A Bayesian Neural Network for Indirect Dating
Author(s): Kenneth Vernon
Year: 2025
Summary
This is an abstract from the "Multiscale Data and the History of Human Development in the US Southwest" session, at the 90th annual meeting of the Society for American Archaeology.
<html>
The two most powerful forces driving long-term development in human societies are climate change and demography, so it should come as no surprise that archaeologists have devoted considerable time and energy to estimating key climate and demographic quantities in the past. To aid in these efforts, specifically demographic reconstruction, this project proposes a new spatio-temporal <b>deep learning</b> (DL) framework for the <b>relative-dating</b> of human settlements. The core of this new framework is a <b>Bayesian Neural Network</b> (BNN) model trained on multi-site archaeological tree-ring date and ceramic assemblage data. In this context, a BNN has two important advantages: (i) unlike other DL models, it is less susceptible to overfitting, and (ii) unlike other archaeological cross-dating methods, it provides a straightforward measure of uncertainty. We also show how to incorporate a <b>Moran Eigenvector Map</b> into a BNN to account for spatial autocorrelation. Training and testing of a BNN for relative-dating are done using the <b>cyberSW</b> database maintained by <b>Archaeology Southwest</b>, which provides an extremely large ceramic assemblage dataset covering much of the US Southwest. Importantly, the approach should be generalizable to any temporally diagnostic artifact type for which count data exist.
</html>
Cite this Record
A Bayesian Neural Network for Indirect Dating. Kenneth Vernon. Presented at The 90th Annual Meeting of the Society for American Archaeology. 2025 ( tDAR id: 509084)
This Resource is Part of the Following Collections
Keywords
Record Identifiers
Abstract Id(s): 50030