Machine Learning (ML) has received much attention over the past decade, particularly in its ability to recognise objects in images. A sub-field of ML known as Deep Learning (DL) performs significantly better at object recognition tasks than previous ML techniques. The combination of lots of data for training plus lots of computing power allows for the multiple (or ‘deep’) layers of computation which makes this approach so effective.
Variations of this high-performance method have quickly become the approach of choice for solving many problems that had previously seemed beyond the capacity of ML. This includes things like handwriting recognition, audio classification, beating humans at Chess and Go, controlling semi-autonomous cars, image processing, and now, detecting archaeology in remote sensing data.
As ML starts to affect many parts of modern life it’s increasingly important that this tool is critically examined and more widely understood. ML and DL models are complicated; they contain many layers of computation and creating a successful model is an iterative process of fine-tuning. Determining what makes a ‘successful’ model is also not straightforwards, as is deciding how to train the model to achieve this success.
The practice of archaeological survey reflects this complexity. Individual interpretive approaches can be hard to untangle, containing many layers of decision-making. Determining the success of survey can also be complicated, as decisions need to be made on where the practicioner derives their ‘truth’ about what is being recorded, or assesses the reliability of their results.