Codes and Expansions (CodEx) Seminar
Andreas Heinecke (Yale-NUS College):
Unsupervised statistical learning in ancient economic history
Ancient coins have been struck from hand-engraved dies. Often hundreds of dies of the exact same design would be commissioned by ancient states to mint millions of coins. The dies themselves are lost and only survive as imprints in the coins struck from them. In principle, it is possible to distinguish different dies by minute differences in the coins, though coins minted from the same die can vary substantially in appearance due to their preservation state. Such "die studies" - the clustering of large samples of coins by the dies they were struck from - are an important tool in ancient economic history. Yet, manual die studies are too labor-intensive to comprehensively study large coinages such as those of the Roman Empire.
From a data science viewpoint, die studies present a challenging unsupervised clustering problem, involving an unknown and large number of highly similar semantic classes of imbalanced sizes, for which training examples would be extremely time-expensive to obtain and are thus not available. We address this problem through determining dissimilarities derived from specifically devised re-weighted Gaussian process-based keypoint features in a Bayesian distance microclustering framework. The resulting method can reduce the time-investment for large-scale die studies by several orders of magnitude, in many cases from years to weeks, and provide data towards longstanding controversial debates on the economic history of ancient states.