Abstract
The increasing volume of data available for archaeological analysis requires automation, and advances in machine learning and artificial intelligence are being deployed to meet these needs. Deep learning approaches, in particular have been very successful in many areas. These methods, however, are something of a 'black box' – while they successfully map from input data to the desired output, they are difficult to interpret. Related to this is the fact that they perform best when trained 'end-to-end' which makes incorporating expert knowledge into them difficult.
In this paper we propose a new direction, incorporating more explicit reasoning into deep learning systems for fine-grained 3D shape analysis. Our target application is the analysis of lithic flakes from pre-contact M¯ aori stone tool manufacture. As well as having significant archaeological interest, this application exemplifies some of the challenges of fine-grained shape analysis. While our work is at an early stage, we believe that the overall approach may be of interest to other members of the community as they explore how to use deep learning to solve archaeological problems.