Loading...
New lens on art history: using complex network analysis and unsupervised machine learning, A
Downie, Khloe N.
Downie, Khloe N.
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2022
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
Visual art is a wholly complex and inherently human creation not easily analyzed and interpreted by digital technology. The subjectivity of art makes its interpretative understanding elusive to machines while virtually instantaneous to the human viewer. This project attempts to demonstrate that there is a relationship between the measurable visual characteristics and the communicative characteristics of art. In doing so, we hope to offer a machine-based software tool that supplements the traditional critical approach to historical art found in art history and art theory. This AI-generated perspective will offer innovative insights to the inherent interpretative information found in art.
This project's methods are seated in the creation of a python-based feature extraction software. The software is an analog to the pluralistic critical approach of art theory. It abstracts images of historical paintings into complex network representations that contain the digital equivalent of formalist elements and principles of design present therein. By measuring the images' network representations, we obtain quantitative descriptions of their innate visual features. We, then, reduce the dimensionality of the measurement data set and find a clustering of the images. From those clusters, we can draw mathematical conclusions about the interpretative characteristics of the images held within. We postulate that the evaluative conclusions enabled by our method's AI-generated art movements will reach beyond those present in traditional art theory.
We measure the interpretative precision of the clustering we obtain using the precision and recall performance measures. We compare the resulting performance from our software to that of a random clustering of images. In doing so, we prove that our software indeed performs better and is statistically distinct from a random grouping of paintings in terms of critical and formalist evaluation. Beyond that, we show that our resulting clustering has greater success in terms of the performance metrics than the critically accepted historical art movements.
These results show that, using complex networks to embody formalist elements and principles of design, measuring those networks, and clustering the paintings based on that data, we are not only able to create distinct groupings of images with common formalist components but common critical interpretations as well. Because of this, we can confirm the existence of an untapped empirical relationship between machine-measurable visual characteristics of images and the communicative concepts held within those images. That we obtained performance better than the historical movements shows that our methods offer the first steps to discerning and building on this correlation.
Associated Publications
Rights
Copyright of the original work is retained by the author.