Computational Formalism: Art History and Machine Learning

Any machine learning project must start with a dataset. In the case of the research profiled in this book, that means collections of art images. How these collections are assembled is thus a core factor in what automated processes will tell us about the data. As the old adage among computer programmers goes, “Garbage in, garbage out.” More data is often equated with better (i.e., more accurate) results. For cultural data, however, this is not necessarily the case. There are pressing issues of bias to consider in the compilation of datasets, no matter how large they are. Construction of even the most inclusive art dataset still reflects the interests and perspective of specific cultural actors. It is important to consider who digitizes artworks and why certain artworks are chosen for digitization. For those artworks that are born digital, which works get collected or gain acceptance into the canon of high art?