 profiles), as they carry the most diagnostic morphological information.
- Preprocessing: Images were cleaned of "noise" (like scan artifacts or template lines), mirrored to a consistent orientation, and resized to 256x256 pixels.
Key Insights and Ideas
- Unbiased Morphometrics: By using an unsupervised approach, the model isn't limited by existing archaeological typologies. It looks at the actual geometry, potentially catching similarities that a human might overlook due to "handcrafted variance" in the drawings.
- Robustness to Noise: Through data augmentation (simulating erosion, dilation, and rotation), the researchers proved that the VAE's latent features are robust to the slight inconsistencies found in hand-drawn archaeological records.
- The "Seed" Concept: The researchers focused on "seeds"—pairs of profiles that the algorithm identified as the closest possible matches at the deepest level of the cluster tree. This provides a manageable starting point for human experts to validate the AI’s findings.
Key Results
The results were both mathematically impressive and archaeologically significant:
- Validation: Out of 1,231 "seeds" identified by the algorithm, archaeologists validated 500 (41%) as genuine positive matches.
- New Discoveries: Most strikingly, of those 500 validated matches, 71% (353 pairs) were entirely new. These were similarities between potsherds from different sites or catalogs that had never been formally linked in the literature.
- Efficiency: The clustering step takes minutes, whereas a manual review of the same dataset would take a specialist months or even years of tedious cross-referencing.
Limitations and Future Work
The authors are transparent about the current scope:
- Rim-Centric: The current model focuses almost exclusively on rims. Future versions aim to incorporate bases and the "main body" of the vessels.
- Human-in-the-Loop: The system is designed to support rather than replace archaeologists. To this end, they developed a MATLAB GUI to help researchers inspect clusters.
- Scaling Up: The next step is to move the tool online, allowing archaeologists to upload new finds from the field and receive instant suggestions for matches within the global ROCOPOT database.
By bridging the gap between deep generative models and heritage science, this work provides a scalable blueprint for organizing the fragmented history of the ancient world.
Summary updated