Seeing like a Neural Network? Possibilities and Predicaments of Automated Virtual Archaeological Prospection

Author(s): Steven Wernke

Year: 2019

Summary

This is an abstract from the "Archaeological Vision in the Age of Big Data" session, at the 84th annual meeting of the Society for American Archaeology.

What might it mean to see like a neural network over vast areas of ancient landscapes? Rapid advances in computer vision—especially approaches using Convolutional Neural Networks (CNNs)—have made automated archaeological site and feature detection from satellite and aerial imagery over very large areas an achievable prospect. Such automated prospections will dwarf pedestrian survey coverage, opening new possibilities for understanding past political landscapes. But with the promise of such big archaeology come several predicaments. Most obviously, machine learning-based virtual prospection will necessarily focus on sites with surficially visible—mostly architectural—remains, and will tend toward the upper tiers of settlement hierarchies. Such unavoidable sample bias may tend analytical framings toward top-down framings (seeing like a state) and occlude local variation. Second, all CNN-based automated prospection systems require large, human-generated training datasets, raising ethical and intellectual issues of authorship. Third, CNN-based computer vision systems are ultimately black boxes—how a CNN learns to identify objects cannot be known with any precision, raising fundamental epistemological problems. These promises and predicaments are explored through an emerging project involving the generation of large training datasets through brute force virtual survey and their deployment in a CNN-based prospection in the Andean region.

Cite this Record

Seeing like a Neural Network? Possibilities and Predicaments of Automated Virtual Archaeological Prospection. Steven Wernke. Presented at The 84th Annual Meeting of the Society for American Archaeology, Albuquerque, NM. 2019 ( tDAR id: 451905)

Spatial Coverage

min long: -82.441; min lat: -56.17 ; max long: -64.863; max lat: 16.636 ;

Record Identifiers

Abstract Id(s): 25845