Abstract
The concept of equifinality is a central issue in taphonomy, conditioning an analyst’s ability to interpret the formation and functionality of palaeontological and archaeological sites. This issue lies primarily in the methods available to identify and characterise microscopic bone surface modifications (BSMs) in archaeological sites. Recent years have seen a notable increase in the number of studies proposing the use of deep learning (DL)-based computer vision (CV) algorithms on stereomicroscope images to overcome these issues. Few studies, however, have considered the possible limitations of these techniques. The present research performs a detailed evaluation of the quality of three previously published image datasets of BSMs, replicating the use of DL for the classification of these images. Algorithms are then subjected to rigorous testing. Despite what previous research suggests, DL algorithms are shown to not perform as well when exposed to new data. We additionally conclude that the quality of each of the three datasets is far from ideal for any type of analysis. This raises considerable concerns on the optimistic presentation of DL as a means of overcoming taphonomic equifinality. In light of this, extreme caution is advised until good quality, larger, balanced, datasets, that are more analogous with the fossil record, are available.